Intelligent dynamic monitoring method for ecological products

Abstract

The invention provides an intelligent dynamic monitoring method for ecological products. The intelligent dynamic monitoring method comprises the following steps: acquiring an index set for dynamically monitoring the ecological products; monitoring at least one of the following ecological data of air quality, water quality, soil, weather, biodiversity and forest fire in a set range of an area where the ecological product is located based on a set Internet of Things sensor; based on a configured edge node, collecting ecological data from the Internet of Things sensor, preprocessing the ecological data to obtain multi-modal feature data, and transmitting the multi-modal feature data to a background server; and on the basis of an ecological detection model deployed on a background server, feature extraction is performed on the multi-modal feature data, so that the ecological value of the ecological product is predicted on the basis of the extracted features, real-time and automatic monitoring and early warning of the ecological product are realized, and the demand of efficient management of the ecological product under rapid change of the ecological environment is met.

Description

Technical Field

[0001] The present invention relates to the field of intelligent processing technology, and in particular to an intelligent dynamic monitoring method for ecological products. Background Art

[0002] In the field of ecological product monitoring, with the deepening promotion of ecological environmental construction and the concept of sustainable development, the demand for dynamic monitoring of ecological products is becoming increasingly urgent. However, existing ecological product monitoring technologies have numerous limitations and are unable to meet practical application needs. Currently, some ecological product monitoring methods rely primarily on manual inspections and single-point data collection. Manual inspections require significant manpower and time, and the subjective judgment of monitoring personnel can significantly affect data accuracy, making it difficult to ensure data objectivity and consistency. Single-point data collection only captures limited information about a local area and fails to reflect the dynamic changes in the overall ecological product and its surrounding environment. This makes it difficult to form comprehensive and systematic monitoring results, and it fails to provide sufficient data support for the comprehensive evaluation of ecological products. In terms of data processing, traditional methods mostly rely on simple statistical and analytical methods. They lack the ability to effectively integrate and process the multi-source, heterogeneous data involved in ecological product monitoring, such as environmental parameters and biological indicators, and are unable to explore potential correlations and inherent patterns between the data. Furthermore, these methods often use static analytical models that are difficult to adapt to the dynamic characteristics of ecological products under different temporal and spatial conditions. This results in poor timeliness of monitoring results and a failure to reflect the true status of ecological products in a timely manner. When assessing the value of ecological products, existing technologies typically rely on fixed indicator systems and empirical formulas. This approach ignores the complex interactions between ecological products and various ecosystem components, as well as the dynamic impact of ecological and environmental changes on their value. This makes it difficult for assessment results to accurately reflect the actual value of ecological products and to provide a scientific and effective basis for decision-making on the rational development, utilization, and protection of ecological resources. Furthermore, traditional ecological product monitoring and assessment technologies lack intelligent tools, making them incapable of real-time, automated monitoring and early warning of ecological products, making them unable to meet the demand for efficient management of ecological products in a rapidly changing ecological environment. Summary of the Invention

[0003] To achieve the above objectives, according to one aspect of the present application, a method for intelligent dynamic monitoring of ecological products is provided, which includes:

[0004] Obtain an index set for dynamically monitoring ecological products, where the index set includes at least one of the following: ecosystem health indicators, environmental quality indicators, ecological diversity indicators, resource utilization indicators, low-carbon attribute indicators; based on the set Internet of Things sensors, monitor at least one of the following ecological data within the set range of the area where the ecological products are located: air quality, water quality, soil, meteorology, biodiversity, forest fires; based on the configured edge nodes, collect ecological data from the Internet of Things sensors and perform preprocessing to obtain multi-modal feature data, and transmit the multi-modal feature data to the background server; based on the ecological detection model deployed on the background server, perform feature extraction on the multi-modal feature data to predict the ecological value of the ecological products based on the extracted features.

[0005] The technical solution in this application offers at least the following technical benefits: Traditional manual inspections and single-point data collection suffer from inefficiency and incomplete data. This application achieves comprehensive, dynamic monitoring of ecological products by acquiring a multi-dimensional dynamic monitoring indicator set encompassing ecosystem health indicators, environmental quality indicators, and other metrics. It also utilizes IoT sensors to monitor various ecological data, such as air quality and water quality, within a defined range within the area where the ecological product is located. Compared to traditional methods, this approach avoids interference from subjective human factors, significantly improves monitoring efficiency, and enables the acquisition of more comprehensive and objective ecological data, providing a rich and accurate data foundation for subsequent analysis. Traditional data processing methods struggle to effectively integrate and analyze multi-source heterogeneous data. This application, based on configured edge nodes, collects ecological data from IoT sensors, preprocesses it, and generates multimodal feature data, which is then transmitted to a backend server. This process systematically integrates and initially processes multi-source ecological data. Compared to traditional, simple statistical analysis, it can better uncover potential correlations between data, improve the effectiveness and relevance of data processing, and make the data more relevant to the needs of ecological product value prediction. Existing methods for assessing the value of ecological products ignore the complex relationships and dynamic changes in ecosystems, resulting in inaccurate assessment results. This application uses an ecological detection model deployed on a background server to extract features from multimodal feature data, and predicts the ecological value of ecological products based on the extracted features. This solution fully considers the interactive relationship between ecological products and various elements of the ecosystem, and uses the model to achieve dynamic prediction of the value of ecological products. Compared with the traditional evaluation method based on a fixed indicator system and empirical formulas, it can more accurately reflect the actual value of ecological products and provide a reliable basis for scientific decision-making on ecological resources. Traditional ecological product monitoring lacks intelligent means and cannot be automatically monitored and warned in real time. This application has constructed a complete intelligent process from indicator set acquisition, IoT sensor data collection, edge node preprocessing to background server model prediction. It can obtain ecological data in real time and automatically analyze and predict the value of ecological products, promptly detect changes in the status of ecological products, and realize intelligent dynamic monitoring and early warning, making up for the shortcomings of traditional technologies and meeting the needs of efficient management of ecological products under the rapid changes of the ecological environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Figure 1 This is a flow chart of an intelligent dynamic monitoring method for ecological products according to an embodiment of the present application. DETAILED DESCRIPTION

[0007] like Figure 1As shown in the figure, an ecological product intelligent dynamic monitoring method provided by an embodiment of the present application includes: obtaining an index set for dynamically monitoring ecological products, where the index set includes at least one of the following: ecosystem health index, environmental quality index, ecological diversity index, resource utilization index, low-carbon attribute index; based on the configured Internet of Things sensors, monitoring at least one of the following ecological data within a set range of the area where the ecological products are located: air quality, water quality, soil, meteorology, biodiversity, forest fire; based on the configured edge nodes, collecting ecological data from the Internet of Things sensors and performing preprocessing to obtain multi-modal feature data, and transmitting the multi-modal feature data to the background server; based on the ecological detection model deployed on the background server, extracting features from the multi-modal feature data to predict the ecological value of the ecological products based on the extracted features.

[0008] Preferably, in a specific application scenario, the above solution is described in an alternative or preferred manner.

[0009] 1. Obtain the dynamic monitoring index set: Determine the index set for dynamically monitoring ecological products. This set covers various types of indexes, providing a direction for subsequent monitoring and analysis. For this purpose, the index set is In practical applications, the present application can determine the specific content by reading relevant information in the configuration file or database. For example, if the ecological product is the ecological service of a forest area, the index set is This means that subsequent monitoring will mainly focus on the health status of the forest ecosystem and biodiversity.

[0010] 2. Internet of Things sensor data monitoring: Use the Internet of Things sensors arranged in the area where the ecological products are located to monitor various ecological data in real time within a set range. Let the sensor monitoring range be Ω, which is an area defined based on geographical coordinates and environmental parameters. For each type of ecological data d ∈ {air quality, water quality, soil, meteorology, biodiversity, forest fire}, let the data sequence collected by the sensor be where t represents time and T is the total monitoring duration. For example, for air quality data, s_air quality(t) represents the concentration value of a certain pollutant in the air monitored at time t. The present application establishes a communication connection with the Internet of Things sensors and reads these data sequences at a set time interval.

[0011] 3. Edge node data preprocessing: The edge nodes collect ecological data from the Internet of Things sensors and generate multi-modal feature data through a series of preprocessing operations for subsequent transmission and analysis: Spatiotemporal alignment (clock synchronization model based on Kalman filter): Let the timestamps of different types of ecological data S d be t d, due to the deviation of different sensor clocks, time and space alignment is required. The state transition equation of the Kalman filter is defined as x k =Ax k-1 +Bu k-1 +w k-1 , where x k is the state vector at time k, which contains information such as clock deviation; A is the state transition matrix, which describes the change of state over time; B is the control input matrix (which can be set to 0 in this scenario because we are mainly concerned with the natural evolution of clock deviation); u k-1 is the control input (negligible); w k-1 is process noise, with mean 0 and covariance Q k-1 Gaussian distribution. The measurement equation is z k =Hx k +v k , where z k is the measurement value at time k, i.e., the difference between the sensor timestamp and the reference time; H is the measurement matrix, which maps the state vector to the measurement space; v k is the measurement noise, which has a mean of 0 and a covariance of R k The state estimation is continuously updated through the Kalman filter algorithm. Thus, the timestamps of different types of ecological data are calibrated to achieve spatiotemporal alignment. The calibrated data series is recorded as

[0012] Anomaly detection (based on the 3σ criterion and the isolation forest model): for ecological data after spatiotemporal alignment First, according to the 3σ criterion, calculate the mean of the data and standard deviation will satisfy The data points are marked as suspected outliers. Then, the isolation forest model is used to further detect anomalies. Assume that the isolation forest model consists of n trees T i Composition, i=1,…,n. For each data point Calculate its i The path length in Defining anomaly scores Where c(T) is a correction factor related to the number of nodes T in the tree. Set a threshold τ, if The data point is determined to be an outlier and is removed. After anomaly detection, multimodal purified data is obtained. Statistical feature extraction in feature engineering: multimodal purification of data Perform statistical feature extraction and calculate the mean median majority Standard deviation Skewness Kurtosis and other statistical quantities. These statistical quantities form a statistical feature vector F d,stat = [μ d,stat , median d , mode d , σ d,stat , skew d , kurtosis d . Time-domain feature extraction: Using the autocorrelation function calculate the autocorrelation values at different delays τ to obtain the autocorrelation feature vector F d,auto-corr = [R d (1), R d (2), …, R d (K)], where K is the set maximum delay value. In addition, adopt the sliding window technique with a window size of w, and calculate statistics such as the mean and standard deviation within each window to form a sliding window time-domain feature vector F d,sliding-window . Concatenate the autocorrelation feature vector and the sliding window time-domain feature vector to obtain the time-domain feature vector F d,time = [F d,auto-corr , F d,sliding-window . Frequency-domain feature extraction: Perform a fast Fourier transform (FFT) on the multi-modal purified data , where f is the frequency, to obtain the spectrum X d (f). After that, calculate the power spectral density Divide different frequency bands [f l,1 , f u,1 , [f l,2 , f u,2 , …, [f l,M , f u,M , and calculate the energy within each band to form the frequency-domain feature vector F d,freq = [E d,1 , E d,2 , …, E d,M . Construction of multi-modal cross-domain feature set: Concatenate the statistical feature vector F d,stat , the time-domain feature vector F d,time and the frequency-domain feature vector F d,freq to obtain the multi-modal cross-domain feature vector F d = [F d,stat , F d,time , F d,freq . For all ecological data types d, combine these multi-modal cross-domain feature vectors into a multi-modal cross-domain feature set Correlation analysis and feature screening: Construct a feature correlation matrix C, where di , d j ∈ {air quality, water quality, soil, meteorology, biodiversity, forest fire}, the Correlation function calculates the correlation between two feature vectors. An adaptive threshold setting method based on deep learning is adopted. By training a small neural network, with the model prediction accuracy as the goal, the correlation threshold θ is dynamically adjusted. For feature pairs with |C ij | > θ, retain the feature that contributes more to the ecological value prediction (the contribution is evaluated through a preliminary linear regression model), and obtain the filtered multi-modal key feature set In feature construction, feature fusion and transformation: Use a deep embedding network to map the features of different modalities in the multi-modal key feature set to a unified feature space. Let the input of the deep embedding network be and the output be The parameters of the network are trained by minimizing the loss function , where the Target function represents the set embedding result (which can be estimated through prior knowledge or data distribution). For the embedded features, use a non-linear transformation method based on the variational autoencoder (VAE). Let the encoder of the VAE encode into the latent variable where z d follows a Gaussian distribution and are the mean and standard deviation output by the encoder. The decoder decodes the latent variable into the transformed features through . The VAE is trained by minimizing the variational lower bound loss function , where KL is the KL divergence, measuring the difference between two Gaussian distributions, is the probability distribution generated by the decoder. After transformation, obtain the compact latent representation feature set Feature derivation and enhancement: Use the generator G of the generative adversarial network (GAN) to generate new derived features based on the compact latent representation feature set . Let the input of the generator be the noise vector n, and the generated derived feature be . The discriminator D is used to distinguish the generated derived features from the real compact latent representation features, and the GAN is trained by minimizing the adversarial loss function . At the same time, use a reinforcement learning algorithm to enhance the generated derived features. Define the agent A of the reinforcement learning, whose state is the current derived feature and the action space is a series of transformation operations on the features (such as scaling, translation, etc.). The agent interacts with the environment (i.e., the ecological value prediction model) and, according to the reward function (where Accuracy represents the prediction accuracy of the ecological value prediction model, and Model represents the ecological value prediction model) Learn the optimal strategy to generate enhanced derivative features Feature combination and optimization: Adopt a hierarchical combination strategy to combine the compact latent representation features transformed by VAE with the derivative features enhanced by reinforcement learning for hierarchical splicing to obtain combined features Use the genetic algorithm to optimize the combined features. Let the individual in the genetic algorithm be the encoded form I of the feature combination, and the fitness function be F(I) = Accuracy(Model(Decode(I))), where the Decode function converts the encoded form into the actual feature combination. Through selection, crossover, and mutation operations, continuously iterate and update the individual to find the optimal feature combination and generate the final multi-modal feature data

[0013] 4. Ecological value prediction by the background server

[0014] The background server uses the deployed ecological detection model to extract features from the multi-modal feature data and predict the ecological value of ecological products based on the extracted features: Spatial feature extraction layer: Let the spatial feature extraction layer of the ecological detection model perform convolution operations on the multi-modal feature data with a convolution kernel of K i , i = 1, …, N, a stride of s, and a padding of p. For the input feature map F, the convolution operation is defined as where (x, y) are the coordinates of the feature map, and M and N are the sizes of the convolution kernel. After the convolution operation, perform downsampling on the feature map through a pooling operation (such as max pooling or average pooling). Let the pooling window size be w pool , with a stride of s pool . The max pooling operation is defined as where (x pool , y pool ) are the starting coordinates of the pooling window on the feature map. By taking the maximum value in each pooling window, downsampling of the feature map is achieved, reducing the data dimension while retaining important spatial feature information. Then, apply the activation function σ (such as the ReLU function, σ(z) = max(0, z)) to the pooled feature map to introduce a non-linear transformation and enhance the model's ability to express innovative spatial features. After these operations, output the spatial feature tensor T space . Temporal feature processing layer: Based on the spatial feature tensor T space , extract temporal information through a gating mechanism. Let the gating mechanism consist of an input gate i t , a forget gate f t , and an output gate o tcomposed of the input feature vector h at time t t-1 (the hidden state at the previous time step) and the input feature x at the current time step t (a slice from the spatial feature tensor T space ), the input gate i t =σ(W ix x t +W ih h t-1 +b i ), the forget gate f t =σ(W fx x t +W fh h t-1 +b f ), the output gate o t =σ(W ox x t +W oh h t-1 +b o ), where W ix , W ih , W fx , W fh , W ox , W oh are weight matrices, and b i , b f , b o are bias vectors. Meanwhile, calculate the cell state c t =f t ⊙c t-1 +i t ⊙tanh(W cx x t +W ch h t-1 +b c ), where W cx , W ch are weight matrices, and b c is a bias vector, and ⊙ represents element-wise multiplication. Finally, the hidden state h t =o t ⊙tanh(c t ). Through such a gating mechanism, a slice of the spatial feature tensor is processed at each time step, and a feature sequence containing time-dependent information is output where T seq is the total number of time steps in the time series. Fully connected layer: Input the feature sequence containing time-dependent information into the fully connected layer. Let the weight matrix of the fully connected layer be W fc , and the bias vector be b fc , then the output of the fully connected layer Among them, the Flatten function unfolds the feature sequence into a one-dimensional vector. Let the ecological value of the ecological product be a scalar V, and through the loss function to measure the difference between the predicted value and the true value, and use the backpropagation algorithm to update the parameters of the ecological detection model to minimize the loss function, so that the model can accurately predict the ecological value of the ecological product based on the extracted features. In practical applications, the present application performs these calculations on the background server, and improves the accuracy of the model in predicting the ecological value of the ecological product through continuous iterative training. For this reason, the above preferred or alternative technical solutions have the following technical advantages.

[0015] 1. Traditional ecological product monitoring often uses simple averaging or thresholding methods to process data from different sensors during data acquisition and preprocessing, failing to fully consider the temporal and spatial characteristics and potential relationships of the data. For example, traditional methods simply align timestamps for ecological data collected by different sensors, ignoring the dynamic changes in clock bias and differences in the spatial distribution of data. This results in poor data fusion and hinders subsequent analysis. This solution utilizes a clock synchronization model based on a Kalman filter for spatiotemporal alignment, dynamically calibrating clock biases across sensors and more accurately integrating multi-source ecological data. In anomaly detection, combining the 3σ criterion with the isolation forest model is more effective than traditional single-source methods in identifying outliers in innovative ecological data. Feature engineering, by comprehensively extracting statistical, temporal, and frequency domain features, and performing correlation analysis and feature screening, can uncover richer and more representative feature information compared to traditional approaches that rely solely on single-type features. This innovative processing provides high-quality, multi-dimensional feature data for subsequent ecological value prediction, enhancing the reliability and integrity of the data foundation. 2. Traditional feature construction methods are often simplistic and straightforward, such as simple feature concatenation or empirical feature selection. These methods fail to fully exploit the nonlinear relationships and underlying structure between features, resulting in insufficient model representation of the innovative ecological phenomena of ecological products. This approach demonstrates a high degree of innovation in the feature construction stage. First, a deep embedding network and variational autoencoder (VAE) are used to map multimodal key features into a unified space and perform nonlinear transformations. This not only reduces dimensionality and removes redundancy, but also explores the underlying structure and distribution patterns of the features, generating compact and representative features. Next, a generative adversarial network (GAN) is used to generate derived features, enriching feature diversity and enhancing them through reinforcement learning to maximize model predictive performance. Finally, hierarchical combination and genetic algorithms are used to optimize feature combinations, finding the optimal feature combination more intelligently and efficiently than traditional methods. These techniques enable the constructed multimodal feature data to better reflect the ecological value of ecological products, significantly improving model prediction accuracy and generalization. 3. Traditional ecological value prediction models, with their simple architectures, such as simple linear regression or shallow neural networks, are unable to effectively handle the high dimensionality, nonlinearity, and spatiotemporal novelty of ecological data, resulting in inaccurate and unstable predictions. The ecological detection model constructed in this solution adopts a layered architecture. The spatial feature extraction layer, through convolution, pooling, and activation operations, effectively extracts spatial features from multimodal feature data, capturing local patterns and important information in the spatial dimension of ecological data. Compared to traditional methods, it can more meticulously depict the spatial distribution of ecological phenomena. The temporal feature processing layer, utilizing gating mechanisms such as LSTM or GRU-like structures, effectively captures temporal information in the spatial feature tensor and processes the dynamic changes and dependencies of ecological data over time, which is difficult for traditional models to achieve.The fully connected layer predicts ecological value based on the previously extracted spatiotemporal features and optimizes parameters through backpropagation, enabling the model to more accurately learn the innovative mapping relationship between ecological features and ecological value. The synergistic effect of this layered architecture enables the model to comprehensively and deeply understand ecological data, significantly improving the accuracy and reliability of ecological value predictions for ecological products. 4. Traditional ecological monitoring and prediction technologies are poorly adaptable to ecological and environmental changes, data noise, and anomalies. Model performance degrades significantly when sudden changes in the ecosystem or data interference occur. This solution, integrating multiple innovative technologies throughout the entire process from data preprocessing to model prediction, demonstrates enhanced adaptability and robustness. For example, effective detection and processing of anomalous data in the data preprocessing stage makes the model more resilient to noise and outliers. The mining and optimization of multiple features during feature construction enhances the model's adaptability to innovative changes in the ecological environment. The layered architecture and parameter optimization methods of the ecological detection model enable stable learning and prediction across diverse ecological scenarios. This comprehensive technical advantage ensures accurate and reliable dynamic monitoring and prediction of the ecological value of ecological products in innovative and changing ecological environments.

[0016] Optionally, the configured edge node collects ecological data from the IoT sensor and preprocesses it to obtain multimodal feature data, including: performing spatiotemporal alignment on different types of ecological data collected from the IoT sensor based on a clock synchronization model based on Kalman filtering; performing anomaly detection on the spatiotemporal aligned ecological data based on the 3σ criterion and the isolation forest model to eliminate abnormal ecological data and obtain multimodal purified data accordingly; and performing feature engineering on the multimodal purified data to generate multimodal feature data.

[0017] Preferably, in a specific application scenario, the above solution is described in an alternative or preferred manner.

[0018] 1. Time and space alignment based on Kalman filter clock synchronization model

[0019] In this application, the system state vector in the process of ecological data collection is assumed to be X k , which is a highly comprehensive vector that comprehensively covers the clock bias, spatial position bias, and other factors that affect spatiotemporal consistency of each sensor. Suppose in an ecological monitoring scenario, involving multiple different types of sensors (such as n air quality sensors, m water quality sensors, etc.), considering the high-order changes in three-dimensional spatial position and clock bias, X k It can be expressed as: X k =[Δt 1,k ,Δt 2,k ,…,Δt n+m,k ,Δx 1,k ,Δy1,k , Δz 1,k , …, Δx n+m,k , Δy n+m,k , Δz n+m,k , α 1,k , α 2,k , …, α p,k T , where Δt i,k represents the clock deviation of the i-th sensor at time k relative to the global time reference; Δx i,k , Δy i,k , Δz i,k respectively represent the position deviations of the i-th sensor in three-dimensional space relative to the global space reference; α j,k represents the other p factors affecting spatio-temporal alignment (such as the drift coefficient of the sensor itself, the influence parameters of environmental factors on the sensor accuracy, etc.). In this application, the state of the system changes over time following the state transition equation X k = A k X k-1 + B k u k-1 + w k-1 . A k is a time-varying state transition matrix, which details how the state transfers from time k - 1 to time k. Considering the dynamic change characteristics of different factors, the elements of A k are functions of time k. For example, for the clock deviation part, due to reasons such as clock aging, its change over time is not constant, and the submatrix elements corresponding to the clock deviation in A k can be expressed as:

[0020] where β i,k is a parameter related to the clock aging or other dynamic changes of the i-th sensor. For the spatial position deviation part, considering the influence of existing environmental factors (such as crustal movement, water flow, etc.) on the sensor position, the submatrix elements corresponding to the spatial position deviation in A k can be expressed as:

[0021] where γ xi,k , γ yi,k , γ zi,k are the dynamic parameters related to the environmental influence on the i-th sensor in the x, y, and z directions respectively. And for the other factor α j,k , its transfer characteristics are also reflected by the corresponding elements in A k . B k is a time-varying control input matrix, u k-1 ​is the control input vector. In most natural ecological monitoring scenarios, without external active control to adjust the clock and position, B k = 0 (zero matrix), u k-1 = 0 (zero vector). w k-1 is the process noise vector, which represents the inevitable random interference during the system state transition. Assume it follows a Gaussian distribution with a mean of 0 and a covariance of Q k-1 i.e., Q k-1 is a time-varying symmetric positive definite matrix, and its elements are determined according to the actual situations such as the stability of the sensor clock and position, and the statistical characteristics of environmental interference. For example, in a seismically active area, the Q k-1 elements related to the spatial position deviation will be relatively large, reflecting the increased uncertainty of position changes. In this application, the relationship between the sensor measurement value and the system state is described by the measurement equation Z k = H k X k + v k Z k is the measurement vector, which synthesizes the spatio-temporal related measurement values actually obtained from the sensors, including the timestamp differences of the data collected by different sensors, the position information differences obtained through high-precision positioning technology, and other related measurement values. Assume that through a variety of advanced measurement technologies, Z k can be expressed as:

[0022] Z k = [Δt meas1，k ,Δt meas2，k ,…,Δt meas(n+m)，k ,Δx meas1，k ,Δy meas1，k ,Δz meas1，k ,Δz meas(n+m)，k ,ω 1，k ,ω 2，k ,…,ω q，k T

[0023] where Δt measi,k is the difference measurement value between the measurement time of the i-th sensor and the global time reference; Δx measi,k ,Δy measi,k ,Δz measi,k are the difference measurement values of the three-dimensional spatial position of the i-th sensor and the global spatial reference; ω j,k is the measurement value related to other factors affecting spatio-temporal alignment. H k is a time-varying measurement matrix, which maps the system state vector to the measurement space, and its form depends on the measurement principle, sensor accuracy, and actual measurement configuration. Due to the different measurement characteristics of different sensors, H k ​is a matrix whose elements are determined according to the measurement model of a specific sensor. For example, for a certain type of high-precision clock sensor with a high precision in measuring clock deviation, the elements of the row vector corresponding to the clock deviation measurement in H k will more accurately reflect the clock deviation information in the state vector. v k is the measurement noise vector, and it is also assumed to follow a Gaussian distribution with a mean of 0 and a covariance of R k , that is R k is a time-varying symmetric positive definite matrix that reflects the intensity and characteristics of the measurement noise. For example, the measurement noise characteristics of different types of sensors are different, and the elements of the sub-matrix corresponding to the measurement values of different sensors in R k will be different.

[0024] The Kalman filter update steps in this application are as follows:

[0025] Prediction step: First, perform state prediction, This application is the predicted value of the state at time k based on the state estimate at time k - 1, [[ID=XX]] is the optimal state estimate value at time k - 1. At the same time, the prediction covariance where P k|k-1 is the covariance of the predicted state, and P k-1|k-1 is the optimal state estimate covariance at time k - 1. This step predicts the state and its uncertainty at the current time through the state transition equation and the optimal estimate at the previous time. In this application, update step: Then, update the predicted state according to the measurement value. Kalman gain It weighs the weights of the predicted value and the measurement value in the update process. Since A k , H k , Q k-1 and R k are all time-varying, the Kalman gain is also dynamically adjusted over time to adapt to changes in the system state and measurement characteristics at different times. Optimal state estimate Through the above formula, combined with the measurement value Z k corrects the predicted state to obtain the optimal state estimate value at time k At the same time, update the covariance P k|k =(I - K k H k )P k|k-1 It should be noted that there seems to be an error in the original text where "XX" is used in the translation. It should be the correct "K" in the context of the Kalman gain formula., where I is the identity matrix. This step optimizes the predicted state based on the measurement values while updating the uncertainty of the state estimate. By continuously iterating the prediction and update steps of the above Kalman filter, this application dynamically and accurately estimates and corrects different sensor clock biases, spatial position biases, and other factors affecting spatio-temporal alignment based on the real-time data collected by the sensors, thereby achieving high-precision alignment of different types of ecological data in time and space. For example, in a large ecological monitoring area containing various meteorological, water quality, and soil sensors, using the above Kalman filter model, it is possible to process the data of each sensor in real time, eliminate the spatio-temporal inconsistency problems caused by clock and position biases, and provide accurate spatio-temporal synchronized data for subsequent ecological data analysis.

[0026] 2. Anomaly detection based on the 3σ criterion and the Isolation Forest model

[0027] In this application, for a certain type of ecological data sequence after spatio-temporal alignment (such as the data sequence of the PM2.5 concentration varying with time collected by a certain air quality sensor), first calculate its mean and standard deviation According to the 3σ criterion, the data points satisfying |x t - μ| > 3σ are marked as suspected anomaly points. This is based on the normal distribution assumption of the data. Under normal circumstances, about 99.7% of the data should fall within the range of the mean ± 3 times the standard deviation, and the data points outside this range are very likely to be abnormal. However, the actual ecological data does not fully conform to the normal distribution, so the 3σ criterion is only used for preliminary screening. In this application, for the data preliminarily screened by the 3σ criterion, the Isolation Forest model is further used for in-depth anomaly detection. The Isolation Forest model consists of n trees For each data point x in the dataset, in each tree T i , starting from the root node, randomly select an attribute dimension according to the eigenvalue of the data point, and randomly select a splitting point within the value range of this dimension to divide the dataset into left and right subtrees. The data point x traverses down the tree until it reaches the leaf node. The path length h i (x) is defined as the number of edges passed from the root node to the leaf node. To make the path lengths of different trees comparable, the path length is standardized, and a correction factor c(T) is introduced, which is a function of the number of nodes T of the tree. For example where H(T - 1) is the harmonic number, Calculate the average path length of the data point x in all trees Then define the anomaly score An appropriate threshold τ is set (e.g., determined through cross-validation or based on domain knowledge). If s(x) > τ, the data point x is considered an outlier and removed. After preliminary screening using the 3σ criterion and in-depth testing using the isolation forest model, multimodal purified data is obtained—a dataset free of outliers. This combined approach can more accurately identify outliers in ecological data and better address the novelty and non-normal distribution characteristics of real-world ecological data compared to single methods.

[0028] 3. Feature Engineering for Multimodal Cleansed Data

[0029] In this application, for each modal data in the multimodal purification data (for example, air quality data, water quality data, etc. as different modalities), comprehensive statistical feature extraction is performed. Calculate the mean median majority Standard deviation Skewness Kurtosis In addition, high-order statistics such as the fifth-order central moment are also calculated. Sixth-order central moment At the same time, considering the quantile information of the data, the values of different quantiles are calculated, such as quartiles Q1, Q2 (i.e., median, which confirms the median calculated above), Q3, and percentile P 10 、P 25 、P 75 、P 90 For ordered data sequences When calculating the quartile Q1, first determine the position If i1 is an integer, then If i1 is not an integer, set i1 = k + f, where k is the integer part and f is the fractional part, then Q1 = (1-f) x k +fx k+1 Similarly, Q3 can be calculated, position The calculation method is similar to Q1. For percentile P p (p means percentage), position Determine P in the same way as for quartiles. pValues. These quantile information can describe the distribution of data at different positions. Combined with other statistics, they can more comprehensively characterize the overall characteristics of the data. For example, the interquartile range IQR = Q3 - Q1 is used to measure the dispersion of data and is insensitive to outliers, which can complement the deficiency of the standard deviation in describing data dispersion. Different percentiles help this application understand the value of data at a specific proportion, which is of great significance for analyzing different degrees of changes in ecological data. Combining these basic statistics, higher-order statistics, and quantile information into a statistical feature vector F stat , which comprehensively summarizes the central tendency, dispersion, distribution pattern, and value characteristics at different positions of the data, providing a rich statistical feature basis for subsequent analysis. In this application, for each modality of ecological data sequence The autocorrelation function is used to analyze the dependence relationship of data in the time series. The autocorrelation function is defined as: where τ is the time delay and μ is the mean of the data sequence. By calculating the autocorrelation function values R x (τ) at different delay τ values, an autocorrelation sequence is obtained where L is the preset maximum delay value. This autocorrelation sequence reflects the similarity degree of data at different time intervals, revealing the periodic and trend characteristics of data changes over time. For example, if the autocorrelation function values are relatively high at certain specific τ values, it indicates that there is a strong correlation in the data at this time interval, suggesting a certain periodic change in ecological phenomena, such as diurnal changes and seasonal changes. In this application, based on the results of autocorrelation analysis, a sliding window technique is used to capture the dynamic change characteristics of data within a local time range. The size of the sliding window is set to w, and the step size is s (s ≤ w). Within each sliding window, a series of statistics are calculated again, such as the mean within the window The standard deviation within the window The peak value within the window The valley value within the window etc. In addition, the slope of the data within the window is calculated to reflect the change trend of the data within the window. Let the data within the window be {x t , x t+1 , …, x t+w-1}, and a straight line y = ax + b is fitted by linear regression method to make minimized, and the obtained slope a is used as the trend characteristic within the window. Combining these characteristics calculated at different sliding window positions t into a sliding window time-domain feature matrix F sliding-window, where each row corresponds to a feature vector at a window position and each column corresponds to a type of feature (such as mean, standard deviation, peak value, etc.). In this application, in addition to autocorrelation analysis and sliding window feature calculation, wavelet transform is also performed on the time-domain data to obtain richer time-domain and frequency-domain features. A suitable wavelet basis function ψ(t) (such as Daubechies wavelet, Haar wavelet, etc.) is selected for the ecological data sequence to perform continuous wavelet transform: where a is the scale parameter that controls the stretching of the wavelet function; b is the translation parameter that controls the translation of the wavelet function; ψ * is the complex conjugate of the wavelet basis function ψ. By calculating at different scales a and displacements b, the wavelet coefficient matrix W x is obtained. Wavelet transform can decompose the time-domain signal into different frequency channels while retaining the time-domain information, which has unique advantages for analyzing non-stationary ecological data. For example, the wavelet coefficients at different scales reflect the change characteristics of ecological data at different time scales, which helps to discover transient changes and hidden periodic components in the data. Some key features are extracted from the wavelet coefficient matrix W x , such as the energy of wavelet coefficients at different scales the entropy of wavelet coefficients , etc. These features further enrich the time-domain feature representation. The autocorrelation feature sequence the sliding window time-domain feature matrix F sliding-window and the features extracted from wavelet transform are combined into a time-domain feature vector F time , which comprehensively reflects the dependence relationship, local dynamic changes and characteristics at different time scales of ecological data in the time series. In this application, for each modality of ecological data sequence fast Fourier transform is performed to convert the time-domain signal into a frequency-domain signal to reveal the energy distribution of the data at different frequencies. where f is the frequency, the spectrum X(f) is efficiently calculated through the FFT algorithm. It is a complex sequence, whose amplitude |X(f)| represents the energy magnitude of the signal at frequency f, and the phase ∠X(f) represents the phase information of the signal at that frequency. Based on the spectrum X(f) obtained by FFT, the Welch method is used for power spectral density (PSD) estimation to more accurately describe the distribution of signal power in the frequency domain. The data sequence is divided into K overlapping sub-segments, each sub-segment has a length of N (usually N < T), and let the k-th sub-segment be FFT is performed on each sub-segment to obtain X k (f), and then the power spectral density estimate value is calculated: This power spectral density estimate It reflects the distribution of signal power in the frequency domain more smoothly and accurately than the estimation directly based on the squared magnitude of the FFT, which helps to analyze the contribution of different frequency components in ecological data to the overall signal energy. In this application, according to the characteristics and analysis requirements of ecological data, different frequency bands [f l,1 , f u,1 , [f l,2 , f u,2 , …, [f l,M , f u,M are divided. Calculate the energy within each frequency band: Approximate the integral value by numerical integration methods (such as the trapezoidal integration method). These frequency band energy characteristics E m (m = 1, …, M) reflect the energy distribution of ecological data in different frequency intervals, which is crucial for understanding the relationship between ecological phenomena and frequency. In addition, calculate high-order frequency domain characteristics, such as frequency domain skewness and frequency domain kurtosis where μ f and σ f are the mean and standard deviation of the power spectral density estimate respectively, and F is the number of frequency samples. These high-order frequency domain characteristics can further characterize the shape of the power spectral density distribution and provide more in-depth information for the frequency domain analysis of ecological data. Combine the frequency band energy characteristics and high-order frequency domain characteristics into a frequency domain feature vector F freq . In this application, the statistical feature vector F stat obtained by statistical feature extraction for each type of modal data, the time domain feature vector F time obtained by time domain feature extraction, and the frequency domain feature vector F freq obtained by frequency domain feature extraction are concatenated to form a multi-modal cross-domain feature vector F. For M types of modal data, the multi-modal cross-domain feature set where F i is the multi-modal cross-domain feature vector of the i-th type of modal data. In this application, construct a feature correlation matrix C, and its element C ij represents the correlation between the i-th feature vector and the j-th feature vector, which is calculated by the Pearson correlation coefficient: where F ik and F jk are the k-th elements of the i-th and j-th feature vectors respectively, and are their means respectively. Adopt an adaptive threshold setting method based on deep learning to train a neural network model with the prediction accuracy of the ecological detection model as the objective function. By continuously adjusting the threshold, the feature combination selected under this threshold can maximize the prediction performance of the ecological detection model. For the absolute value of the correlation |C ijFeature pairs with > θ (θ is an adaptive threshold) are considered to have a large redundancy in information expression. According to the contribution degree of the feature pairs to the ecological value prediction (evaluated by preliminary training on the ecological detection model), one feature with a higher contribution degree is selected to be retained, so as to obtain the screened multi-modal key feature set. In this application, a deep embedding network is used. The multi-modal key feature set The different modal features in are mapped to a unified feature space. The deep embedding network Has multiple hidden layers. Let the weight matrix from the input layer to the first hidden layer be W1, the bias vector be b1, and the weight matrices and bias vectors between the hidden layers be W l , b l (l = 2,…, L, L is the total number of hidden layers), the weight matrix of the output layer is W out , and the bias vector is b out . For the input multi-modal key feature vector F in , after the transformation of the network: h1 = σ(W1F in + b1), h l = σ(W l h l-1 + b l ), l = 2,…, L, F embedded = W out h L + b out , where σ is the activation function, such as the ReLU function σ(z) = max(0, z). The deep embedding network is trained by minimizing the loss function , where Target i is the set embedding result of the i-th modal feature in the unified feature space (which can be obtained by prior knowledge or data distribution estimation). For the embedded features, a non-linear transformation method based on the variational autoencoder (VAE) is adopted. Let the encoder of the VAE encode F embedded into the latent variable z = Encoder(F embedded ), where z follows a Gaussian distribution μ and σ are the mean and standard deviation vectors output by the encoder. The decoder decodes the latent variable into the transformed feature through F transformed = Decoder(z). The VAE is trained by minimizing the variational lower bound loss function , where KL is the KL divergence, which measures the difference between two Gaussian distributions, is the probability distribution generated by the decoder. After the transformation, a compact latent representation feature set is obtained In this application, the generator G of the generative adversarial network (GAN) is used, based on the compact latent representation feature set Generate new derivative features. Let the input of the generator be the noise vector n, and the generated derivative feature be F derived = G(n, F compact ). The discriminator D is used to distinguish the generated derivative features from the real compact latent representation features, and the GAN is trained by minimizing the adversarial loss function . In this application, meanwhile, a reinforcement learning algorithm is used to enhance the generated derivative features. Define the agent A of the reinforcement learning, whose state is the current derivative feature F derived , and the action space is a series of transformation operations on the feature (such as linear and non-linear transformations like scaling, translation, rotation, etc.). The agent interacts with the environment (i.e., the ecological value prediction model), and learns the optimal policy according to the reward function R = Accuracy(Model(F derived )) - Accuracy(Model(F compact ). Among them, Accuracy(Model(·)) represents the prediction accuracy of the ecological value prediction model when inputting the corresponding feature. This reward function measures the improvement degree of the prediction accuracy of the ecological value prediction model when using the generated derivative feature F derived compared with using the compact latent representation feature F compact . In this application, at each time step t, the agent selects an action a from the action space according to the current state t . The action a t acts on the current derivative feature to generate a new derivative feature where the ApplyAction function represents the operation of applying the action a t to the feature . In this application, then, the agent inputs the new derivative feature into the ecological value prediction model Model and calculates the reward . The goal of the agent is to maximize the long-term cumulative reward by continuously trying different actions where γ is the discount factor, and its value range is [0, 1], which determines the importance of future rewards relative to the current reward. A smaller γ means the agent pays more attention to immediate rewards, while a larger γ makes the agent pay more attention to long-term rewards. In this application, in order to learn the optimal policy, the agent uses a reinforcement learning algorithm such as the Deep Q-Network (DQN). In DQN, the agent maintains a Q-network Q(s, a; θ), which estimates the expected long-term cumulative reward of taking the action a in the state s, and θ is the parameter of the Q-network. At each time step, the agent selects an action according to the ∈-greedy policy, that is, randomly selects an action with probability ∈, and selects the action that maximizes Q(s t, the action that maximizes. After collecting sufficient experience samples (s t , a t , R t+1 , s t+1 ), the agent updates the parameters θ of the Q-network by minimizing the loss function , where is the experience replay buffer that stores the experience samples of the agent's interaction with the environment, and θ - are the parameters of the target Q-network, which are periodically copied from the current Q-network to stabilize the learning process. By continuously interacting with the environment and updating the Q-network, the agent gradually learns the optimal policy, that is, to find a series of transformation operations that can maximize the accuracy of the ecological value prediction model for the generated derivative features F derived . Finally, the enhanced derivative features F enhanced are generated. These features can more effectively reflect the ecological value-related information of ecological products and provide better materials for subsequent feature combination and optimization. In this application, when performing hierarchical feature combination, based on domain knowledge and data characteristics, a hierarchical combination strategy is adopted to combine the compact latent representation features F compact and the enhanced derivative features F enhanced . First, the compact latent representation features and the enhanced derivative features are grouped according to the modality. For example, for data of different modalities such as air quality, water quality, and soil, the corresponding compact latent representation features and enhanced derivative features are processed respectively. For each modality, the compact latent representation features and the enhanced derivative features are hierarchically concatenated. Specifically, a multi-layer feature combination structure is constructed. In the first layer, and are simply concatenated to obtain . In the second layer, a non-linear transformation is performed on the combined features obtained in the first layer, for example, processed by a multi-layer perceptron (MLP). Let the input of the MLP be , and after passing through the hidden layer transformation of the MLP: where σ is the activation function (such as the ReLU function), and are the weight matrix and bias vector of the l-th hidden layer respectively. Then the output of the last hidden layer is concatenated with the combined features of the first layer again to obtain . Through this hierarchical combination method, the advantages of the compact latent representation features and the enhanced derivative features can be fully utilized, the deeper relationships between the features can be mined, and a feature representation with a more hierarchical structure and expressive ability can be formed It helps the model better understand the innovative relationship between the ecological value of ecological products and various characteristics. In this application, when optimizing features based on the genetic algorithm, the genetic algorithm is used to optimize the features after hierarchical combination to find the optimal feature combination method and further improve the performance of the ecological value prediction model. First, encode the feature combination. Each feature dimension is regarded as a gene, and the feature combination is encoded as a chromosome I m . For example, if is a D-dimensional feature vector, then chromosome I m is an encoded vector of length D, and the value of each gene indicates whether the feature dimension is retained or undergoes a certain transformation (such as a scaling factor, etc.). Define the fitness function F(I m ) = Accuracy(Model(Decode(I m ))), where the Decode(I m ) function decodes the encoded chromosome I m into the actual feature combination, and Accuracy(Model(·)) is the prediction accuracy of the ecological value prediction model when inputting the corresponding feature combination. The fitness function measures the contribution of each chromosome (i.e., feature combination) to the performance of the ecological value prediction model. In each generation of the genetic algorithm, perform the following operations: Selection: According to the value of the fitness function, use selection strategies such as roulette wheel selection or tournament selection to select a certain number of chromosomes from the current population as parents. Chromosomes with higher fitness have a greater probability of being selected, which means that better feature combinations have a greater chance of participating in the reproduction of the next generation. Crossover: Perform a crossover operation on the selected parent chromosomes to generate offspring chromosomes. For example, use single-point crossover or multi-point crossover methods, randomly select one or more crossover points, and exchange the gene segments on both sides of the crossover points of the parent chromosomes, thus generating new feature combinations. Mutation: Mutate the genes of each offspring chromosome with a certain mutation probability. Mutation randomly changes the value of a certain gene. For example, for the gene representing the feature dimension scaling factor, randomly adjust its value to introduce new genetic diversity into the population and avoid the algorithm falling into a local optimum. After multiple generations of evolution, the genetic algorithm continuously searches the space of feature combinations and gradually finds the chromosome that maximizes the fitness function, that is, the optimal feature combination. Combine the feature combinations optimized by the genetic algorithm for all modalities again to generate the final multi-modal feature data These feature data can maximize the prediction accuracy and efficiency of the ecological value prediction model and provide strong support for the ecological value prediction of ecological products.

[0030] Therefore, the above preferred or alternative technical solutions have the following technical advantages.

[0031] 1. Traditional spatio-temporal alignment methods often adopt simple timestamp matching or calibration methods based on fixed parameters. For example, in some simple ecological monitoring systems, only by periodically manually synchronizing the sensor clocks, the dynamic changes of clock deviations and the innovative characteristics of spatial position deviations of different sensors are ignored. This method cannot adapt to the innovative and changeable ecological monitoring environment. When facing clock and position deviations caused by factors such as sensor aging and environmental interference, it is difficult to achieve precise spatio-temporal alignment, thus affecting the accuracy and consistency of data. The spatio-temporal alignment technology based on the innovative Kalman filter model in this solution has significant advantages. From the definition of the state space model, it comprehensively considers various factors affecting spatio-temporal alignment, including clock deviations of multiple sensors, three-dimensional spatial position deviations, and other potential influencing factors, and can accurately depict the spatio-temporal changes in the innovative ecological monitoring scenario. The time-varying state transition matrix and measurement matrix enable the model to dynamically adapt to the changes in different sensor characteristics and environmental factors. For example, by modeling dynamic parameters such as clock aging and the impact of the environment on sensor positions, the state transition and measurement processes can be adjusted in real time. Compared with traditional fixed-parameter methods, it can more accurately track the spatio-temporal deviations of sensors. In the Kalman filter update step, the time-varying Kalman gain dynamically adjusts the weights of the predicted value and the measured value according to the real-time changes of the system state and measurement noise, ensuring that data can be optimally fused at different times to achieve high-precision spatio-temporal alignment. This enables this application to perform subsequent analysis based on accurately aligned data when processing ecological data collected by different sensors, greatly improving the quality and usability of the data, and providing a solid foundation for ecological analysis. 2. Traditional anomaly detection methods usually rely solely on some simple rules or basic statistical methods, such as only using fixed thresholds to detect anomalies, or only based on simple statistical distributions. In the field of ecological data, the distribution of ecological data is often innovative and non-standard, and traditional methods are prone to high false alarm rates and missed alarm rates. For example, simply judging whether water quality data is abnormal based on a simple fixed threshold cannot adapt to the natural fluctuations of water quality in different seasons and regions, misjudging normal fluctuations as anomalies, or missing real anomalies. This solution combines the 3σ criterion and the isolation forest model for anomaly detection, which has obvious advantages. The 3σ criterion is used as a preliminary screening. Based on the basic statistical characteristics of the data, it quickly marks the data points that deviate significantly from the mean, narrowing the scope for subsequent more accurate detection. The isolation forest model deeply explores the distribution characteristics of the data and evaluates the isolation degree of data points by constructing multiple trees. Its advantage lies in not relying on specific distribution assumptions of the data and being able to effectively process ecological data with innovative distributions. For example, when facing sudden and rare abnormal events in the ecological system, the isolation forest model accurately judges the degree of abnormality of data points through the path length of the data points in the tree. Compared with traditional methods based on fixed distribution assumptions, it can more accurately identify these anomalies.This combination gives full play to the advantages of the two methods, greatly improves the accuracy and robustness of anomaly detection, reduces misjudgments and missed judgments, makes the data processed by this application more reliable, and provides a pure data foundation for subsequent feature engineering and ecological analysis. 3. Traditional feature engineering methods are relatively simple, usually only extracting a few common statistical features, or only performing feature extraction in a single field (such as only the time domain or frequency domain). For example, in traditional ecological data analysis, only simple statistics such as mean and standard deviation are calculated, ignoring the high-order statistical characteristics of the data, quantile information, and characteristics at different time scales. In addition, traditional methods are also relatively rough in terms of feature combination and optimization. They often use simple feature splicing and lack in-depth exploration of innovative relationships between features. It is difficult to give full play to the advantages of multimodal data, resulting in limited description of ecological phenomena by the model. This scheme performs comprehensive and in-depth operations in the feature extraction stage. In statistical feature extraction, not only basic statistics are calculated, but also high-order statistics and rich quantile information are considered to comprehensively characterize the distribution characteristics of the data. Time-domain feature extraction combines autocorrelation analysis, sliding window techniques, and wavelet transforms to capture data dependencies, local dynamics, and characteristics at different time scales from different perspectives. Frequency-domain feature extraction uses FFT and the Welch method to estimate power spectral density, and further calculates band energy and high-order frequency-domain features to accurately analyze the energy distribution and morphology of the data in the frequency domain. This comprehensive feature extraction approach can uncover richer and more representative features, providing multi-faceted information for ecological analysis. Regarding feature screening, a deep learning-based adaptive threshold setting method uses a neural network to dynamically adjust the correlation threshold based on the predictive accuracy of the ecological detection model. Compared with traditional fixed threshold methods, this method more intelligently retains features that contribute most to ecological value prediction and effectively removes redundant features. In the feature construction stage, a deep embedding network and variational autoencoder are used for feature fusion and transformation, exploring the underlying structure and distribution patterns among features to generate compact and representative features. Generative adversarial networks are used to generate derived features, which are enhanced through reinforcement learning to further enrich the diversity and effectiveness of features. Finally, hierarchical feature combination and feature optimization based on genetic algorithms can fully utilize the advantages of different types of features and search for the optimal feature combination, so that the generated multimodal feature data can maximize the performance of the ecological value prediction model.

[0032] Optionally, the feature engineering of the multimodal purified data to generate multimodal feature data includes: performing statistical feature extraction, time domain feature extraction, and frequency domain feature extraction on the multimodal purified data to obtain a multimodal cross-domain feature set; performing correlation analysis and feature importance evaluation on the multimodal cross-domain feature set to screen out multimodal key features therefrom; and performing feature construction based on the multimodal key features to generate multimodal feature data.

[0033] Preferably, in a specific application scenario, the above solution is described in an alternative or preferred manner.

[0034] 1. Feature extraction of multimodal purified data

[0035] When extracting statistical features, basic statistics are calculated: for each modal data X in the multimodal purified data m (m=1,2,…,M, M represents the total number of modes. For example, when M=5, it can correspond to five different types of data modes: air quality, water quality, soil, meteorology, and biodiversity.) Calculate its mean μ m , the formula is where x m,i Represents the i-th data point in the m-th modal data, N m is the total number of data points for the mth modal data. The mean reflects the average level of the modal data. For example, in air quality data, the mean can represent the average status of the air quality index over a period of time.

[0036] Calculate the median m , firstly transform the data x m,i Sort in ascending order, if N m If is an odd number, If N m If is an even number, The median can reflect the middle position of the data, is not affected by extreme values, and has good robustness to the interference of outliers in ecological data. m , which is the value with the highest frequency in the data set. In actual calculations, we can count the number of times each value appears, and the value with the highest number of occurrences is the mode. The mode reflects the most common state in the data. For example, in the water quality data of a certain area, the mode represents the most common water quality condition in that area. Calculate the standard deviation σ m , the formula is The standard deviation measures the degree of dispersion of the data, that is, the fluctuation of the data around the mean. In ecological monitoring, the standard deviation can reflect the stability of ecological data. For example, a large standard deviation of air quality data indicates that the air quality fluctuates more violently. m , the formula is Skewness describes the asymmetry of the data distribution. Positive skewness indicates that the data has a longer tail on the right side (in the direction of larger values), while negative skewness indicates the opposite. In ecological data, skewness helps understand the shape of the data distribution, for example, some ecological indicators have more extreme values in a certain direction. Calculate Kurtosis m , the formula is Kurtosis is used to describe the peak shape of data distribution. Compared with the normal distribution, positive kurtosis indicates that the peak of the data is sharper, while negative kurtosis indicates that the peak of the data is flatter. The kurtosis analysis of ecological data helps to discover the differences between the data distribution and the normal distribution. For example, certain ecological phenomena may lead to abnormal peaks or flat distributions in the data. Higher-order statistic calculation: To more deeply characterize the data features, higher-order statistics are calculated. For example, the fifth central moment The sixth central moment etc. These higher-order statistics can capture more subtle distribution characteristics of the data. Although their physical meanings are not as intuitive as those of lower-order statistics, they can provide additional information dimensions when analyzing innovative ecological data, helping to reveal the hidden laws in the data. Quantile calculation: Considering the quantile information of the data, the values at different quantiles are calculated. For example, the quartiles Q 1,m 、Q 2,m (i.e., the median, which corroborates with the median calculated previously), Q 3,m . When calculating Q 1,m , first determine the position If i1 is an integer, then If i1 is not an integer, let i1 = k + f, where k is the integer part and f is the decimal part, then Q 1,m =(1 - f)x m,k +fx m,k+1 . Similarly, calculate Q 3,m , the position Also calculate the percentiles P 10,m 、P 25,m 、P 75,m 、P 90,m etc., and the calculation method is similar to that of quartiles. Quantile information can describe the distribution of data at different positions. For example, P 90,m means that 90% of the data is less than or equal to this value, which is of great significance for analyzing different degrees of changes in ecological data. Combine all the above statistical features into a statistical feature vector F stat,m =[μ m , Median m , Mode m , σ m , Skewness m , Kurtosis m , μ 5,m , μ 6,m , Q 1,m , Q 2,m , Q 3,m , P 10,m , P 25,m , P 75,m , P 90,m T , for all modalities m = 1, …, M, obtain the statistical feature set​

[0037] When extracting time-domain features, autocorrelation analysis: For the ecological data sequence of each mode (T m represents the length of the time series of the m-th mode data), the autocorrelation function is used to analyze the dependence relationship of the data in the time series. The autocorrelation function is defined as where τ is the time delay, and the value range is 0 ≤ τ ≤ L m (L m is the preset maximum delay value, which is determined according to the data characteristics and analysis requirements. For example, for ecological data with obvious seasonal changes, L m can be set to the number of time steps corresponding to one year). By calculating the autocorrelation function values R m (τ) at different delay τ values, the autocorrelation sequence is obtained. It reflects the similarity degree of the data at different time intervals and reveals the periodic and trend characteristics of the data changing with time. For example, if R m (τ) peaks at τ = 24 (assuming the time step is in hours), it implies that the ecological data has a daily variation pattern with a period of 24 hours. Sliding window feature calculation: Based on the results of autocorrelation analysis, the sliding window technique is adopted to capture the dynamic change characteristics of the data within a local time range. Set the size of the sliding window to w m , and the step size to s m (s m ≤ w m ). In each sliding window, a series of statistics are calculated. The mean within the window The standard deviation within the window The peak within the window The valley within the window In addition, calculate the slope of the data within the window. Let the data within the window be Fit a straight line y = a m (t)x + b m (t) through linear regression method, so that is minimized, and the obtained slope a m (t) is used as the trend feature within the window. Combine these features calculated at different sliding window positions t into a sliding window time-domain feature matrix F sliding-window,m , where each row corresponds to a feature vector at a window position, and each column corresponds to a feature type (such as mean, standard deviation, peak, etc.). Specifically, F sliding-window，m = [μ w，m (1), σ w，m (1), Peak w，m (1), Valley w，m (1), a m(1); μ w，m (2), σ w，m (2), Peak w,m (2), valley w，m (2), a m (2); …; μ w，m (n m ), σ w，m (n m ), Peak w，m (n m ), Valley w，m (n m ), a m (n m ), where is the total number of sliding windows. Wavelet transform feature extraction: Perform wavelet transform on the time-domain data to obtain richer time-domain and frequency-domain features. Select an appropriate wavelet basis function ψ m (t) (such as Daubechies wavelet, Haar wavelet, etc., selected according to the data characteristics. For example, for ecological data with mutation characteristics, the Daubechies wavelet is more suitable), and perform continuous wavelet transform on the ecological data sequence : where a is the scale parameter that controls the dilation of the wavelet function; b is the translation parameter that controls the translation of the wavelet function; is the complex conjugate of the wavelet basis function ψ m . By calculating at different scales a and displacements b, the wavelet coefficient matrix W m is obtained. Extract some key features from the wavelet coefficient matrix W m , such as the wavelet coefficient energy the entropy of the wavelet coefficients , etc. These features further enrich the time-domain feature representation. For example, the wavelet coefficient energy E a,m reflects the energy distribution of the data at different time scales and helps to discover transient changes and hidden periodic components in the data. Combine the autocorrelation feature sequence the sliding window time-domain feature matrix F sliding-window,m and the features extracted from the wavelet transform into the time-domain feature vector F time,m . The specific combination method is to flatten the autocorrelation feature sequence into a one-dimensional vector, and then concatenate it with the column-wise concatenation result of the sliding window time-domain feature matrix and the wavelet transform feature vector, that is where vec(·) represents the operation of expanding the matrix into a one-dimensional vector by columns.

[0038] When extracting frequency-domain features, for each mode of the ecological data sequence Perform a fast Fourier transform to convert the time-domain signal into a frequency-domain signal to reveal the energy distribution of the data at different frequencies. where f is the frequency, The spectrum X(f) is efficiently calculated by the FFT algorithm, which is a complex sequence. The magnitude |X(f)| represents the energy magnitude of the signal at frequency f, and the phase ∠X(f) represents the phase information of the signal at that frequency. For example, in meteorological data, the energy distribution at different frequencies corresponds to meteorological changes with different periods. The high-frequency part corresponds to short-term weather fluctuations, and the low-frequency part corresponds to seasonal climate changes. Power spectral density estimation: Based on the spectrum X(f) obtained by the FFT, the Welch method is used to estimate the power spectral density (PSD) to more accurately describe the distribution of signal power in the frequency domain. The data sequence m (f), it is a complex sequence, and its magnitude |X m (f)| represents the energy magnitude of the signal at frequency f, and the phase ∠X m (f) represents the phase information of the signal at that frequency. For example, in meteorological data, the energy distribution at different frequencies corresponds to meteorological changes with different periods. The high-frequency part corresponds to short-term weather fluctuations, and the low-frequency part corresponds to seasonal climate changes. Power spectral density estimation: Based on the spectrum X m (f) obtained by the FFT, the Welch method is used to estimate the power spectral density (PSD) to more accurately describe the distribution of signal power in the frequency domain. Divide the data sequence into K m overlapping sub-segments, each with a length of N m (usually N m < T m ). Let the k-th sub-segment be Perform FFT on each sub-segment to obtain X m,k (f), and then calculate the power spectral density estimate: This power spectral density estimate is smoother than the estimate directly based on the square of the FFT magnitude and more accurately reflects the distribution of signal power in the frequency domain, which helps to analyze the contribution of different frequency components in ecological data to the overall signal energy. Band energy feature extraction and high-order frequency domain features: According to the characteristics and analysis requirements of ecological data, different frequency bands (S m represents the number of frequency bands divided for the m-th mode data) are divided. Calculate the energy within each frequency band: Approximate the integral value by a numerical integration method (such as the trapezoidal integration method). These band energy features E s,m (s = 1,..., S m ) reflect the energy distribution of ecological data in different frequency intervals and are crucial for understanding the relationship between ecological phenomena and frequencies. In addition, calculate high-order frequency domain features, such as frequency domain skewness frequency domain kurtosis where μ f,m and σ f,m are the mean and standard deviation of the power spectral density estimate respectively, and F mis the number of frequency samples. These high-order frequency domain features can further characterize the form of power spectrum density distribution and provide more in-depth information for frequency domain analysis of ecological data. The frequency band energy features and high-order frequency domain features are combined into a frequency domain feature vector

[0039] Multimodal cross-domain feature set construction: The statistical feature vector F obtained by extracting statistical features from each modal data is stat,m , the time domain feature vector F obtained by time domain feature extraction time,m And frequency domain feature extraction to obtain multimodal cross-domain feature set construction: the statistical feature vector F obtained by statistical feature extraction of each modal data stat,m , the time domain feature vector F obtained by time domain feature extraction time,m And the frequency domain feature vector F obtained by frequency domain feature extraction freq,m Splice to form a multimodal cross-domain feature vector F m That is In this application, T represents the transposition operation of the vector, which converts the row vector into a column vector for concatenation. For M types of modal data, the multimodal cross-domain feature set The construction is thus completed. This multimodal cross-domain feature set integrates rich feature information from different modal data in the statistical, time domain and frequency domain, comprehensively depicts the characteristics of ecological data in different aspects, and provides a multi-dimensional data foundation for the subsequent in-depth analysis of the ecological value of ecological products. For example, when analyzing a forest ecological product, the multimodal cross-domain feature set composed of the multimodal cross-domain feature vector F1 of the air quality modal data and the F2 of the biodiversity modal data is It can reflect the innovative relationship and dynamic changes between forest ecosystems and the surrounding environment from multiple perspectives.

[0040] 2. Correlation Analysis and Feature Importance Assessment of Multimodal Cross-Domain Feature Sets

[0041] Correlation analysis: Construct a feature correlation matrix C with a dimension of D×D, where That is, the total length of the multimodal cross-domain feature vectors of all modes after splicing. ij Represents the correlation between the i-th feature and the j-th feature, calculated by the Pearson correlation coefficient: Among them F ik and F jk are the values of the i-th and j-th features on all samples k=1,…,N (in this application, N is the total number of samples, for example, the number of samples obtained by multiple monitoring of ecological products over a period of time), and are their means respectively. The Pearson correlation coefficient measures the degree of linear correlation between two variables, C ijThe closer the value is to 1 or -1, the stronger the linear correlation between the two features; the closer it is to 0, the weaker the linear correlation. In this application, to analyze the correlation more intuitively, the correlation matrix C is visualized, such as by drawing a heatmap. Through the heatmap, this application can clearly see the strength distribution of the correlation between different features, helping researchers quickly identify highly correlated feature clusters. Feature importance evaluation: An ensemble learning model is used to evaluate the importance of each feature. This application selects Random Forest, Gradient Boosting Tree, and Adaptive Fusion Ensemble Model (AFEM). In this application, for the Random Forest model, it consists of n rf decision trees . During the construction of each decision tree, a part of the samples is randomly drawn from the training samples with replacement (bootstrap sampling). At the same time, when splitting each node, a part of the features is randomly selected from all the features to determine the best splitting condition. For each feature f, its importance is evaluated by calculating the degree of contribution of this feature to the node splitting in all decision trees. Specifically, for each tree Calculate the sum of the information gains of feature f on all nodes in the tree Then sum and average the information gains of all trees to obtain the importance score of feature f under the Random Forest model The Gradient Boosting Tree model gradually improves the model performance by iteratively training a series of weak learners (usually decision trees). In each iteration, the new weak learner fits the residual between the prediction results of all previous weak learners and the true values. For each feature f, its importance is evaluated according to its contribution to reducing the residual in each iteration. Suppose a total of n gb iterations are performed. In the t-th iteration, calculate the information gain of feature f for the current weak learner (decision tree ) Then sum the information gains of all iterations to obtain the importance score of feature f under the Gradient Boosting Tree model In this application, the Adaptive Fusion Ensemble Model (AFEM) can dynamically adjust the weights of each sub-model according to the performance of different features on different sample subsets. First, the training samples are divided into multiple non-overlapping subsets (n af is the number of subsets). For each subset S i , multiple base models (such as decision trees, linear regression, etc.) are trained respectively to form a model ensemble (n b is the number of base models trained on each subset). During prediction, for each sample x, according to the subset S it belongs toj , calculate each base model The predicted value of Then, through an adaptive weight function The weighted sum of these predicted values is used to obtain the final predicted value Weight function According to each basic model in the subset S j Dynamically adjust the historical prediction performance based on the dataset S, for example, by calculating the performance of each model in the subset S j The weight is determined by the prediction accuracy, mean square error and other indicators on the dataset. For each feature f, the importance is evaluated by analyzing its influence on each basic model on different subsets. Specifically, the feature f is calculated in each subset S i For each base model The feature importance score of (For example, using a method similar to that used in random forests or gradient boosting trees to calculate information gain), and then comprehensively considering all subsets and basic models to obtain the importance score of feature f under the adaptive fusion integration model where α i,k According to the subset S i and the base model The weight coefficient is determined by the importance of the feature, and is optimized by cross-validation and other methods. Finally, the feature importance scores obtained by the random forest, gradient boosting tree and adaptive fusion integration model are combined to calculate the comprehensive importance score of each feature S(f) = β rf S rf (f)+β gb S gb (f)+β af S af (f), where β rf , β gb and β af It is the weight coefficient determined by cross-validation based on the model performance, satisfying β rf +β gb +β af =1.

[0042] 3. Feature construction based on multimodal key features

[0043] Feature Fusion and Transformation: Using Deep Embedding Networks The selected multimodal key feature sets are fused and transformed to explore the potential relationships between features and map them into a unified feature space. There are multiple hidden layers. Let the weight matrix from the input layer to the first hidden layer be W1, the bias vector be b1, and the weight matrix and bias vector between the hidden layers be W l , b l(l = 2, …, L, where L is the total number of hidden layers), the weight matrix of the output layer is W out , and the bias vector is b out . For the input multi-modal key feature vector F key (which is a vector composed of the key features selected from the multi-modal cross-domain feature set through correlation analysis and feature importance evaluation), after being transformed by the network: h1 = σ(W1F key + b1), h l = σ(W l h l-1 + b l ), l = 2, …, L, F embedded = W out h L + b out , where σ is the activation function. In this application, the Leaky ReLU function is selected, and its definition is α is a small positive number (such as α = 0.01). Compared with the traditional ReLU function, the Leaky ReLU function also has a non-zero output when z < 0, which helps to solve the gradient vanishing problem and enables the network to learn better. By minimizing the loss function to train the deep embedding network, where and are the embedded feature vector and the input key feature vector of the i-th sample respectively, and Target i is the set embedded result of the i-th sample in the unified feature space (which can be obtained through prior knowledge or data distribution estimation, for example, using methods such as principal component analysis to reduce the dimension of the data and taking the reduced result as a reference for the set embedding). In this application, for the embedded features, a non-linear transformation method based on the variational autoencoder (VAE) is adopted. Let the encoder of the VAE encode F embedded into the latent variable z = Encoder(F embedded ), where z follows a Gaussian distribution μ and σ are the mean and standard deviation vectors output by the encoder. The specific implementation of the encoder is a multi-layer neural network, and its outputs are μ and log(σ 2 ) respectively (to ensure the non-negativity of σ, usually log(σ 2 ) is output, and then σ is obtained through exponential operation). In this application, the decoder decodes the latent variable into the transformed feature through F transformed = Decoder(z). The decoder is also a multi-layer neural network. In this application, the VAE is trained by minimizing the variational lower bound loss function , where is the KL divergence, which measures the Gaussian distribution output by the encoder and the standard normal distribution The differences prompt the latent variables to learn a meaningful distribution; is the reconstruction loss, hoping that the decoder can reconstruct the original embedded features as accurately as possible according to the latent variables. After transformation, a set of compact latent representation features is obtained

[0044] Feature derivation and enhancement: Using the generator G of the generative adversarial network (GAN), based on the set of compact latent representation features generate new derived features. Let the input of the generator be the noise vector n, which follows the standard normal distribution (I is the identity matrix), and the generated derived feature is F derived = G(n, F compact ). The generator G is a multi-layer neural network that takes the noise vector and the compact latent representation features as inputs and outputs derived features with a similar dimension to the compact latent representation features. The discriminator D is used to distinguish the generated derived features from the real compact latent representation features, and the GAN is trained by minimizing the adversarial loss function . The discriminator D is also a multi-layer neural network that takes a feature vector as input and outputs a scalar value representing the probability that the feature vector is a real feature. During the training process, the generator and the discriminator conduct an adversarial game. The generator tries to generate more realistic derived features to deceive the discriminator, while the discriminator tries to more accurately identify the generated features, and finally reaches a balanced state, making the generated derived features of high quality. In this application, at the same time, a reinforcement learning algorithm is used to enhance the generated derived features. Define the agent A of reinforcement learning, whose state is the current derived feature F derived , and the action space is a series of transformation operations on the feature (such as linear and non-linear transformations like scaling, translation, rotation, etc.). The agent interacts with the environment (i.e., the ecological value prediction model) and learns the optimal policy according to the reward function R = Accuracy(Model(F derived )) - Accuracy(Model(F compact ). Among them, Accuracy(Model(·)) represents the prediction accuracy of the ecological value prediction model when inputting the corresponding features. This reward function measures the improvement degree of the prediction accuracy of the ecological value prediction model when using the generated derived feature F derived compared with using the compact latent representation feature F compact . In this application, at each time step t, the agent selects an action a from the action space according to the current state t . The action a t acts on the current derived feature to generate a new derived feature Among them, the ApplyAction function represents applying action a t to the feature . Then, the agent inputs the new derived feature into the ecological value prediction model Model to calculate the reward The goal of the agent is to maximize the long-term cumulative reward by continuously trying different actions where γ is the discount factor, and its value range is [0, 1]. It determines the importance of future rewards relative to the current reward. A smaller γ indicates that the agent pays more attention to immediate rewards, while a larger γ makes the agent focus more on long-term rewards. In this application, in order to learn the optimal policy, the agent uses reinforcement learning algorithms such as the Deep Q-Network (DQN). In DQN, the agent maintains a Q-network Q(s, a; θ), which estimates the expected long-term cumulative reward for taking action a in state s, and θ is the parameter of the Q-network. At each time step, the agent selects an action according to the ε-greedy policy, that is, randomly selects an action with probability ε and selects the action that maximizes Q(s t , a; θ) with probability 1 - ε. This policy trades off exploring new actions (with probability ε) and exploiting the current optimal action (with probability 1 - ε), which helps the agent fully explore the action space at the beginning of learning to discover potential better policies and gradually focus on the currently considered optimal action as learning progresses. After collecting enough experience samples (s t , a t , R t+1 , s t+1 ), the agent stores these samples in the experience replay buffer . The experience replay buffer allows the agent to randomly sample from past experiences, break the correlation between samples, and make the learning process more stable. Then, the agent updates the parameter θ of the Q-network by minimizing the loss function L(θ), and the loss function is defined as: where represents the expectation sampled from the experience replay buffer , and (s, a, R, s′) represent the state, action, reward, and next state respectively. θ - is the parameter of the target Q-network, which is periodically copied from the current Q-network and remains relatively stable for calculating the target Q-value R + γ max a′ Q(s′, a′; θ - ). This double Q-network structure (the current Q-network and the target Q-network) helps to stabilize the learning process and avoid over-optimism or instability in Q-value estimation. By continuously interacting with the environment, collecting experience samples, and updating the Q-network parameters, the agent gradually learns the optimal policy, that is, finds a series of actions that can make the generated derived feature Fderived Transformation operations to maximize the accuracy of the ecological value prediction model. Finally, the enhanced derivative feature F is generated enhanced .

[0045] When combining and optimizing features, based on domain knowledge and data characteristics, a hierarchical combination strategy is adopted for the compact latent representation feature F compact and the enhanced derivative feature F enhanced for combination. First, group the compact latent representation feature and the enhanced derivative feature according to the modality. For example, for data of different modalities such as air quality, water quality, and soil, process the corresponding compact latent representation feature and the enhanced derivative feature respectively. For each modality, concatenate the compact latent representation feature and the enhanced derivative feature hierarchically. Specifically, construct a multi-layer feature combination structure. In the first layer, simply concatenate and to obtain In the second layer, perform a non-linear transformation on the combined feature obtained in the first layer, for example, process it through a multi-layer perceptron (MLP). Let the input of the MLP be After the transformation of the hidden layer of the MLP: where σ is the activation function (such as the ReLU function σ(z) = max(0, z)), and are the weight matrix and bias vector of the l-th hidden layer respectively. Then concatenate the output of the last hidden layer with the combined feature of the first layer again to obtain Through this hierarchical combination method, the advantages of the compact latent representation feature and the enhanced derivative feature can be fully utilized, the deeper relationships between features can be mined, and a more hierarchical and expressive feature representation can be formed It helps the model better understand the innovative relationship between the ecological value of ecological products and each feature. Use the genetic algorithm to optimize the hierarchically combined features to find the optimal feature combination method and further improve the performance of the ecological value prediction model.

[0046] First, encode the feature combination. Treat each feature dimension as a gene, and encode the feature combination as a chromosome I m . For example, if is a D-dimensional feature vector, then the chromosome I m is an encoded vector of length D, and the value of each gene indicates whether the feature dimension is retained or undergoes a certain transformation (such as a scaling factor, etc.). Define the fitness function F(I m)=Accuracy(Model(Decode(I m ))), where Decode(I m ) function will encode the chromosome I m Decoded as the actual feature combination, Accuracy(Model(·)) is the prediction accuracy of the ecological value prediction model when the corresponding feature combination is input. The fitness function measures the contribution of each chromosome (i.e., feature combination) to the performance of the ecological value prediction model. In each generation of the genetic algorithm, the following operations are performed: Selection: Based on the value of the fitness function, a certain number of chromosomes are selected from the current population as parents using the tournament selection method. Specifically, k chromosomes are randomly selected from the population each time (the tournament size is k), and then the chromosome with the highest fitness is selected as the parent. This selection method can select chromosomes with better fitness with a higher probability, while also retaining a certain degree of randomness to avoid premature convergence to the local optimal solution. Crossover: Perform a crossover operation on the selected parent chromosome to generate a daughter chromosome. Using the multi-point crossover method, multiple crossover points are randomly selected, and the gene fragments of the parent chromosome on both sides of the crossover point are exchanged to generate a new feature combination. For example, suppose chromosome and There are two parent chromosomes, and the crossover points c1, c2, ..., c are randomly selected. n , then the offspring chromosomes and Generated by:

[0047] Mutation: with a certain mutation probability p m Perform mutation operations on the genes of each daughter chromosome. Mutation is to randomly change the value of a gene. For example, for the gene representing the feature dimension scaling factor, its value is randomly adjusted to introduce new genetic diversity to the population and prevent the algorithm from falling into local optimality. For example, for a gene g, with probability p m The mutation method is g=g+δ, where δ is a normal distribution. The random number, σ m Adjust according to the characteristics of the problem. After multiple generations of evolution, the genetic algorithm continuously searches the space of feature combinations and gradually finds the chromosome that maximizes the fitness function, that is, the optimal feature combination. The feature combinations of all modalities optimized by the genetic algorithm are combined again to generate the final multimodal feature data. These feature data can maximize the prediction accuracy and efficiency of the ecological value prediction model, providing strong support for the ecological value prediction of ecological products. When performing these operations, this application needs to efficiently process a large amount of mathematical calculations and data storage to achieve the conversion from raw ecological data to high-quality multimodal feature data.

[0048] Therefore, the above preferred or alternative technical solutions have the following technical advantages.

[0049] 1. Traditional feature extraction methods are usually rather single and simple. In terms of statistical feature extraction, only basic means, standard deviations, etc. are calculated, ignoring high-order statistics and quantile information, and unable to comprehensively capture the subtle features of data distribution. In time-domain feature extraction, only simple autocorrelation analysis is relied on, without making full use of the sliding window technique and wavelet transform, and insufficiently mining the local dynamic changes and features of different time scales of the data. In frequency-domain feature extraction, only the spectrum is obtained through simple Fourier transform, lacking accurate estimation of power spectral density and analysis of high-order frequency-domain features. This simple feature extraction method leads to the loss of a large amount of useful information and is difficult to accurately depict the innovative characteristics of ecological data. The feature extraction method of this solution has comprehensiveness and in-depthness. In statistical feature extraction, not only basic statistics are covered, but also high-order statistics and rich quantile information are deeply calculated, which can more comprehensively describe the central tendency, dispersion degree, distribution form of the data, and the value characteristics at different positions. For example, high-order statistics can reveal hidden characteristics such as asymmetry and heavy-tailedness in the data, and quantile information helps analyze the situation of ecological data under different degrees of change. Time-domain feature extraction combines multiple methods. Autocorrelation analysis reveals the time-dependent relationship of the data, the sliding window technique captures local dynamic changes, and wavelet transform obtains features of different time scales, comprehensively and meticulously depicting the change law of the data in the time series. Frequency-domain feature extraction, through accurate power spectral density estimation, band energy analysis, and high-order frequency-domain feature calculation, more accurately analyzes the energy distribution and form of the data in the frequency domain, helping to discover the innovative relationship between ecological phenomena and frequency. Combining these feature extraction methods can provide rich and multi-dimensional information for subsequent analysis, significantly improving the understanding and representation ability of ecological data. 2. Traditional correlation analysis only relies on the simple Pearson correlation coefficient, and in terms of feature importance assessment, often a single model (such as the simple linear regression coefficient or the feature importance of decision trees) is adopted, unable to comprehensively consider the non-linear relationships between features and the differences in feature importance assessment among different models. This single assessment method leads to misjudgment of feature importance, missing some features that are crucial for ecological value prediction but non-linearly correlated. This solution first constructs a comprehensive correlation matrix, uses the Pearson correlation coefficient to analyze the linear correlation between all features in detail, and further assists the analysis through visualization. In feature importance assessment, multiple ensemble learning models (random forest, gradient boosting tree, and adaptive fusion ensemble model) are used for comprehensive assessment. Each model evaluates the contribution of features to the model performance from different perspectives. The random forest evaluates the contribution of features to node splitting through the construction of decision trees; the gradient boosting tree evaluates the importance according to the role of features in reducing residuals during the iterative process; the adaptive fusion ensemble model evaluates the importance according to the influence of features on multiple basic models in different sample subsets. Finally, by synthesizing the evaluation results of these models, a more accurate and comprehensive feature importance score is obtained.This multi-model fusion evaluation approach fully considers innovative relationships between features, more accurately identifying features truly important for ecological value prediction, avoiding misselection and omission of features, and improving the accuracy and efficiency of subsequent analysis. 3. Traditional feature construction methods are relatively simple and straightforward, relying solely on simple feature concatenation or a small number of empirically based transformations. These methods fail to deeply explore the underlying relationships between features and struggle to generate features with strong representational capabilities. Furthermore, they lack intelligent optimization mechanisms for feature enhancement, making it difficult to dynamically adjust and optimize features based on actual application scenarios (such as ecological value prediction). This solution utilizes a deep embedding network and variational autoencoders for feature fusion and transformation. This approach exploits the underlying structure and distribution patterns between features, maps multimodal key features into a unified feature space, and performs nonlinear transformations using a variational autoencoder to generate compact and representative features. Generative adversarial networks are used to generate derived features, ensuring high quality and diversity through adversarial game-playing. A reinforcement learning algorithm enhances these derived features, intelligently searching for optimal feature transformation strategies to dynamically generate features most conducive to ecological value prediction, guided by improving the accuracy of the ecological value prediction model. The hierarchical combination strategy and genetic algorithm further optimize the feature combination, and explore the deep relationship between features through multi-layer splicing and nonlinear transformation. The genetic algorithm searches for the optimal solution in the feature combination space, so that the multimodal feature data finally generated can maximize the performance of the ecological value prediction model, significantly enhancing the accuracy and robustness of the model in predicting the ecological value of ecological products.

[0050] In a follow-up manner, the multimodal purified data is subjected to statistical feature extraction, time domain feature extraction, and frequency domain feature extraction to obtain a multimodal cross-domain feature set, including: calculating the basic statistics of the multimodal purified data to estimate the central tendency, degree of dispersion, and extreme characteristics of the multimodal purified data based on the basic statistics; performing empirical mode decomposition on the multimodal purified data to determine the dependency characteristics, local dynamic change characteristics, and characteristics of different time scales of the multimodal purified data in the time dimension; transforming the multimodal purified data into the frequency domain and performing power spectrum density statistics and frequency band feature extraction to determine the frequency domain energy distribution of the data and the frequency domain characterization characteristics of the ecological phenomenon. Preferably, in a specific application scenario, the above scheme is described in an alternative or preferred manner.

[0051] 1. Calculate basic statistics of multimodal purified data

[0052] Definition and calculation of basic statistics: For each modal data X in the multimodal purified data m(m = 1, 2, …, M, where M represents the total number of modalities. For example, in the ecological monitoring scenario, M corresponds to different monitoring modalities such as air quality, water quality, and soil quality), this application calculates a series of basic statistics. Mean: The mean is used to measure the central tendency of the data, and the formula is where x m,i represents the i-th data point in the data of the m-th modality, and N m is the total number of data points in the data of the m-th modality. In the air quality monitoring modality, the mean reflects the average level of air quality indicators (such as PM 2.5 concentration) over a period of time. Median: The median is also used to describe the central tendency, especially when there are extreme values in the data, it can better reflect the central position of the data. Arrange the data x m,i in ascending order. If N m is odd, then If N m is even, then For example, in water quality monitoring data, if there are individual extremely high or low pollution values, the median can more robustly represent the general level of water quality. Mode: The mode is the value that appears most frequently in the data set. In actual calculation, a frequency distribution table can be constructed to count the number of times each value appears, and the value with the most occurrences is the mode. The mode reflects the most common state in the data. For example, in the soil pH data of a certain area, the mode represents the most common soil pH condition in that area. Standard Deviation: The standard deviation measures the degree of dispersion of the data, and the formula is It represents the average deviation degree of the data points around the mean. The larger the standard deviation, the higher the degree of dispersion of the data. For example, in meteorological temperature data, a larger standard deviation means greater temperature fluctuations. Coefficient of Variation: To more accurately compare the degree of dispersion of data in different modalities (even if the means are different), calculate the coefficient of variation (when μ m ≠ 0). The coefficient of variation eliminates the influence of the mean on the measurement of the degree of dispersion and is very useful when comparing the stability of different ecological indicators (such as biodiversity indices and average species numbers in different regions). Skewness: Skewness describes the asymmetry of the data distribution, and the formula is Positive skewness means that the right side (the direction of larger values) of the data has a longer tail, while negative skewness means that the left side (the direction of smaller values) has a longer tail. In ecological data, due to special ecological processes, the distributions of some ecological indicators are asymmetric, and skewness helps this application understand this distribution characteristic. Kurtosis: Kurtosis is used to describe the peak shape of the data distribution, and the formula is Compared with the normal distribution, positive kurtosis indicates a sharper peak of the data, while negative kurtosis indicates a flatter peak of the data. Through kurtosis analysis, this application discovers the differences between the ecological data distribution and the normal distribution. For example, certain ecological phenomena lead to abnormal peaks or flat distributions in the data. InterQuartile Range (IQR): Calculate the quartiles Q 1,m and Q 2,m (i.e., the median), and Q 3,m . First, determine the position If i1 is an integer, then If i1 is not an integer, let i1 = k + f, where k is the integer part and f is the fractional part. Then Q 1,m =(1 - f)x m,k [[ID=E14]]+ fx m,k+1 . Similarly, calculate Q 3,m , and the position The interquartile range IQR m = Q 3,m - Q 1,m . It measures the dispersion of the middle 50% of the data, is not sensitive to extreme values, and can more robustly reflect the discrete characteristics of the data. In ecological data, IQR helps this application identify the main fluctuation range of the data and exclude the interference of extreme values. Extreme Value Ratio (EVR): To more directly measure the extreme characteristics of the data, define the extreme value ratio (when μ m ≠ 0). It reflects the ratio of the difference between the maximum and minimum values in the data relative to the mean. A larger EVR indicates a larger extreme value difference in the data, which is of great significance when analyzing extreme situations (such as sudden severe pollution events or rare ecological prosperity phenomena) in ecological data.

[0053] Comprehensive analysis and eigenvector construction: Combine the above basic statistics into the basic statistical feature vector F basic-stat,m =[μ m , Median m , Mode m , σ m , CV m , Skewness m , Kurtosis m , IQR m , EVR m T . For all M types of modal data, obtain the basic statistical feature set These basic statistics measure the central tendency, dispersion, and extreme characteristics of multi-modal purified data from different perspectives, providing a basic data feature description for subsequent analysis. ​

[0054] 2. Empirically decompose the multi-modal purified data

[0055] Principle of Empirical Mode Decomposition (EMD): Empirical Mode Decomposition is a method for decomposing innovative time series data into multiple Intrinsic Mode Functions (IMFs), which is applicable to non-linear and non-stationary data and is consistent with the characteristics of ecological data. For each ecological data sequence of each mode (T m (representing the length of the time series of the m-th mode data), the EMD process is as follows. Screening process: First, determine all local extreme points (maximum and minimum points) of the data sequence x m,t . Connect all the maximum and minimum points respectively through cubic spline interpolation to obtain the upper envelope e u,m (t) and the lower envelope e l,m (t). Calculate the average envelope Then subtract the average envelope from the original data to obtain h m (t) = x m,t - e m (t). IMF judgment condition: Judge whether h m (t) satisfies the two conditions of IMF: one is that within the entire data length, the number of extreme points and the number of zero-crossing points must be equal or differ by at most one; the other is that at any time, the mean of the upper envelope formed by local maximum points and the lower envelope formed by local minimum points is zero. If h m (t) does not meet these two conditions, then take h m (t) as the new x m,t , and repeat the above screening process until the satisfied h m (t) is obtained, denoted as IMF 1,m (t). Multiple decompositions: Subtract IMF m,t (t) from the original data x 1,m to obtain the remaining data r 1,m (t) = x m,t - IMF 1,m (t). Then take r 1,m (t) as the new original data and repeat the above screening process to obtain the second IMF, that is, IMF 2,m (t), and so on, until the remaining data r n,m (t) becomes a monotonic function and no more IMFs can be extracted. Finally, the original data x m,t is expressed as

[0056] When performing feature extraction and analysis, time-dependent features: By analyzing the phase information of each IMF i,m (t), the time-dependent features of the data in the time dimension are obtained. Define the phase function where H[IMF i,m (t)] is the Hilbert transform of IMF i,m (t). The change in phase reflects the relative relationship between data at different time points. For example, periodic changes in phase imply periodic dependencies in ecological phenomena, such as circadian rhythms or seasonal changes. Local dynamic change features: The amplitude change of each IMF i,m (t) reflects the local dynamic change features of the data. Calculate the instantaneous amplitude of IMF i,m (t) By observing the changes in a i,m (t) at different time points, understand the dynamic changes of ecological data within a local time range. For example, rapid fluctuations or slow changes in ecological indicators within a short period. Different time-scale features: Different IMFs i,m (t) correspond to different time scales. High-frequency IMF components reflect short-term fluctuations in the data, while low-frequency IMF components reflect long-term trends in the data. By performing frequency analysis on each IMF i,m (t) (e.g., calculating its spectrum through Fourier transform), determine the characteristics of ecological data at different time scales. For example, high-frequency IMFs correspond to short-term disturbances or rapid change processes in the ecosystem, and low-frequency IMFs are related to the long-term evolution or seasonal changes of the ecosystem. Feature vector construction: Integrate the phase function values, instantaneous amplitudes, and frequency information of each IMF i,m (t). For each IMF i,m (t), calculate the mean phase mean instantaneous amplitude and the dominant frequency f i,m (determined by spectrum analysis as the frequency at which the energy is maximum) over the entire time series. Then construct the empirical mode decomposition feature vector(for all M types of modal data, obtain the empirical mode decomposition feature set These feature vectors comprehensively characterize the time-dependent features, local dynamic change features, and different time-scale features of the multi-modal purified data, providing rich information for in-depth understanding of the time characteristics of ecological data.

[0057] 3. Transform the multi-modal purified data into the frequency domain and perform power spectral density statistics and frequency band feature extraction

[0058] For each ecological data sequence of each mode perform a fast Fourier transform to convert the time-domain signal into a frequency-domain signal to reveal the energy distribution of the data at different frequencies. where \(f\) is the frequency, the spectrum \(X(f)\) is efficiently calculated by the FFT algorithm, m (f), which is a complex sequence, and its magnitude \(|X\) m (f)|\) represents the energy magnitude of the signal at frequency \(f\), and the phase \(\angle X\) m (f)\) represents the phase information of the signal at that frequency. In ecological monitoring, the energy distribution at different frequencies corresponds to ecological changes with different periods. For example, the high-frequency part corresponds to short-term ecological fluctuations, and the low-frequency part corresponds to long-term ecological trends or seasonal changes. Power Spectral Density (PSD) estimation: Based on the spectrum \(X\) m (f) obtained by FFT, the Welch method is used for power spectral density estimation to more accurately describe the distribution of signal power in the frequency domain. The data sequence is divided into \(K\) m overlapping sub-segments, each with a length of \(N\) m (usually \(N\) m )]]\(< T\) m ), and let the \(k\)-th sub-segment be Perform FFT on each sub-segment to obtain \(X\) m,k (f), and then calculate the power spectral density estimate value: This power spectral density estimate is smoother than the estimate directly based on the square of the FFT magnitude and more accurately reflects the distribution of signal power in the frequency domain, which helps to analyze the contribution of different frequency components in ecological data to the overall signal energy.

[0059] Frequency band feature extraction: According to the characteristics and analysis requirements of ecological data, different frequency bands (\(S\) m represents the number of frequency bands divided for the \(m\)-th mode data) are divided. Calculate the energy within each frequency band: The integral value is approximately calculated by a numerical integration method (such as the trapezoidal integration method). These frequency band energy features \(E\) s,m (\(s = 1,\cdots,S\) m ) reflect the energy distribution of ecological data within different frequency intervals and are crucial for understanding the relationship between ecological phenomena and frequency.

[0060] Calculation of high-order frequency domain features: In addition to frequency band energy features, high-order frequency domain features are also calculated, such as frequency domain skewness frequency domain kurtosis where \(\mu\) f,m and \(\sigma\) f,m are the mean and standard deviation of the power spectral density estimate value respectively, and \(F\) mis the number of frequency samples. These high-order frequency domain features can further characterize the shape of the power spectral density distribution and provide more in-depth information for the frequency domain analysis of ecological data.

[0061] Feature vector construction: Combine the band energy features and high-order frequency domain features into a frequency domain feature vector For all M modal data, the frequency domain feature set is obtained These eigenvectors determine the frequency domain energy distribution of the multimodal purified data and the frequency domain representation characteristics of the ecological phenomena, providing key information for understanding ecological data from a frequency domain perspective. Construction of multimodal cross-domain feature set: The basic statistical feature vector F obtained by extracting the above statistical features of each modal data is basic-stat,m , the empirical mode decomposition eigenvector F obtained by empirical mode decomposition EMD,m And the frequency domain feature vector F obtained by frequency domain feature extraction freq,m Splice to form a multimodal cross-domain feature vector F m .Right now In this application, T represents the transposition operation of the vector, which converts the row vector into a column vector for concatenation. For M types of modal data, the multimodal cross-domain feature set The construction is thus completed. This multimodal cross-domain feature set integrates rich feature information from different modal data in statistics, time domain and frequency domain, and comprehensively characterizes the characteristics of ecological data in different aspects. In this application, first of all, this application will calculate various feature vectors of each modal data. When calculating basic statistics, statistics such as mean and median are obtained by traversing the data points for summation and sorting. For empirical mode decomposition, this application will repeatedly perform operations such as determining extreme points, interpolating envelope calculations, and judging IMF conditions. In frequency domain analysis, this application will perform FFT transformation, power spectrum density estimation, and frequency band energy calculation. Finally, this application will splice different types of feature vectors of each mode into multimodal cross-domain feature vectors, and summarize them to obtain a multimodal cross-domain feature set. This feature set provides a comprehensive and in-depth feature representation for subsequent in-depth analysis of ecological data, such as pattern recognition of ecological phenomena and prediction of the ecological value of ecological products, which helps to explore potential innovative relationships and laws in ecological data. For example, in a large-scale ecological monitoring project, a multimodal cross-domain feature set constructed by combining multiple modal data such as air quality, water quality, and biodiversity helps researchers understand the operating mechanism of the ecosystem from multiple dimensions and provide strong support for ecological protection and management decisions.

[0062] To this end, the above preferred or alternative technical solutions have the following technical advantages.

[0063] 1. When calculating basic statistics using traditional methods, usually only simple statistics such as the common mean and standard deviation are calculated. For ecological data, this approach cannot comprehensively reflect the distribution characteristics of the data. For example, only focusing on the mean and standard deviation and ignoring the asymmetry (skewness), peak shape (kurtosis) of the data, as well as the influence of extreme values, results in inaccurate characterization of the central tendency, dispersion degree, and extreme characteristics of the data, losing a large amount of useful information. This solution not only covers common statistics but also calculates the coefficient of variation, interquartile range, extreme value ratio, etc. The coefficient of variation can effectively compare the dispersion degree among data with different means and is crucial for analyzing the stability of ecological indicators of different magnitudes, such as comparing the fluctuations of the number of species and biodiversity indices in different regions. The interquartile range is insensitive to extreme values and can robustly measure the dispersion degree of the middle part of the data. When ecological data is affected by occasional extreme events, it can more accurately reflect the fluctuation range of the main body of the data. The extreme value ratio directly measures the extreme characteristics of the data and helps to capture sudden extreme situations in ecological data, such as rare ecological disasters or prosperous events. These rich statistics comprehensively and meticulously estimate various characteristics of the multimodal purified data, providing a solid data foundation for subsequent analysis. 2. Traditional time-domain analysis techniques often rely on simple autocorrelation functions or statistical analysis with fixed windows and are difficult to handle the non-linear and non-stationary characteristics of ecological data. A simple autocorrelation function can only reflect the linear correlation of data at a fixed delay and cannot capture time-dependent relationships; fixed-window analysis cannot adaptively adjust the analysis scale according to the data characteristics and is insufficient in mining the dynamic change characteristics of ecological data at different time scales. Empirical mode decomposition, aiming at the innovative characteristics of ecological data, decomposes it into multiple intrinsic mode functions (IMFs). By analyzing the phase function of the IMFs to obtain time-dependent characteristics, it can reveal time relationships such as circadian rhythms and seasonal changes in ecological phenomena, which are difficult to achieve with traditional methods. The instantaneous amplitude change of the IMFs reflects local dynamic change characteristics, enabling this application to observe the rapid fluctuations or slow changes of ecological indicators in a short period, which cannot be accurately captured by fixed-window analysis. Different IMFs correspond to different time scales, from high-frequency short-term disturbances to low-frequency long-term evolutions, comprehensively characterizing the characteristics of ecological data at different time scales and providing a powerful tool for in-depth understanding of the dynamic processes of ecological systems. 3. Traditional frequency-domain analysis only obtains the frequency spectrum through simple Fourier transform, lacking accurate estimation of the power spectral density and analysis of higher-order frequency-domain characteristics. Simple frequency spectrum analysis cannot accurately describe the distribution of signal power in the frequency domain and is difficult to deeply analyze the contributions of different frequency components to ecological phenomena. At the same time, ignoring higher-order frequency-domain characteristics will miss important information about the distribution pattern of the power spectral density, limiting the comprehensive understanding of the frequency-domain characteristics of ecological data.This solution uses the Welch method for power spectral density estimation. Compared with simple spectral analysis, it can more smoothly and accurately reflect the distribution of signal power in the frequency domain, which helps to accurately analyze the energy contributions of different frequency components of ecological data. For example, it can clarify the energy-level manifestations of ecological changes with different periods. By dividing frequency bands and calculating the energy characteristics of the frequency bands, it is possible to specifically analyze the energy distribution of ecological data within different frequency ranges, providing intuitive and crucial information for understanding the relationship between ecological phenomena and frequencies. In addition, calculating high-order frequency-domain characteristics such as frequency-domain skewness and kurtosis further characterizes the shape of the power spectral density distribution, uncovering hidden information that cannot be obtained by traditional methods and providing a richer perspective for in-depth analysis of ecological phenomena from the frequency-domain angle. 4. When dealing with multimodal data, traditional methods simply splice a small number of features from different modalities, failing to fully exploit the deep features of each modality of data in different domains (statistical, time-domain, frequency-domain), nor considering the mutual relationships between features. This simple processing method cannot effectively integrate the advantages of multimodal data, resulting in insufficient overall representation ability of ecological data and being difficult to meet the requirements of innovative ecological analysis tasks. This solution obtains rich feature vectors of each modality of data in different domains through comprehensive basic statistic calculation, empirical mode decomposition, and frequency-domain feature extraction, and cleverly splices them into a multimodal cross-domain feature set. This construction method fully integrates the information at different levels of multimodal data, comprehensively characterizing the innovative characteristics of ecological data. The multimodal cross-domain feature set provides a multi-dimensional and deep-level feature representation for subsequent ecological analysis, helping to uncover potential innovative relationships and laws in ecological data, greatly improving the accuracy and reliability of tasks such as ecological phenomenon pattern recognition and ecological value prediction of ecological products, and providing more powerful support for ecological protection and management decision-making.

[0064] Optionally, perform correlation analysis and feature importance evaluation on the multimodal cross-domain feature set to screen out multimodal key features, including: calculating the pairwise correlation of each feature in the multimodal cross-domain feature set based on the constructed correlation matrix to determine the associated eigenvalue between features; estimating the contribution degree of each feature in the multimodal cross-domain feature set to the prediction result to determine the importance score of each feature; calculating the information gain of each feature for the ecological value prediction of ecological products based on the associated eigenvalue between features; based on the information gain of each feature for the ecological value prediction of ecological products, statistically analyze the fluctuation trend of the corresponding importance score, and screen out multimodal key features from the features of the multimodal cross-domain feature set based on the fluctuation trend. Preferably, in a specific application scenario, illustrate the above solution in an alternative or preferred manner.

[0065] 1. Calculate the pairwise correlation of each feature in the multimodal cross-domain feature set based on the constructed correlation matrix

[0066] Set the multi-modal cross-domain feature set where F m is the multi-modal cross-domain feature vector of the m-th type of modal data. Concatenate the feature vectors of all modalities in order to form a large feature vector whose dimension is Construct a D×D correlation matrix C, and the matrix element C ij represents the correlation between the i-th feature and the j-th feature.

[0067] Correlation calculation: Use an improved partial correlation analysis method to calculate C ij , and the formula is: where, f i and f j are the i-th and j-th eigenvalues in the feature vector F respectively, Cov(f a , f b ) represents the covariance of the features f a and f b , and the calculation formula is N is the number of samples, f a,n and f b,n are the values of the features f a and f b in the n-th sample respectively, and are the means of the features f a and f b respectively. Var(f k ) represents the variance of the feature f k , and the calculation formula is In the ecological data scenario, through this correlation calculation, the associated eigenvalue between each feature can be more accurately determined. For example, when analyzing a forest ecosystem, the true correlation between features such as tree density and soil nutrient can be accurately judged, avoiding problems such as spurious correlation or omission of potential correlation relationships in simple correlation analysis.

[0068] 2. Estimate the contribution degree of each feature in the multi-modal cross-domain feature set to the prediction result to determine the importance score of each feature

[0069] To comprehensively evaluate the contribution degree of each feature to the ecological value of ecological products, a method integrating multiple innovative models is adopted. This application combines three models: Random Forest (RF), Gradient Boosting Regression Tree (GBRT), and Deep Neural Network (DNN). The random forest consists of T RF decision trees composed. For each feature f i , its importance is evaluated by calculating its contribution degree to node splitting in all decision trees. When constructing each decision tree When, a part of the samples are randomly drawn from the training samples with replacement (bootstrap sampling). At the same time, when each node is split, a part of the features are randomly selected from all the features to determine the best splitting condition. Feature f i In the decision tree The importance score is calculated based on the Gini impurity. Let the sample set of node n be S n , and the class set be Gini impurity where |S n,y | is the number of samples belonging to class y in node n. When splitting node n with feature f i , the reduction in Gini impurity after splitting, ΔGini n (f i ) is the contribution of feature f i to the splitting of this node. The importance score of feature f i in the decision tree is where |S| is the total number of training samples. Then, under the random forest model, the importance score of feature f i is Gradient Boosting Regression Tree Model: The Gradient Boosting Regression Tree gradually improves the model performance by iteratively training a series of weak learners (usually decision trees). In each iteration, the new weak learner fits the residuals between the predicted results of all previous weak learners and the true values. Let a total of T GBRT iterations be performed. In the t-th iteration, the importance score of feature f i for the current weak learner (decision tree ) is also calculated based on the reduction in Gini impurity, similar to the random forest. Finally, under the Gradient Boosting Regression Tree model, the importance score of feature f i is Deep Neural Network Model: Construct a deep neural network with multiple hidden layers. Let the input layer have D neurons corresponding to D features, and the hidden layers have H1, H2, …, H L neurons respectively, and the output layer has 1 neuron for predicting the ecological value of ecological products.

[0070] The forward propagation process of the network is: h1 = σ1(W1F + b1)h l = σ l (W l h l-1 + b l ), l = 2, …, L where F is the input feature vector, W1, W l , W outis the weight matrix, b1, b l , b out is the bias vector, σ1, σ l is the activation function (e.g., the ReLU function σ(z) = max(0, z)). By calculating the influence of the small change of the feature f i on the predicted value to evaluate its importance. Specifically, add a small perturbation ∈ to the feature f i to obtain F′, whose i-th element is f i +∈, and the other elements remain unchanged. Calculate the predicted value Then the importance score of the feature f i under the deep neural network model Comprehensive importance score: Combine the importance scores obtained from the three models, and calculate the comprehensive importance score S(f i ) = ω RF S RF (f i ) + ω GBRT S GBRT (f i ) + ω DNN S DNN (f i ), where ω RF , ω GBRT , ω DNN are weight coefficients determined by cross-validation according to the performance of the model on the validation set, and satisfy ω RF + ω GBRT + ω DNN = 1. In this way, by fusing multiple models, the contribution degree of each feature to the predicted ecological value of ecological products can be estimated more comprehensively and accurately, and a more reliable importance score can be obtained. For example, when predicting the ecological value of wetland ecological products, different models evaluate the contributions of features such as water level change characteristics and biodiversity index characteristics to the final ecological value from different angles, and the comprehensive score can more accurately reflect the importance of these features.

[0071] 3. Calculate the information gain of each feature for the prediction of the ecological value of ecological products according to the correlation eigenvalue between features

[0072] Information gain calculation: Information gain is used to measure the usefulness of a feature for a classification or prediction task. Let X represent the category of the ecological value of ecological products (or the range division of the predicted value), and f i be the i-th feature. First, calculate the information entropy H(X) of X, and the formula is where is the set of all values of X, and p(x) is the probability that X takes the value of x, which is obtained through sample statistics |S x| is the number of samples with the ecological value of ecological products being x. Then, for feature f i , assume it has V different values {v1, v2, …, v V}, and calculate the conditional entropy of X under the condition that feature f i takes the value of v j : where p(x|f i = v j ) is the conditional probability that the ecological value of ecological products takes the value of x when feature f i takes the value of v j , which is obtained through sample statistics is the number of samples where feature f i takes the value of v j and the ecological value of ecological products takes the value of x, is the number of samples where feature f i takes the value of v j . Then the information gain IG(f i ) of feature f i for predicting the ecological value of ecological products is:

[0073] In the actual prediction scenario of the ecological value of ecological products, for example, predicting the carbon sink value category of the ecological products of a forest, by calculating the information gain of each feature (such as the tree species diversity feature, forest area feature, etc.), understand the contribution degree of each feature to accurately predicting the carbon sink value category. The greater the information gain, the greater the help of the feature to the prediction.

[0074] 4. Based on the information gain of each feature for predicting the ecological value of ecological products, statistically analyze the fluctuation trend of the corresponding importance scores, and select multi-modal key features from the features of the multi-modal cross-domain feature set based on the fluctuation trend

[0075] Fluctuation trend statistics: Sort all features in descending order of information gain to obtain the feature sequence {f (1) , f (2) , …, f (D)}, and at the same time record the comprehensive importance score S(f (k) ) corresponding to each feature, k = 1, …, D. Calculate the difference sequence ΔS k = S(f (k+1) ) - S(f (k) ), k = 1, …, D - 1. This application uses the locally weighted regression scatterplot smoothing method (LOWESS) to smooth the difference sequence {ΔS k} to obtain the smoothed difference sequence The LOWESS method fits the data by performing weighted linear regression within a local neighborhood around each point k. The weight function usually uses a Gaussian kernel function where d is the distance between the point and the center point, and h is the bandwidth parameter, which is determined by cross-validation. Key feature screening: Analyze the fluctuation trend of the smoothed difference sequence When suddenly drops from a large value and remains at a low level, it is considered that the features before this position have high importance and good stability for the ecological value prediction of ecological products. These features are screened as multimodal key features. Specifically, a threshold τ (determined by experiments on the validation set) is set. When and several subsequent (l = 1, …, L, where L is the set window length, also determined by experiments) are all less than 0, the feature f (k) and the features before it are determined as multimodal key features.

[0076] Therefore, the above preferred or alternative technical solutions have the following technical advantages.

[0077] 1. Traditional correlation analysis often relies on the simple Pearson correlation coefficient, which can only measure the linear relationship between two variables. In scenarios such as ecological data, the relationships between features are often non-linear. The simple Pearson correlation coefficient will miss a large amount of important correlation information, leading to misjudgment of the true relationships between features. For example, there are causal or synergistic relationships between some ecological factors, but due to the non-simple linear correlation, traditional methods cannot accurately capture them. This application uses an improved partial correlation analysis method to calculate the elements of the correlation matrix. When calculating the correlation between two features, this method takes into account the influence of other features, can eliminate the interference of other variables, and more accurately reveal the true associations between features. In ecological scenarios, such as when studying forest ecosystems, multiple ecological variables interact with each other. The improved partial correlation analysis accurately judges the true correlation between tree density and soil nutrients after excluding other factors (such as climate conditions, precipitation, etc.), avoids spurious correlations, discovers potential relationships, and provides a more reliable basis for subsequent feature screening and model construction. 2. Traditional feature importance assessment usually relies on a single model, such as only using the feature importance of decision trees or linear regression coefficients to evaluate. The limitation of a single model is that its assumptions and applicable scenarios are limited, and it cannot comprehensively consider the innovative and variable characteristics of ecological data. For example, decision trees have limited ability to capture non-linear relationships in data, and linear regression assumes a linear relationship between features and the target variable. In the prediction of the ecological value of ecological products, the evaluation results of such a single model are inaccurate, missing important features or overestimating the role of some unimportant features. This application combines three models, random forest, gradient boosting regression tree, and deep neural network, to evaluate feature importance. Random forest calculates feature importance based on the node splitting of decision trees, can handle non-linear relationships and has good robustness to noise; gradient boosting regression tree emphasizes learning from difficult samples by iteratively fitting residuals and can effectively capture functional relationships; deep neural network has a powerful non-linear mapping ability and can learn high-order relationships between features. By integrating the evaluation results of these three models and determining the weight coefficients according to the performance of the models on the validation set, the comprehensive importance score obtained can more comprehensively and accurately reflect the contribution of features to prediction. When predicting the ecological value of wetland ecological products, different models evaluate features from different perspectives, and the comprehensive score avoids the limitations of a single model and lays a foundation for accurately screening key features. 3. Traditional information gain calculation is relatively simple in data processing and probability estimation, and does not fully consider the innovative distribution and uncertainty of ecological data. For example, when calculating conditional probability, it only relies on simple frequency statistics and does not reasonably model the uncertainty of data, resulting in inaccurate information gain calculation and unable to accurately reflect the true contribution of features to the prediction of the ecological value of ecological products. In this application, when calculating information gain, according to the definitions of information entropy and conditional entropy, probabilities are calculated through accurate sample statistics. Considering different categories of the ecological value of ecological products and different values of features, information gain is accurately calculated.When predicting the forest carbon sink value category, this precise calculation can accurately measure the degree to which each feature (such as tree species diversity, forest area, etc.) helps in classification, helping to determine which features are most critical for predicting the ecological value of ecological products and providing strong support for feature screening. 4. Traditional key feature screening is based on fixed thresholds or simple rankings, without fully considering the mutual influence between features and the dynamic changes in importance scores. This method often fails to adapt to the innovation of ecological data, easily missing some features that, although having low importance scores, have significant synergistic effects with other features, or misselecting some features that are important under specific conditions but overall unstable. After ranking features based on information gain in this application, the difference sequence of importance scores is processed by the locally weighted regression scatterplot smoothing method, and its fluctuation trend is analyzed to screen key features. This method considers the dynamic changes in feature importance scores and can identify features with relatively high importance and stability. By setting thresholds and window lengths, key features are screened according to the fluctuation trend, avoiding interference from excessive redundant or unimportant features. In an actual ecological product ecological value evaluation system, the key features that have the greatest impact on the evaluation results can be accurately determined, providing a precise decision-making basis for ecological protection and resource management and improving prediction accuracy and efficiency.

[0078] Optionally, feature construction is performed on the basis of the multi-modal key features to generate multi-modal feature data, including: mapping the multi-modal key features to a unified feature space to determine the potential relationships between different multi-modal key features and obtain an embedded feature representation; performing a non-linear transformation on the embedded feature representation to generate a structure mining feature; performing a derivative process on the structure mining feature to obtain an ecological diversity derivative feature; and performing hierarchical splicing on the ecological diversity derivative feature to generate multi-modal feature data. Preferably, in a specific application scenario, the above solution is described in an alternative or preferred manner.

[0079] 1. Map multi-modal key features to a unified feature space to determine potential relationships and obtain an embedded feature representation

[0080] Construct a deep embedding network: In order to map multi-modal key features to a unified feature space, this application constructs a deep embedding network Let the multi-modal key feature vector be where D is the total dimension of the multi-modal key features. The deep embedding network has L hidden layers, and the number of neurons in the l-th layer is H l (l = 1, …, L), and the weight matrix from the input layer to the first hidden layer is The bias vector is The weight matrix from the l-th hidden layer to the l + 1-th hidden layer is The bias vector is The output layer weight matrix is The bias vector is where E is the dimension of the embedded feature representation. Forward propagation calculates the embedded features: Forward propagation is performed through the following formula: h1 = σ1(W1F key + b1)h l = σ l (W l h l-1 + b l ), l = 2, …, L F embedded = W out h L + b out , where σ l is the activation function of the l-th layer. In this application, the Leaky ReLU function is selected, and its definition is α l is a small positive number (e.g., α l = 0.01). While maintaining the advantages of the ReLU function, it solves the problem that the gradient of the ReLU function is 0 when z < 0, enabling the network to better learn the innovative relationships between multi-modal key features. To train the deep embedding network, this application defines a loss function L embed , which is used to measure the difference between the embedded feature representation and the expected latent relationship representation. Considering the innovation and multi-modality of ecological data, this application adopts a hybrid loss function based on reconstruction error and contrast learning. For the reconstruction error part, let a set reconstruction function which reconstructs the embedded feature F embedded back to a representation similar to the original multi-modal key features The reconstruction error loss L recon is defined as: where N is the number of training samples, and are the multi-modal key feature vector and the embedded feature vector of the i-th training sample respectively. For the contrast learning part, this application randomly selects positive sample pairs (which come from the same ecosystem or samples with similar ecological characteristics) and negative sample pairs (which come from different ecosystems or samples with significantly different ecological characteristics) from the training data. For the embedded features and The contrast learning loss L contrast is defined as:

[0081] where sim(·, ·) is the similarity function, such as cosine similarity τ is the temperature parameter, which is used to adjust the intensity of contrast learning. The final loss function L embed = λrecon L recon + λ contrast L contrast , where λ recon and λ contrast are weight parameters determined by cross - validation, used to balance the reconstruction error loss and the contrastive learning loss. By minimizing L embed , the deep embedding network can learn the potential relationships between multi - modal key features and generate effective embedded feature representations F embedded .

[0082] 2. Perform a non - linear transformation on the embedded feature representation to generate structure - mining features

[0083] Non - linear transformation based on variational auto - encoder: To further mine the potential structure in the embedded feature representation, this application uses a variational auto - encoder (VAE) to perform a non - linear transformation on the embedded feature F embedded . The variational auto - encoder consists of an encoder q φ (z|F embedded ) and a decoder p θ (F embedded |z). The encoder maps the embedded feature F embedded to the latent space where Z is the dimension of the latent space. The encoder outputs the mean μ and the log variance logσ 2 of the latent variable z, that is where μ = μ(F embedded ; φ) and logσ 2 = logσ 2 (F embedded ; φ) are the outputs of the encoder, and φ is the parameter of the encoder. The specific calculation is: where sample z = μ + σ ⊙ ∈ from the latent variable z, where is the random noise of the standard normal distribution, and ⊙ represents element - wise multiplication. The decoder decodes the latent variable z back to the embedded feature space to obtain the reconstructed embedded feature where Loss function and training of VAE: The variational auto - encoder is trained by minimizing the variational lower - bound loss function L vae . This loss function consists of two parts: reconstruction loss and KL divergence. The reconstruction loss measures the difference between the reconstructed embedded feature and the original embedded feature F embedded , and is defined as: The KL divergence measures the difference between the distribution of the latent variable output by the encoder and the standard normal distribution , and promotes the latent variable to learn a meaningful distribution, and is defined as: The final variational autoencoder loss function L vae =L recon-vae +βL KL , where β is a hyperparameter used to balance the reconstruction loss and KL divergence. By minimizing L vae , the variational autoencoder can learn the potential structure in the embedded feature representation and generate the structure mining feature F structure ,This feature can better reflect the intrinsic structural relationship between multimodal key features.

[0084] 3. Derivative processing of structural mining features to obtain ecological diversity derivative features

[0085] In order to derive the structure mining features, this application adopts the Generative Adversarial Network (GAN). The Generative Adversarial Network consists of a generator G and a discriminator D. The generator G is a noise vector (N is the dimension of the noise vector) and the structure mining feature F structure As input, generate ecological diversity derived features F derived The generator G is a multi-layer neural network with M hidden layers and K neurons in the mth layer. m (m=1,…,M). The weight matrix from the input layer to the first hidden layer is The bias vector is The weight matrix from the mth hidden layer to the m+1th hidden layer is The bias vector is The output layer weight matrix is The bias vector is The output of the generator is calculated as: h G1 =σ G1 (W G1 [n; F structure ]+b G1 ), h Gm =σ Gm (W Gm h G(m-1) +b Gm ),m=2,…,M,F derived =W Gout h GM +b Gout where σ Gm is the activation function of the mth layer of the generator, such as the ReLU function. The discriminator D is used to distinguish the generated ecological diversity derived features F derived and the true structure mining feature F structure The discriminator D is also a multi-layer neural network with P hidden layers, and the number of neurons in the p-th layer is Q p (p=1,…,P). The weight matrix from the input layer to the first hidden layer is The bias vector is The weight matrix from the p-th hidden layer to the (p + 1)-th hidden layer is The bias vector is The output layer weight matrix is The bias vector is The output of the discriminator is calculated as: h D1 = σ D1 (W D1 F + b D1 ), h Dp = σ Dp (W Dp h D(p-1) + b Dp ), p = 2, …, P, where σ Dp is the activation function of the p-th layer of the discriminator, such as the ReLU function, is the sigmoid activation function, which is used to map the output of the discriminator to the interval [0, 1], representing the probability that the input features are real features. Loss function and training of GAN: The generative adversarial network is optimized through adversarial training. The goal of the generator is to generate as realistic ecological diversity-derived features as possible, making it difficult for the discriminator to distinguish the generated features from the real structure-mined features; the goal of the discriminator is to accurately distinguish the generated features from the real features. The loss function L G of the generator is defined as: The loss function L D of the discriminator is defined as: During the training process, the generator and the discriminator are alternately optimized. By continuously adjusting their respective parameters, the ecological diversity-derived features generated by the generator become more and more realistic, and the discrimination ability of the discriminator also becomes stronger. Finally, the ecological diversity-derived features F derived generated by the generator can enrich the feature representation of ecological data and reflect the diversity information in the ecosystem.

[0086] 4. Perform hierarchical splicing on the ecological diversity-derived features to generate multi-modal feature data

[0087] Hierarchical splicing strategy: To generate multi-modal feature data, the present application uses a hierarchical splicing method to process the ecological diversity-derived features. Suppose the present application has S different levels of splicing structures. At the first level, the present application splices the ecological diversity-derived feature F derived with the original multi-modal key feature F key to obtain the first-level spliced feature F concat1 = [F key ; F derived . At the second level, the present application performs splicing on the first-level spliced feature F concat1Perform a non - linear transformation and process it through a multi - layer perceptron (MLP). Let the MLP have R hidden layers, and the number of neurons in the r - th layer be U r (r = 1,…,R). The weight matrix from the input layer to the first hidden layer is The bias vector is The weight matrix from the r - th hidden layer to the (r + 1)-th hidden layer is The bias vector is The output layer weight matrix is The bias vector is where V is the feature dimension before concatenation in the second layer. The transformation process through the MLP is as follows: h MLP1 = σ MLP1 (W MLP1 F concat1 + b MLP1 ), h MLPr = σ MLPr (W MLPr h MLP(r-1) + b MLPr ), r = 2,…,R, F transformed = W MLPout h MLPR + b MLPout . In this application, σ MLPr is the activation function of the r - th layer of the MLP. For example, the ReLU function σ MLPr (z)= max(0,z) is selected to introduce non - linear transformation and enhance the model's ability to capture innovative relationships between features. Then, the feature F transformed after being transformed by the MLP is concatenated again with the first - layer concatenated feature F concat1 to obtain the second - layer concatenated feature F concat2 = [F concat1 ; F transformed . In this way, in each subsequent layer s (s = 3,…,S), the above process is repeated: first, perform a similar MLP transformation on the concatenated feature F concat(s-1) of the previous layer to obtain F transformed(s-1) , and its transformation process is similar to that of the second - layer MLP transformation, except that the weight matrix and bias vector are adjusted according to the current layer, that is:

[0088] h MLP1(s-1) = σ MLP1(s-1) (W MLP1(s-1) F concat(s-1) + b MLP1(s-1) )

[0089] h MLPr(s-1) = σ MLPr(s-1) (W MLPr(s-1) h MLP(r-1)(s-1) + b MLPr(s-1) ), r = 2,…,R

[0090] F transformed(s-1) = W MLPout(s-1) h MLPR(s-1) + b MLPout(s-1)

[0091] Then splice F transformed(s-1) with F concat(s-1) to obtain F concats = [F concat(s-1) ; F transformed(s-1) . Finally, after hierarchical splicing through S layers, the obtained F concatS is the generated multi-modal feature data. This way of hierarchical splicing can fully explore the relationships at different levels between multi-modal key features and ecological diversity-derived features, generate multi-modal feature data with rich information and strong representation ability, and provide more powerful data support for the subsequent analysis and prediction of the ecological value of ecological products. From the perspective of this application, it needs to perform a large number of numerical calculation operations such as matrix multiplication, addition, and activation function operations in sequence to complete the generation process from ecological diversity-derived features to multi-modal feature data.

[0092] Therefore, the above preferred or alternative technical solutions have the following technical advantages.

[0093] 1. When dealing with multimodal data, traditional methods simply concatenate different-modal data or use simple linear transformations to attempt to integrate features. This approach cannot deeply explore the potential relationships between key features of different modalities, resulting in a lack of effective capture of the internal structure of the data in the feature representation. For example, in ecological data, there are interactions between data of different modalities (such as meteorology, soil, and biodiversity), but simple concatenation or linear transformation is difficult to reveal these relationships, making it impossible to fully utilize the advantages of multimodal data in subsequent analysis. This application constructs a deep embedding network and trains it using a hybrid loss function based on reconstruction error and contrastive learning. The multi-layer structure of the deep embedding network can automatically learn the non-linear relationships between multimodal key features and map them to a unified feature space. The reconstruction error loss ensures that the embedded features can restore the original multimodal key features as much as possible, while the contrastive learning loss prompts the network to learn a discriminative feature representation, strengthening the aggregation of similar samples in the embedding space and the separation of different samples. In an ecological scenario, this helps to discover the hidden associations between different ecological factors (such as temperature, soil nutrients, and species number), provides a more powerful feature representation for understanding the internal mechanisms of ecosystems, and improves the accuracy of subsequent ecological analysis tasks (such as ecological value prediction and ecosystem health assessment). 2. Traditional feature transformations use simple functions (such as standardization, normalization) or shallow non-linear transformations (such as simple polynomial transformations). These methods cannot fully explore the potential structure in the features and are difficult to capture the deep feature relationships for ecological data. For example, simple standardization can only adjust the scale of the features and cannot discover the high-order dependence relationships between the features, while the expressive ability of shallow non-linear transformations is limited and cannot adapt to the innovation of ecological data. This application uses a variational autoencoder (VAE) to perform non-linear transformations on the embedded features. The VAE maps the embedded features to the latent space through the encoder, learns the latent distribution of the data, and the decoder then reconstructs the embedded features from the latent space. In this process, the VAE can capture the potential structure in the embedded features and generate more representative structure-mining features. The reconstruction loss and KL divergence term in the variational lower-bound loss function respectively ensure the accuracy of reconstruction and the rationality of the latent variable distribution. In ecological applications, the VAE discovers the hidden patterns of ecological data in the latent space, such as the potential causal relationships or co-variation patterns between different ecological indicators, provides a new perspective and more valuable feature representation for ecological research, and helps to improve the prediction ability and interpretability of ecological models. 3. Traditional feature derivation methods are based on empirical rules or simple data transformations, such as performing addition, subtraction, multiplication, and division operations on existing features to obtain new features. The derived features generated in this way often lack innovation and diversity and are difficult to reflect the innovative diversity of ecosystems. For example, simple empirical rule derivation cannot capture the new features generated by the interactions between different factors in an ecosystem, resulting in the omission of important ecological information.This application uses a generative adversarial network (GAN) for feature derivation. The generator of the GAN takes a noise vector and structure-mined features as inputs to generate ecologically diverse derived features, while the discriminator distinguishes the generated features from the real structure-mined features. Through this adversarial training mechanism, the generator can learn the distribution of real features and generate diverse and authentic derived features. In the ecological field, the GAN generates new features reflecting the diversity of ecosystems, such as simulating new combinations of ecological indicators that appear under different ecological conditions, providing more data dimensions and properties for ecological research, and helping to discover new ecological laws and potential ecological value influencing factors. 4. Traditional feature splicing methods are usually a one-time simple splicing, that is, directly connecting all relevant features together without considering the hierarchical structure and interaction between features. This simple splicing results in a too-high feature dimension, information redundancy, and it is difficult to effectively utilize the feature relationships at different levels. In ecological data processing, simple splicing cannot highlight the importance and mutual influence of different ecological factors at different levels, reducing the model's ability to understand the innovative relationships of ecosystems. This application adopts a hierarchical splicing strategy. Through multi-layer splicing and non-linear transformation, the feature relationships at different levels are gradually mined. In each layer, first perform a non-linear transformation (such as through an MLP) on the spliced features of the previous layer to capture the innovative non-linear relationships between features, and then splice them with the features of the previous layer, continuously enriching the feature representation. This method can effectively integrate information at different levels, avoid information redundancy, and highlight the hierarchical structure and interaction between features. In an ecological scenario, hierarchical splicing simulates the relationships between different levels of ecological processes (such as individual - population - community - ecosystem) in an ecosystem, and the generated multi-modal feature data can more accurately reflect the innovation of the ecosystem, providing more targeted and accurate information support for ecological analysis and decision-making.

[0094] Optionally, based on the ecological detection model deployed on the background server, perform feature extraction on the multi-modal feature data to predict the ecological value of the ecological product based on the extracted features, including: Based on the spatial feature extraction layer, perform convolution, pooling, and activation operations on the multi-modal feature data to extract the spatial features in the data and output a spatial feature tensor; Based on the temporal feature processing layer, extract the temporal information of the spatial feature tensor through a gating mechanism to output a feature sequence containing time-dependent information; Based on the fully connected layer, predict the ecological value of the feature sequence containing time-dependent information. Preferably, in a specific application scenario, illustrate the above solution in an alternative or preferred manner.

[0095] 1. Based on the spatial feature extraction layer, perform convolution, pooling, and activation operations on the multi-modal feature data to extract the spatial features in the data and output a spatial feature tensor

[0096] Let the multi-modal feature data be Where N is the number of samples, representing the number of multimodal ecological data samples collected under different conditions such as time or location; C is the number of feature channels, and different channels correspond to different types of ecological characteristics, such as temperature, humidity, species richness and other different ecological indicators; H and W represent the height and width of the feature map, respectively, which can be understood as the number of grids used to divide the ecological area in the spatial dimension, and the corresponding ecological feature values are recorded at different grid positions. Convolution operation: Define a set of convolution kernels Where l represents the lth group of convolution kernels, and each group of convolution kernels is used to extract spatial features of different levels or types. i and j are the position indexes where the convolution kernel slides on the feature map, k h,l and k w,l are the height and width of the lth group of convolution kernels, which determine the size of the receptive field of the convolution kernels in space. The receptive field sizes of different groups of convolution kernels can be different to capture spatial features of different scales; C is the number of input feature channels, C out,l is the number of feature channels output after the lth group of convolution kernel operations. The convolution calculation is implemented by the following formula: where n∈{1,…,N}, c out,l ∈{1,…,C out,l}, h′∈{1,…,H′ l}, w′∈{1,…,W′ l}, H′ l =Hk h,l +1, W′ l =Wk w,l +1, is a bias term used to adjust the overall offset of the output features. l,m and ω l,m is an additional parameter introduced, α l,m Controls the intensity of the sinusoidal modulation, ω l,m Determine the frequency of the sinusoidal modulation, and through this sinusoidal modulation, capture more spatial frequency characteristics in the convolution process. This helps to explore the periodicity or fluctuation of ecological characteristics in space in ecological data processing. After L groups of convolution kernel operations, the feature map is obtained. Concatenate the outputs of all groups according to the channel dimension to obtain Pooling operation: This application adopts an adaptive hybrid pooling method that combines average pooling and maximum pooling. The pooling window size is defined as p h ×p w , the step size is s h and s w For feature map Y, within each pooling window, calculate the average pooling value And the maximum pooling value maxY n,c,h″,w″ : where n ∈ {1, …, N}, h″ ∈ {1, …, H″}, w″ ∈ {1, …, W″}, Then, a learnable weight γ n,c,h″,w″ is used to fuse the average pooling value and the max pooling value: The weight γ n,c,h″,w″ is calculated through a small neural network which takes the feature values within the pooling window as input and outputs a value between [0, 1]. In this way, according to the feature distribution at different positions, the results of average pooling or max pooling are adaptively selected to better retain feature information. This application tries a variant of the improved activation function - Scaled Exponential Linear Unit (SELU). Define the activation function σ(x) as: where λ and α are two learnable parameters that are optimized during training through backpropagation. For the feature map Z output by the pooling operation, after the activation operation, the spatial feature tensor S is obtained: S n,c,h″,w″ = σ(Z n,c,h″,w″ ) This activation function can automatically adjust the activation characteristics according to the data distribution when processing ecological data, which helps the model better learn the ecological feature relationships.

[0097] 2. Based on the temporal feature processing layer, the temporal information of the spatial feature tensor is extracted through a gating mechanism to output a feature sequence containing time-dependent information

[0098] Reshape the spatial feature tensor S into a form suitable for temporal processing. Let the reshaped tensor be where T = H″ × W″, that is, the spatial dimension is flattened into the temporal dimension, representing the feature dimension at each time step. In the ecological scenario, this transformation regards the ecological features at different spatial positions as a sequence that appears sequentially in time, which is convenient for analyzing the dynamic changes of the ecosystem in space and time. Multi-Head Attention Gated Recurrent Unit (MHAGRU): The multi-head attention gated recurrent unit is used to process the reshaped feature sequence to extract richer temporal information. MHA GRU combines the advantages of the multi-head attention mechanism and the gated recurrent unit. First, define the multi-head attention mechanism. Let the input feature sequence be (corresponding to the feature vector at the t-th time step in S reshaped ), and project it into the query (Query), key (Key), and value (Value) spaces respectively: Q t = W q x t + b q K t= W k x t + b k V t = W v x t + b v , where is the weight matrix, is the bias vector, D q 、D k and D v are the dimensions of the query, key, and value spaces, respectively. The query, key, and value are split on the head dimension, and there are h heads: where i ∈ {1, …, h}. Calculate the attention scores for each head: Concatenate the attention results of all heads and project back to the original dimension: A t = Concat(Attention1(Q t,1 , K t,1 , V t,1 ), …, Attention h (Q t,h , K t,h , V t,h ))W o + b o , where Then, combine the output A t of the multi-head attention with a gated recurrent unit. Let the hidden state at the previous time step be Define the weight matrix Bias vector Reset gate r t and update gate z t are calculated as follows: r t = σ sigmoid (W xr A t + W hr h t-1 + b r ), z t = σ sigmoid (W xz A t + W hz h t-1 + b z ). The candidate hidden state is calculated as: The hidden state h at the current time step t is: After being processed by the MHAGRU layer, for each sample \(n\in\{1,\ldots,N\}\), a feature sequence containing time-dependent information is output where \(H\) n \( = [h_1,\ldots,h\) T , and \(h\) t is the hidden state of the \(n\)-th sample at the \(t\)-th time step. This feature sequence can better capture the long-range dependencies in the feature sequence by combining the multi-head attention mechanism, providing richer and more accurate context information for subsequent ecological value prediction.

[0099] 3. Based on the fully connected layer, predict the ecological value of the feature sequence containing time-dependent information

[0100] To input the feature sequence containing time-dependent information into the fully connected layer for prediction, a weighted attention pooling method is used to aggregate the feature sequence. Define the attention weight \(\beta\) t , which is calculated through a small neural network . This neural network takes the hidden state sequence \(H\) n \( = [h_1,\ldots,h\) T as input and outputs an attention weight vector \(\beta\) n \( = [\beta\) n,1 ,\ldots,\beta\) n,T of length \(T\), where The aggregated feature vector is calculated as follows: This weighted attention pooling can perform weighted aggregation according to the importance of features at different time steps, highlighting the time step information that is more critical for ecological value prediction. Define the weight matrix of the fully connected layer and the bias vector where \(O\) is the output dimension of the ecological value prediction. For example, if predicting multiple value indicators of ecological products (such as carbon sequestration value, biodiversity value, water conservation value, etc.), \(O\) is the number of these indicators. The aggregated feature vector is used for ecological value prediction through the fully connected layer, and the calculation formula is: where is the ecological value prediction result of the \(n\)-th sample, is the Softmax function, which converts the output of the fully connected layer into a probability distribution, representing the prediction probability of each ecological value indicator. In the above formula, is the main fully connected layer weight matrix, is the bias vector, and their functions are the same as those of a conventional fully connected layer, performing a linear transformation on the aggregated feature vector to initially generate output values corresponding to the number of ecological value indicators. And this part is the introduced additional non-linear term. Among them is the m-th additional weight matrix, ω m is the angular frequency vector, is the phase vector. represents the dot product operation of the vector ω m and , and the result is a scalar. Then, adding the phase makes it the independent variable of the sine function. Through this additional term, a more non-linear relationship between the feature vector and the ecological value is captured, further enhancing the expressive power of the model. For example, in ecological value prediction, there are periodic or fluctuating correlations between certain values of the ecosystem and multiple ecological factors, and this non-linear term helps to explore and reflect these relationships. When this application is executed, first perform the matrix multiplication to obtain a vector of dimension O, and then add the bias vector b fc , and then calculate This involves multiple matrix multiplications, vector dot products, sine function operations, and summation operations. Finally, input the above result into the Softmax function to obtain the final ecological value prediction probability distribution

[0101] Therefore, the above-mentioned preferred or alternative technical solutions have the following technical advantages.

[0102] 1. Traditional spatial feature extraction only uses simple convolution kernels for convolution operations, and the parameters of the convolution kernels are fixed, lacking the ability to effectively capture spatial features at different scales. Pooling operations mostly use a single max pooling or average pooling, and cannot adaptively select the appropriate pooling method according to the data characteristics. Activation functions also often choose relatively simple ones such as ReLU, with limited ability to depict innovative non-linear relationships in the data. In ecological data processing, simple convolution cannot fully explore the innovative relationships between different ecological indicators in space, a single pooling method loses important information, and simple activation functions are difficult to accurately fit the innovative changes of ecological features. This application introduces multiple groups of convolution kernels with different parameters and combines sine modulation. Multiple groups of convolution kernels can simultaneously capture spatial features at different levels and scales, such as the interaction of ecological indicators in different sized regions. Sine modulation, through the additional parameters α l,m and ω l,m , can explore the periodic or fluctuating laws of ecological features in space, such as the spatial periodic changes existing in certain phenomena in the ecosystem. This enables the convolution operation to extract spatial features in ecological data more comprehensively and deeply. This application adopts an adaptive hybrid pooling method, combining average pooling and max pooling, and calculates the weight γ through a small neural network n,c,h",w″To fuse the results of both. This method can adaptively select to retain the mean information (average pooling) or highlight the local maximum information (max pooling) according to the feature distributions at different positions, avoiding the limitations of a single pooling method and better retaining the key information of ecological data in the spatial dimension. This application uses a variant SELU activation function with learnable parameters. In ecological data, the data distribution is innovatively variable, and the learnable λ and α parameters enable the activation function to automatically adjust the activation characteristics according to the data characteristics, so that the model can better learn the non-linear relationships between ecological features and has stronger adaptability compared to traditional activation functions with fixed parameters. 2. Traditional temporal feature extraction only relies on simple recurrent neural networks (RNNs) or gated recurrent units (GRUs), lacking the ability to effectively capture long-range dependencies. When dealing with multi-modal ecological data, it is difficult to fully utilize the innovative correlation information between different features. For example, a simple GRU cannot accurately capture the mutual influence between ecological factors separated by a long time step in an ecosystem. This application combines the multi-head attention mechanism and GRU to form MHAGRU. The multi-head attention mechanism can capture long-range dependencies in the feature sequence by projecting the input into query, key, and value spaces and calculating attention scores in multiple heads. In the ecological data scenario, this helps to mine the interactions between ecological factors at different time points, such as the correlations between ecological indicators in different seasons. Then, the output of the multi-head attention is combined with GRU, enabling the model to both capture long-range dependencies and effectively process temporal information using GRU's gating mechanism, thus more comprehensively and accurately extracting the temporal features in ecological data. 3. Traditional fully connected layer prediction usually just performs a simple linear transformation followed by the Softmax function, making it difficult to capture the non-linear relationship between features and ecological value. In ecological value prediction, the innovation of the ecosystem makes the relationship between ecological features and value not simply linear, and traditional methods cannot accurately fit this innovative relationship, resulting in limited prediction accuracy. This application uses weighted attention pooling to aggregate the feature sequence, calculating the attention weight β n,t . This enables the model to perform weighted aggregation according to the importance of features at different time steps for ecological value prediction, highlighting the information at key time steps and being able to utilize the useful information in temporal features more effectively compared to simple average pooling or max pooling. This application introduces an additional non-linear term in the fully connected layer prediction formula. This part of the non-linear term can capture a more non-linear relationship between the feature vector and ecological value through the sine function and additional weight matrix, angular frequency vector, and phase vector, further enhancing the model's expressive ability and thus more accurately predicting the ecological value of ecological products.

[0103] The scope of the invention involved in this application is not limited to the technical solutions formed by the specific combination of the above technical features, and should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept.

Claims

1. An intelligent dynamic monitoring method for ecological products, characterized in that, including obtaining an index set for dynamically monitoring ecological products, where the index set includes at least one of the following: ecosystem health indicators, environmental quality indicators, ecological diversity indicators, resource utilization indicators, low-carbon attribute indicators; monitoring at least one of the following ecological data within a set range of the area where the ecological products are located based on the set Internet of Things sensors: air quality, water quality, soil, meteorology, biodiversity, forest fires; collecting ecological data from the Internet of Things sensors based on the configured edge nodes, preprocessing the collected ecological data to obtain multi-modal feature data, and transmitting the multi-modal feature data to the background server; extracting features from the multi-modal feature data based on the ecological detection model deployed on the background server to predict the ecological value of the ecological products based on the extracted features.

2. The method according to claim 1, wherein The step of collecting ecological data from the Internet of Things sensors based on the configured edge nodes and preprocessing the collected ecological data to obtain multi-modal feature data includes: performing spatio-temporal alignment on different types of ecological data collected from the Internet of Things sensors based on the Kalman filter-based clock synchronization model; performing anomaly detection on the spatio-temporally aligned ecological data based on the 3σ criterion and the isolation forest model to remove abnormal ecological data and thereby obtain multi-modal purified data; performing feature engineering on the multi-modal purified data to generate multi-modal feature data.

3. The method according to claim 2, wherein The step of performing feature engineering on the multi-modal purified data to generate multi-modal feature data includes: performing statistical feature extraction, time-domain feature extraction, and frequency-domain feature extraction on the multi-modal purified data to obtain a multi-modal cross-domain feature set; performing correlation analysis and feature importance evaluation on the multi-modal cross-domain feature set to screen out multi-modal key features therefrom; performing feature construction based on the multi-modal key features to generate multi-modal feature data.

4. The method according to claim 3, wherein The step of performing statistical feature extraction, time-domain feature extraction, and frequency-domain feature extraction on the multi-modal purified data to obtain a multi-modal cross-domain feature set includes: calculating the basic statistics of the multi-modal purified data to estimate the central tendency, dispersion degree, and extreme features of the multi-modal purified data based on the basic statistics; performing empirical mode decomposition on the multi-modal purified data to determine the dependence features, local dynamic change features, and features at different time scales of the multi-modal purified data in the time dimension; transforming the multi-modal purified data into the frequency domain, performing power spectral density statistics and frequency band feature extraction to determine the frequency domain energy distribution of the data and the frequency domain characterization features of ecological phenomena.

5. The method according to claim 3, characterized in that, The step of performing correlation analysis and feature importance evaluation on the multi-modal cross-domain feature set to screen out multi-modal key features therefrom includes: performing pairwise correlation calculation on each feature in the multi-modal cross-domain feature set based on the constructed correlation matrix to determine the associated feature values between the features; estimating the contribution degree of each feature in the multi-modal cross-domain feature set to the prediction result to determine the importance scores of each feature; calculating the information gain of each feature for predicting the ecological value of ecological products according to the associated feature values between the features. Based on the information gain of the ecological value prediction of ecological products for each feature, the fluctuation trend of the corresponding importance scores is statistically analyzed, and based on the fluctuation trend, multi-modal key features are screened out from the features of the multi-modal cross-domain feature set.

6. The method according to claim 3, characterized in that Feature construction is performed on the basis of the multi-modal key features to generate multi-modal feature data, including: Mapping the multi-modal key features to a unified feature space to determine the potential relationships between different multi-modal key features and obtain an embedded feature representation; Performing a non-linear transformation on the embedded feature representation to generate a structure mining feature; Performing a derivative process on the structure mining feature to obtain an ecological diversity derivative feature; Performing hierarchical splicing on the ecological diversity derivative features to generate multi-modal feature data.

7. The method according to claim 2, wherein Based on the ecological detection model deployed on the background server, feature extraction is performed on the multi-modal feature data, and based on the extracted features, the ecological value of the ecological product is predicted, including: Based on the spatial feature extraction layer, performing convolution, pooling, and activation operations on the multi-modal feature data to extract spatial features in the data and output a spatial feature tensor; Based on the temporal feature processing layer, extracting the temporal information of the spatial feature tensor through a gating mechanism to output a feature sequence containing time-dependent information; Based on the fully connected layer, predicting the ecological value of the feature sequence containing time-dependent information.

Images