Space-air-ground integrated ecological monitoring method and system based on multi-source data fusion

Abstract

The invention provides a space-air-ground integrated ecological monitoring method and system based on multi-source data fusion, and the method comprises the steps: collecting a multi-source heterogeneous data set, carrying out the space-time alignment, multi-scale fusion and dynamic deduction, and generating an ecological change dynamic deduction result; and carrying out ecological monitoring early warning and decision support based on the result. According to the method, cross-platform data is safely cleaned and enhanced through a federated learning framework, and data modal differences are eliminated; deep fusion static feature regression and time sequence trend prediction are carried out by using a double-model architecture, and microenvironment detail changes and long-term ecological trends are synchronously captured; and finally, through a decision strategy model driven by a double-target reward function, quantifying an ecological restoration cost-benefit ratio and a stability gain as an optimal path, forming a'data fusion-dynamic deduction-decision support 'full-link closed loop, improving ecological monitoring space-time continuity, trend prediction accuracy and management decision scientificity, and improving ecological monitoring efficiency. The contradiction that in the prior art, a user can see widely but cannot see finely, and the user can measure finely but does not tend to measure is overcome.

Description

Technical Field

[0001] This invention relates to the field of ecological monitoring technology, and in particular to an integrated air-space-ground ecological monitoring method and system based on multi-source data fusion. Background Technology

[0002] Traditional ecological monitoring technologies primarily rely on single data sources. For example, satellite remote sensing acquires large-scale land cover information through optical or radar sensors; aerial photogrammetry uses drones or fixed-wing aircraft to collect high-resolution images; and ground sensor networks monitor microenvironmental parameters such as temperature, humidity, and soil moisture through IoT nodes. These technologies follow a linear process of "data acquisition - independent analysis - result output." For instance, satellite data is used to invert vegetation cover using the NDVI index, aerial data is used to identify land surface types through texture analysis, and ground sensors use threshold alarms to detect anomalies. Their technical principles are based on the physical characteristics of different data sources: satellite data has the advantage of wide-area coverage but limited resolution; aerial data provides detailed information but is costly; and ground sensors enable real-time monitoring but have limited spatial coverage. Existing technologies have formed mature single-source data processing systems, such as ENVI-based remote sensing image classification and ArcGIS-based GIS spatial analysis.

[0003] However, existing technologies suffer from fragmented data sources, leading to fragmented ecological monitoring. For example, satellite data cannot capture micro-environmental changes monitored by ground sensors, and aerial data is difficult to reflect long-term time-series dynamics, creating a contradiction of "being able to see the large area but not the details, and being able to measure the details but not the trend." This results in problems such as an imbalance in spatiotemporal resolution and weak correlation between multi-source data in ecological monitoring results. Summary of the Invention

[0004] This invention aims to at least address the technical problems of unbalanced spatiotemporal resolution and weak correlation of multi-source data in existing ecological monitoring results. In particular, it innovatively proposes an integrated air-space-ground ecological monitoring method and system based on multi-source data fusion.

[0005] To achieve the above-mentioned objectives of this invention, this invention provides an integrated air-space-ground ecological monitoring method based on multi-source data fusion, the method comprising: S1. Collect raw satellite data, raw aerial data, and ground sensor data to obtain a multi-source heterogeneous dataset; S2. Perform spatiotemporal alignment on the multi-source heterogeneous dataset to obtain spatiotemporal sequence aligned data; S3. Perform multi-scale fusion on the spatiotemporal sequence aligned data to obtain a fused feature vector; S4. Based on the fused feature vector, perform dynamic inference to generate dynamic inference results of ecological changes; S5. Based on the dynamic simulation results of ecological changes, conduct ecological monitoring, early warning, and decision support, wherein the ecological monitoring and early warning includes real-time monitoring and early warning of abnormal changes in the ecological environment, and the decision support includes providing targeted ecological protection and management suggestions based on the dynamic simulation results of ecological changes.

[0006] On the other hand, the present invention also provides an integrated air-space-ground ecological monitoring system based on multi-source data fusion, the system comprising: processor; Memory used to store processor-executable instructions; The processor is configured to implement the integrated air-space-ground ecological monitoring method based on multi-source data fusion when executing the executable instructions.

[0007] The beneficial effects of this invention are as follows: This invention effectively solves the problems of spatiotemporal resolution imbalance and weak correlation caused by fragmented data sources in traditional ecological monitoring by using multi-source heterogeneous data fusion and spatiotemporal alignment technology. First, this method achieves secure cleaning and enhancement of cross-platform data through a federated learning framework, eliminating modal differences between data from different sources. Then, it employs a dual-model architecture (random forest + LSTM) for deep fusion of static feature regression and time-series trend prediction, achieving a breakthrough in simultaneously capturing micro-environmental detail changes and long-term ecological trends while maintaining wide-area coverage advantages. Finally, through a decision-making strategy model driven by a dual-objective reward function, it quantifies the cost-benefit ratio of ecological restoration and stability gain into an executable optimal path, forming a closed-loop chain of "data fusion - dynamic inference - decision support." This significantly improves the spatiotemporal continuity of ecological monitoring, the accuracy of trend prediction, and the scientific nature of management decisions, fundamentally overcoming the core contradiction of existing technologies that "can see broadly but not in detail, and measure in detail but not in trends."

[0008] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0009] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a flowchart of an integrated air-space-ground ecological monitoring method based on multi-source data fusion, according to the present invention. Detailed Implementation

[0010] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0011] Example 1 like Figure 1 As shown, an integrated air-space-ground ecological monitoring method based on multi-source data fusion is proposed, the method comprising: S1. Collect raw satellite data, raw aerial data, and ground sensor data to obtain a multi-source heterogeneous dataset; In step S1, it is necessary to explain in detail that, in this embodiment, the raw satellite data is acquired by satellite and includes multispectral imagery, hyperspectral imagery, and synthetic aperture radar imagery, with a spatial resolution covering the range of 0.5 meters to 30 meters and a temporal resolution supporting daily to monthly dynamic updates. The raw aerial data is acquired using a UAV platform equipped with a hyperspectral imager and LiDAR, with a spectral band range covering 400-2500 nm, a spatial resolution of 0.1 meters, and a flight altitude set between 100-500 meters. The ground sensor network deploys monitoring nodes for temperature and humidity, soil moisture, gas concentration, and biomass, with a sampling frequency set to 15 minutes / time. Data transmission uses the LoRa wireless communication protocol to ensure stable transmission in complex terrain. All data undergoes preliminary verification through edge computing devices to remove outliers caused by equipment failure or environmental interference, forming a multi-source heterogeneous dataset containing spatiotemporal coordinates and data quality identifiers.

[0012] S2. Perform spatiotemporal alignment on multi-source heterogeneous datasets to obtain spatiotemporally aligned sequence data; S3. Perform multi-scale fusion on the spatiotemporal sequence aligned data to obtain the fused feature vector; S4. Dynamic simulation based on fused feature vectors to generate dynamic simulation results of ecological changes; S5. Based on the dynamic simulation results of ecological changes, conduct ecological monitoring, early warning and decision support. Ecological monitoring and early warning includes real-time monitoring and early warning of abnormal changes in the ecological environment, and decision support includes providing targeted ecological protection and management suggestions based on the dynamic simulation results of ecological changes.

[0013] In this embodiment, the principle of an integrated air-space-ground ecological monitoring method based on multi-source data fusion is as follows: First, a three-dimensional data acquisition network is constructed using satellite, aerial, and ground sensors. Satellite data provides a wide-area macroscopic background, aerial data enables high-precision characterization of key areas, and the ground sensor network captures detailed changes in the microenvironment. Second, a federated learning framework is used to securely clean and enhance the multi-source heterogeneous data, eliminating data differences between different modalities and constructing a standardized dataset with spatiotemporal alignment. Subsequently, Gabor filters, wavelet transforms, and empirical mode decomposition are used to extract multi-scale features. The feature weights are evaluated using the ReliefF algorithm, and nonlinear dimensionality reduction is performed using the t-SNE algorithm. Finally, the multi-modal data is processed... The Transformer fusion model achieves cross-modal feature fusion. Based on this, a dual-model architecture of random forest regression and LSTM time series prediction is constructed to perform static feature regression analysis and dynamic trend prediction, respectively. After generating preliminary inference results through the fusion layer, uncertainty quantification and calibration are performed through a Bayesian optimization framework. Finally, based on the dynamic inference results of ecological changes, on the one hand, an ecological parameter probability distribution model is constructed through kernel density estimation to achieve anomaly early warning. On the other hand, key driving factors are extracted to construct a multi-dimensional causal inference network. A decision strategy model driven by a dual-objective reward function is trained using a reinforcement learning framework to generate executable suggestions including ecological restoration priority planning, dynamic adjustment of protected areas, and resource allocation schemes.

[0014] As an optional embodiment of the present invention, step S2 may involve spatiotemporal alignment of the multi-source heterogeneous dataset to obtain spatiotemporally aligned sequence data; including: S201. Use the federated learning framework to perform multimodal data cleaning on multi-source heterogeneous datasets, and perform encrypted sample alignment, normalization and dimensionality reduction to obtain a cleaned and enhanced subset of data. The mathematical expression for the federated learning framework is: in, This represents the federal cleaning framework function. Indicates the federal batch normalization layer. This indicates the number of sources for a multi-source heterogeneous dataset. This represents a privacy set intersection protocol based on elliptic curve Diffie-Hellman. Indicates the first The original dataset of each client, Indicates the use of a public key Perform RSA encryption. Represents the normalized function. Indicates client Original dataset Specific numerical characteristics in Indicates client The average of the data. Indicates client standard deviation This indicates dimensionality reduction using kernel principal component analysis. Represents the kernel function.

[0015] In step S201, it is necessary to explain in detail that the federated learning framework completes global data augmentation through distributed node collaborative processing while ensuring the privacy of data from each client. Specifically, firstly, the Diffie-Hellman elliptic curve protocol is used to align encrypted samples, ensuring that data from different sources only perform intersection calculations in the encrypted space, avoiding leakage of original data; then, a federated batch normalization layer is used to normalize the data from each client, eliminating dimensional bias caused by device differences; for high-dimensional heterogeneous data, key features are extracted using kernel principal component analysis dimensionality reduction technology, where the kernel function is a Gaussian radial basis function, which preserves the nonlinear structure of the data while reducing computational complexity. This process optimizes the federated cleaning framework function through multiple rounds of iteration, so that the cleaned and augmented data subset significantly improves data quality and modal consistency while maintaining the original distribution characteristics.

[0016] S202. Construct a distance matrix based on Euclidean distance, calculate the similarity of the cleaned and enhanced data subset, and obtain the similarity matrix. In step S202, it is necessary to explain in detail that the distance matrix is ​​constructed with each data point as a node. A symmetric matrix is ​​formed by calculating the Euclidean distance between each pair of nodes, where the element values ​​reflect the spatial similarity between data points. The similarity calculation uses a Gaussian kernel function to map the Euclidean distance to the [0,1] interval, generating a similarity matrix. This process achieves multi-scale similarity characterization by dynamically adjusting the σ value—when σ is large, it captures global structural features, and when σ is small, it focuses on local detail changes. The final generated similarity matrix not only preserves the spatial topological relationships of the original data, but also enhances the fusionability of heterogeneous data through the nonlinear transformation of the kernel function.

[0017] S203. Perform spatiotemporal alignment based on the similarity matrix to match data from different sources and at different times in the time and space dimensions to obtain spatiotemporal sequence aligned data.

[0018] In step S203, it is necessary to explain in detail that the spatiotemporal alignment operation adopts a method combining the existing dynamic time warping algorithm with a greedy matching strategy based on the similarity matrix. Specifically, firstly, the dynamic time warping algorithm is used to flexibly align the time series. This algorithm finds the optimal matching path by constructing a distance matrix, which can handle the time series misalignment problem caused by inconsistent sampling frequencies of different data sources. Then, based on the greedy matching strategy, data points are paired in the spatial dimension. Using the similarity matrix as the weight, the spatial nodes with the highest similarity are matched first, ensuring accurate correspondence of cross-platform data in geographic coordinates. For time segments with missing values, a spatiotemporal interpolation method based on K nearest neighbors is used for completion, where the K value is determined to be 5 through cross-validation, which ensures interpolation accuracy and avoids overfitting. The final spatiotemporal sequence aligned dataset contains a unified timestamp and geographic coordinate system, and each data point is labeled with a data source identifier and quality score. This process, through a spatiotemporal dual-dimensional alignment mechanism, effectively solves the problem of fusion error accumulation caused by inconsistent spatiotemporal benchmarks in traditional methods, and significantly improves the spatiotemporal correlation and fusion quality of multi-source heterogeneous data.

[0019] As an optional embodiment of the present invention, optionally, performing multi-scale fusion on the spatiotemporal sequence aligned data in step S3 to obtain the fused feature vector includes: S301. Extract satellite image texture features from raw satellite data using Gabor filter, extract frequency domain features of aerial hyperspectral data from raw aerial data using wavelet transform, and analyze the intrinsic mode components of ground sensor data using empirical mode decomposition to obtain a multi-source feature set. In step S301, it is necessary to explain in detail that the Gabor filter, as an effective tool for joint analysis of the spatial and frequency domains, can accurately capture texture features such as vegetation cover and surface roughness in satellite imagery by setting kernel functions of different directions and scales. Specifically, an 8-direction, 5-scale Gabor filter bank is used to perform convolution operations on the multispectral imagery, generating a 40-dimensional texture feature vector, where each component corresponds to the response intensity of a specific direction and frequency. For airborne hyperspectral data, wavelet transform maps the signal to different frequency bands through multi-resolution decomposition. A 4-level decomposition is performed using the db4 wavelet basis to extract the energy features and statistical parameters of each sub-band, forming a feature set containing 32-dimensional frequency domain features. The time-series characteristics of ground sensor data are adaptively processed using Empirical Mode Decomposition (EMD). This algorithm does not require preset basis functions and can automatically decompose non-stationary signals into several intrinsic mode functions (IMFs) and residual terms. During implementation, EMD decomposition was performed on 12 types of sensor data, including temperature, humidity, and soil moisture. The energy proportions of the first three IMFs and Hilbert marginal spectral features were extracted to construct a 48-dimensional time-frequency feature vector. Finally, a multi-source feature set containing 120 dimensions was formed by feature concatenation. This feature set not only retains the macroscopic texture information of satellite data and the frequency domain details of aerial data, but also incorporates the dynamic changes of the microenvironment of ground sensors.

[0020] S302. Use the ReliefF algorithm to evaluate the feature weights in the multi-source feature set, and use the t-SNE algorithm to perform nonlinear dimensionality reduction on the multi-source feature set based on the feature weights to obtain the optimized feature subset. In step S302, it is necessary to explain in detail that the existing ReliefF algorithm, as a classic method for feature weight evaluation, dynamically adjusts the importance weight of each feature by calculating its ability to distinguish between similar samples and dissimilar samples. Specifically, for the 120-dimensional features in the multi-source feature set, a sample is first randomly selected as a benchmark. Then, k nearest neighbors (k=5) are found among similar samples and k nearest neighbors are found among dissimilar samples. The distance difference between each feature and similar / dissimilar samples is calculated, and the feature weights are updated through a weighted voting mechanism. This process is repeated 1000 times, ultimately generating a feature importance vector containing 120 weight values, where a larger weight value indicates a stronger ability to distinguish ecological change patterns. Based on the feature weight ranking results, the existing t-SNE algorithm is used for nonlinear dimensionality reduction. This algorithm constructs a probability distribution mapping from a high-dimensional space to a low-dimensional space, preserving local structural features of the data while achieving dimensionality compression. In implementation, the dimensionality reduction target dimension was set to 16, Euclidean distance was used to calculate the similarity between samples, and the KL divergence loss function was optimized using gradient descent. After 500 iterations, an optimized feature subset was generated. This subset significantly reduced computational complexity while retaining key feature components.

[0021] S303. The optimized feature subset is fused across modalities using a multimodal Transformer fusion model to obtain primary fused features; the expression for the Transformer fusion model is: ;in, Represents the primary fusion feature vector. Representation layer normalization, Represents the texture feature vector of satellite imagery. Indicates a feedforward network. Indicates the number of heads of attention. This represents the attention mechanism. Represents the satellite modality query vector. Represents the aerospace mode key vector. Represents the ground modal value vector; In step S303, it is necessary to explain in detail that the multimodal Transformer fusion model achieves cross-modal information interaction by introducing modality-specific query, key, and value vectors. Specifically, the satellite modality query vector retrieves relevant features from the airborne modality key vector through an attention mechanism, and simultaneously combines it with the ground modality value vector for weighted summation to generate an attention weight matrix containing multi-source information. This model employs an 8-head attention mechanism, where each attention head independently learns the association patterns between different modalities. For example, one head may focus on the synergistic changes in vegetation cover and soil moisture, while another head captures the nonlinear relationship between topographic relief and spectral reflectance. A feedforward network is used to perform a nonlinear transformation on the attention output, and residual connections and layer normalization are introduced to enhance the model's ability to fit complex ecological patterns. In application, the hidden layer dimension is set to 256, and the Adam optimizer is used for end-to-end training with a learning rate of 0.001, ultimately generating a 16-dimensional primary fusion feature vector. This vector dynamically adjusts the contribution of each modality feature through a self-attention mechanism, effectively solving the problem of strong subjectivity in modality weight allocation in traditional fusion methods, and significantly improving the robustness of feature representation and ecological semantic consistency.

[0022] S304. By introducing ecological parameters as physical constraints and combining them with the random forest algorithm to establish an empirical model, the primary fusion features are enhanced to obtain the fusion feature vector.

[0023] In step S304, it is necessary to explain in detail that, firstly, key parameters strongly correlated with the monitoring target are selected from ecological theory, such as net primary productivity (NPP), soil organic carbon content (SOC), and albedo, to construct an ecological constraint library containing 12 core parameters. Subsequently, an empirical model is built using the random forest algorithm. This model uses a primary fusion feature vector as input and measured values ​​of ecological parameters as supervision signals, achieving feature enhancement through ensemble learning of 500 decision trees. During implementation, the maximum depth of each tree is set to 15, the Gini index is used as the splitting criterion, and the model complexity is dynamically adjusted through out-of-bag error estimation (OOBError). During training, an ecological consistency loss function is introduced, which consists of two parts: mean squared error loss (MSE) to ensure numerical matching between features and ecological parameters, and structural similarity loss (SSIM) to ensure consistency between the spatial distribution of features and the spatial autocorrelation of ecological parameters. Through 1000 rounds of iterative optimization, the final fused feature vector not only retains the integrity of the original multimodal information, but also embeds clear ecological semantics—for example, the third dimension of the feature vector directly corresponds to the vegetation productivity level, and the seventh dimension reflects the soil carbon cycle intensity.

[0024] As an optional embodiment of the present invention, step S304 may involve introducing ecological parameters as physical constraints and combining them with a random forest algorithm to establish an empirical model, including: S3041. Based on primary fusion features, extract ecological parameter benchmark values, including leaf area index, net primary productivity and carbon storage, and construct a set of physical constraint parameters; In step S3041, it is necessary to explain in detail that the method for extracting the baseline values ​​of ecological parameters in this embodiment is as follows: First, satellite remote sensing inversion technology is used to obtain the large-scale leaf area index (LAI) distribution. This technology generates a leaf area index product with a spatial resolution of 30 meters by regression modeling of multispectral vegetation indices (such as NDVI) and ground-measured data, combined with atmospheric correction and geometric correction processing. For the estimation of net primary productivity (NPP), the CASA model is used. This model integrates photosynthetically active radiation, vegetation indices, temperature stress factors, and water stress factors. By calculating the difference between photosynthetic rate and respiration consumption on a pixel-by-pixel basis, a monthly NPP distribution map is obtained. Carbon storage assessment combines ground plot survey data and LiDAR point cloud data, and uses the random forest algorithm to establish a quantitative relationship model between forest biomass, canopy height, and vegetation indices. Finally, spatial interpolation is used to generate the spatial distribution of regional carbon storage. During implementation, feature components strongly correlated with the aforementioned parameters were extracted from the primary fusion feature vector (e.g., dimensions 2-5 correspond to vegetation spectral features, and dimensions 8-10 correspond to topographic features). Combined with concurrent ground observation data (including measured biomass values ​​from 12 forest plots, temperature and humidity data from 20 meteorological stations, and CO2 flux data from 5 flux towers), a physical constraint parameter set was constructed, encompassing three core parameters: leaf area index, net primary productivity, and carbon storage. This parameter set was validated using K-fold cross-validation (K=5) to ensure parameter estimation accuracy. The R² for leaf area index reached 0.89, the root mean square error (RMSE) for net primary productivity was 1.2 gC / m² / day, and the relative error for carbon storage was controlled within 8%.

[0025] S3042. Based on the physical constraint parameter set, perform Z-score standardization on the primary fusion features to obtain a standardized feature set; It should be noted in step S3042 that the standardization process is to eliminate the influence of different units and numerical ranges on model training, so that each feature is comparable.

[0026] S3043. A random forest model is trained by standardizing the feature set, and a dual-objective loss function is introduced to integrate ecological constraints to obtain an empirical model.

[0027] In step S3043, it should be noted that the random forest model consists of multiple decision trees, and its generalization ability is improved through ensemble learning. During training, a standardized feature set is used as input, and a dual-objective loss function is introduced to incorporate ecological constraints. The dual-objective loss function consists of two parts: one is the traditional mean squared error loss, which measures the numerical difference between the model's predicted values ​​and the true values, ensuring that the model can accurately fit the data; the other is the ecological consistency loss, which is constructed based on ecological principles, such as considering the spatial autocorrelation between ecological parameters and the dynamic changes of ecological processes. By introducing the ecological consistency loss, the model not only focuses on numerical accuracy during training but also follows the inherent logic of ecology. When constructing the random forest model, the number of decision trees is set to 300, the maximum depth of each tree is 20, and information gain is used as the splitting criterion. Through 1000 rounds of iterative training, the model parameters are continuously adjusted until the dual-objective loss function reaches its minimum value. The final empirical model can not only make effective predictions using primary fused features but also ensure that the prediction results conform to ecological laws.

[0028] As an optional embodiment of the present invention, optionally, in step S4, dynamic inference is performed based on the fused feature vector to generate dynamic inference results of ecological changes, including: S401. Construct a dual-model inference architecture based on fused feature vectors. The dual-model inference architecture includes a random forest regression model and an LSTM time series prediction model. In step S401, it is necessary to explain in detail that the design of the dual-model inference architecture aims to fully leverage the advantages of the random forest regression model in handling complex nonlinear relationships, and the ability of the LSTM time series prediction model in capturing the long-term dependency characteristics of time series data. The random forest regression model, by constructing a large number of decision trees and integrating their prediction results, can effectively address nonlinear mapping problems in high-dimensional feature spaces, and is particularly suitable for handling complex feature vectors after the fusion of multi-source heterogeneous data in ecological monitoring. This model generates multiple training subsets through bootstrap sampling, each subset independently constructing a decision tree, and finally determining the regression result through a voting mechanism. This ensemble learning approach significantly improves the model's stability and anti-overfitting ability. The LSTM time series prediction model, through its unique gating mechanism (input gate, forget gate, output gate), achieves long-term memory of time series data, effectively capturing the time-series dependency patterns in the process of ecological change. When constructing the LSTM model, a three-layer stacked structure is adopted to enhance feature extraction capabilities, with each layer containing 64 hidden units. Parameter optimization is performed using the backpropagation algorithm combined with backpropagation time (BPTT). The dual-model architecture employs a parallel extrapolation strategy. A random forest regression model generates spatial ecological change predictions, while an LSTM model focuses on temporal trend extrapolation. Finally, a weighted fusion mechanism integrates the outputs of both models. Weight allocation is dynamically adjusted based on the model's performance on historical data. Cross-validation is used during training to determine the optimal weight ratio, ensuring that the extrapolation results possess both spatial accuracy and temporal continuity. This architecture, by fusing two models with different mechanisms, effectively addresses the limitations of a single model in handling complex spatiotemporal ecological problems, significantly improving the reliability and accuracy of dynamic extrapolation.

[0029] S402. Use the random forest regression model to perform static feature regression analysis on the fused feature vector to obtain the predicted values ​​of ecological parameters. At the same time, use the LSTM time series prediction model to perform time series modeling on the fused feature vector to capture long-term dependencies and generate ecological dynamic trend sequences. In step S402, it is necessary to explain in detail that when performing static feature regression analysis using the random forest regression model, the fused feature vector is first input into the pre-trained random forest regression model. Each decision tree in this model independently makes predictions based on the input features, and finally, the prediction results of all decision trees are combined through a voting mechanism to obtain the predicted values ​​of ecological parameters. These predicted values ​​cover key ecological indicators such as vegetation cover, soil moisture, and biomass. Simultaneously, the LSTM time-series prediction model performs time-series modeling on the fused feature vector. Through its unique gating structure, the LSTM model can effectively capture long-term dependencies in time-series data. During the modeling process, the model learns time-series patterns in historical data and predicts future ecological dynamic trends based on these patterns. Through optimization of the multi-layer stacked structure and backpropagation algorithm, the LSTM model can generate accurate ecological dynamic trend sequences, reflecting the changing patterns of ecological parameters over time.

[0030] S403. Input the predicted values ​​of ecological parameters and the ecological dynamic trend sequence into the integration and fusion layer, and use the weighted average algorithm combined with the model confidence to fuse the results and generate preliminary inference results. In step S403, it is necessary to explain in detail that the integration and fusion layer, as the core component of the dual-model inference architecture, is designed to organically integrate spatial dimension prediction results with temporal dimension trend analysis. Specifically, the ecological parameter predictions (such as vegetation cover and soil moisture) output by the random forest regression model are first aligned with the ecological dynamic trend sequence generated by the LSTM model to ensure consistency in both temporal and spatial resolution. Subsequently, a model confidence evaluation mechanism is introduced. Dynamic weight coefficients are generated by calculating the prediction errors (such as root mean square error (RMSE) and mean absolute error (MAE) of each model on the historical validation set. The weights of the random forest model are determined based on its spatial prediction accuracy, while the weights of the LSTM model are allocated according to its temporal trend fit. For example, when the spatial prediction RMSE of vegetation cover in a certain area is below 0.15, the corresponding weight of the random forest model is set to 0.6, while the weight of the LSTM model can be increased to 0.7 when a significant seasonal change trend is detected. The weighted average algorithm employs an adaptive fusion strategy, emphasizing spatial prediction results during periods of stable ecological parameter changes (weight ratio 6:4) and enhancing the contribution of the time-series model during periods of sudden ecological events (such as drought or fire) (weight ratio 4:6). An ecological consistency verification module is also introduced during the fusion process. By comparing the physical rationality of the fusion results with the set of ecological constraint parameters (such as leaf area index and net primary productivity), outliers are corrected. For example, when the predicted soil moisture exceeds the field capacity limit, the system automatically triggers a constraint correction mechanism to adjust the value to a reasonable range. The final preliminary inference results are stored in the form of a spatiotemporal matrix, with each grid cell containing the predicted values ​​and confidence intervals of 12 key ecological parameters, along with time-series variation curves. This fusion method effectively solves the problems of dimensional differences and logical conflicts between the outputs of multiple models through dynamic weight allocation and physical constraint verification.

[0031] S404. Based on the Bayesian optimization framework, the uncertainty of the preliminary inference results is quantified and calibrated. Ecological prior knowledge is introduced as a regularization constraint to generate calibrated dynamic inference results of ecological changes.

[0032] In step S404, it is necessary to explain in detail that the Bayesian optimization framework quantifies the uncertainty in the preliminary projection results by constructing a probabilistic model. Its core lies in continuously updating the posterior distribution using prior distribution and observational data, thereby obtaining more reliable prediction results. In the dynamic projection scenario of ecological change, each predicted value of ecological parameter in the preliminary projection results is first treated as a random variable, and it is assumed that it follows a Gaussian distribution. The mean and variance of the prior distribution are determined through historical data and expert experience. For example, for the prediction of vegetation cover, the mean of the prior distribution can be set as the regional historical average, and the variance is set according to the spatial variability of vegetation types. Subsequently, ecological prior knowledge is introduced as a regularization constraint. This knowledge includes the reasonable range of ecological parameters (e.g., soil moisture should not exceed field capacity), the synergistic change patterns between parameters (e.g., vegetation cover and biomass are usually positively correlated), and the dynamic equilibrium relationship of ecological processes (e.g., the balance between absorption and release in the carbon cycle). When quantifying uncertainty, a Monte Carlo simulation method is used to randomly sample from the prior distribution to generate a large set of candidate parameters, each representing a possible ecological state. These candidate parameter sets are input into pre-constructed ecological process models (such as the CASA model or the CENTURY model) to simulate the dynamic response of the ecosystem and generate corresponding ecological change sequences. By comparing the simulation results with actual observational data (such as vegetation indices retrieved from satellite remote sensing and CO2 flux data from ground flux towers), the likelihood function value of each candidate parameter set is calculated. This value reflects the degree of consistency between the parameter set and the observational data. Based on Bayes' theorem, and combining the prior distribution and the likelihood function, the posterior probability of each parameter set is updated; parameter sets with higher posterior probabilities are closer to the true ecological state. During the uncertainty calibration phase, outliers in the preliminary simulation results (such as predicted soil moisture exceeding the physically reasonable range) are corrected by introducing an ecological consistency loss function. This loss function consists of two parts: a data fitting term to ensure that the calibrated results are as close as possible to the observed data; and an ecological constraint term to penalize predicted values ​​that violate ecological principles (such as the negative correlation between vegetation cover and biomass). By minimizing the total loss function, the parameter values ​​in the preliminary simulation results are adjusted to conform to both the observed data and ecological prior knowledge. The final calibrated dynamic projection results of ecological changes are presented in the form of probability distributions. Each ecological parameter not only provides a predicted value but also a confidence interval (e.g., a 95% confidence interval), intuitively reflecting the degree of uncertainty in the prediction results. For example, if the predicted vegetation cover value for a certain area is 0.65, and the 95% confidence interval is [0.60, 0.70], it indicates that under the existing data and model conditions, the actual vegetation cover has a 95% probability of falling within this interval. Furthermore, the calibration results also include information on the coordinated changes among ecological parameters, such as the joint distribution of vegetation cover and soil moisture, providing ecological managers with a more comprehensive basis for decision-making.By introducing a Bayesian optimization framework and ecological prior knowledge, this method significantly improves the reliability and ecological rationality of dynamic inference results, and effectively solves the problems of insufficient uncertainty quantification and lack of physical constraints in traditional methods.

[0033] As an optional embodiment of the present invention, the ecological monitoring and early warning based on the dynamic simulation results of ecological changes in step S5 may include: S501. Based on the dynamic extrapolation results of historical ecological changes, a probability distribution model of ecological parameters is constructed through kernel density estimation to generate a baseline threshold range; In step S501, it is necessary to explain in detail that kernel density estimation, as a non-parametric estimation method, can automatically fit the probability density function of ecological parameters based on sample data from historical dynamic ecological change projections without requiring a preset distribution. Specifically, firstly, dynamic ecological change projection data over a historical period is collected and organized. This data covers the predicted values ​​of multiple key ecological parameters at different time points and spatial locations. For example, for the parameter of vegetation cover, monthly predicted values ​​for the past five years are collected. Subsequently, the kernel density estimation algorithm is used to process this data. The algorithm calculates the data density contribution within a certain range around each data point using a kernel function (such as a Gaussian kernel function), and then sums up the contributions of all data points to obtain the probability density curve of the entire ecological parameter. Based on this probability density curve, the baseline threshold intervals at different confidence levels can be determined. For example, selecting a 95% confidence level, the integral probability density curve is used to find two boundary values ​​that make the area under the curve reach 95%. The interval formed by these two boundary values ​​is the baseline threshold interval for the ecological parameter. For multiple ecological parameters, the above process is repeated to obtain the baseline threshold interval for each parameter. These baseline threshold ranges can reflect the fluctuations of ecological parameters within normal ranges of variation. For example, when the predicted vegetation cover value of a certain area exceeds its baseline threshold range, it may mean that the vegetation growth in that area is abnormal and requires further attention and analysis.

[0034] S502. Real-time comparison of the current dynamic ecological change projection results with the baseline threshold range. When the monitored value exceeds the range, a primary warning signal is triggered. In step S502, it is necessary to explain in detail that during real-time ecological monitoring, the system continuously acquires the dynamic projection results of ecological changes at the current moment. These results include the latest predicted values ​​of multiple key ecological parameters. Subsequently, the system compares these real-time predicted values ​​one by one with the baseline threshold intervals generated in advance through kernel density estimation. For example, for the parameter of vegetation cover, the system checks whether its real-time predicted value falls within its corresponding 95% confidence level baseline threshold interval. If the real-time predicted value of a certain ecological parameter exceeds its baseline threshold interval, the system will immediately trigger a primary warning signal. The triggering of a primary warning signal means that the ecological parameter may have undergone abnormal changes, requiring the attention of ecological managers. For example, when the predicted value of soil moisture in a certain area suddenly falls below the lower limit of its baseline threshold interval, the system will issue a primary warning, indicating that the area may be experiencing drought. The triggering of the primary warning signal not only depends on the anomaly of a single parameter, but can also be comprehensively judged by combining the joint changes of multiple parameters. For example, when the predicted value of vegetation cover decreases while the predicted value of soil moisture also falls below the baseline threshold interval, this joint abnormal change will enhance the reliability of the warning, indicating that the area may be facing serious ecological problems. After triggering a primary warning signal, the system will promptly push relevant warning information (including the name of the abnormal parameter, the real-time predicted value, the baseline threshold range, and the warning level) to the ecological manager so that they can take swift countermeasures to prevent further deterioration of the ecological problem.

[0035] S503. The gradient change rate of the monitored value is calculated through a sliding window mechanism. If the gradient change rate of three consecutive time windows exceeds the preset mutation threshold, a secondary mutation warning signal is triggered. In step S503, it is necessary to explain in detail that in the ecological monitoring and early warning scenario, the size of the sliding window must first be determined. This size should be reasonably set according to the change cycle of the ecological parameters and the monitoring needs. For example, for the parameter of vegetation cover, if its change cycle is in months, the sliding window size can be set to one month, that is, the data changes within one month are analyzed each time. After determining the sliding window size, the system will slide the window sequentially according to time sequence and calculate the gradient rate of change of the monitored values ​​within each window. The gradient rate of change reflects the speed and direction of change of the monitored values ​​within the window, and can be obtained by calculating the difference between the monitored value at the last time point and the monitored value at the first time point within the window, and then dividing it by the window time length. For example, for the vegetation cover monitoring data of a certain month, if the predicted value at the beginning of the month is 0.5 and the predicted value at the end of the month is 0.6, then the gradient rate of change within this window is (0.6-0.5) / 1=0.1 (assuming the time unit is months), indicating that the vegetation cover shows an upward trend within that month, and the rate of increase is 0.1 per month. The system will continuously perform sliding window operations and calculate the gradient rate of change for each window. If the gradient change rate exceeds a preset mutation threshold for three consecutive time windows, it indicates that the ecological parameter has undergone a significant abrupt change during that period. For example, if the preset mutation threshold is 0.15, and the gradient change rates of a certain ecological parameter for three consecutive months are 0.2, 0.25, and 0.3, all exceeding the threshold, the system will trigger a secondary mutation warning signal. The triggering of a secondary mutation warning signal signifies that the ecological parameter may have experienced a sudden and drastic change, often closely related to ecological disasters or emergencies. For example, if soil moisture decreases rapidly for three consecutive months, exceeding a preset threshold, it may indicate a severe drought in the region. After triggering the secondary mutation warning signal, the system will immediately push relevant warning information (including the name of the abnormal parameter, the gradient change rate, the warning level, and potential ecological problems) to ecological managers and recommend emergency response measures to mitigate or avoid losses caused by ecological disasters.

[0036] S504 integrates primary warning signals and secondary abrupt change warning signals, generates a comprehensive warning level through evidence theory synthesis rules, and outputs a warning report that includes spatiotemporal location, abnormal parameter type, and warning level.

[0037] In step S504, it is necessary to explain in detail that evidence theory, as an effective method for handling uncertain information, can fuse warning signals from different sources to generate a more accurate and comprehensive integrated warning level. When fusing primary warning signals and secondary abrupt change warning signals, it is first necessary to clarify the basic probability allocation represented by each warning signal. Primary warning signals mainly reflect whether ecological parameters exceed the normal fluctuation range, and their basic probability allocation can be determined based on the degree to which the parameter exceeds the baseline threshold range. For example, when the parameter exceeds the range but the deviation is small, a lower basic probability value can be assigned; when the deviation is large, a higher basic probability value is assigned. Secondary abrupt change warning signals focus on abrupt changes in ecological parameters, and their basic probability allocation can be determined based on the degree to which the gradient change rate exceeds the preset abrupt change threshold. For example, when the gradient change rate slightly exceeds the threshold, the basic probability value is low; when it significantly exceeds the threshold, the basic probability value is high. After obtaining the basic probability allocations of the two warning signals, they are fused using the synthesis rules of evidence theory. The synthesis rules determine the comprehensive warning level by calculating the trust function and likelihood function under different combinations of warning signals. Specifically, all possible combinations of primary and secondary mutation warning signals are listed, such as (low primary warning, no secondary mutation), (low primary warning, low secondary mutation), (high primary warning, high secondary mutation), etc. For each combination, the corresponding confidence function and likelihood function are calculated. The confidence function reflects the degree of confidence that the combination is true, while the likelihood function reflects the degree to which the combination is not rejected. By comparing the confidence function and likelihood function values ​​of different combinations, the most likely comprehensive warning level is determined. For example, if the confidence function and likelihood function values ​​of the (high primary warning, high secondary mutation) combination are significantly higher than those of other combinations, the comprehensive warning level is determined to be high. The final warning report not only includes spatiotemporal location information, clearly indicating the specific area and time of the anomaly, but also lists the types of anomaly parameters in detail, such as vegetation cover and soil moisture. Simultaneously, the warning report provides a comprehensive warning level, such as low, medium, or high, providing clear decision-making basis for ecological managers. For example, when a region simultaneously triggers a primary warning and a secondary abrupt change warning, and the overall warning level is high, ecological managers should immediately activate the emergency response mechanism and take urgent measures to prevent the further expansion of the ecological disaster. Warning reports generated using evidence theory synthesis rules can effectively integrate multi-source warning information, improving the accuracy and reliability of warnings.

[0038] As an optional embodiment of the present invention, the decision support based on the dynamic simulation results of ecological changes in step S5 may include: A501. Extract key driving factors from the dynamic inference results of ecological changes, associate them with entity relationships in the ontology library of the ecological field, and construct a multidimensional causal reasoning network. In step A501, it is necessary to explain in detail that the extraction of key driving factors is the foundation for constructing a multidimensional causal inference network. These factors are elements that significantly influence ecological change. For example, when studying the dynamic changes of forest ecosystems, precipitation, temperature, light intensity, and the intensity of human activities can all become key driving factors. Through data mining and statistical analysis methods, these influential factors are selected from the dynamic inference results of ecological change. An ecology ontology is a knowledge base containing ecology-related concepts, entities, and their interrelationships. The extracted key driving factors are associated with entities in the ontology to clarify their position and role in the ecological knowledge system. For example, precipitation is associated with the "climate factor" entity in the ontology, and the intensity of human activities is associated with the "anthropogenic disturbance factor" entity. Based on this association, a multidimensional causal inference network is constructed. This network uses key driving factors as nodes and the causal relationships between them as edges, intuitively demonstrating the complex causal mechanisms behind ecological change. For example, in a forest ecosystem, reduced rainfall may lead to decreased soil moisture, which in turn affects the growth of vegetation. This series of causal relationships is clearly presented in a multidimensional causal reasoning network.

[0039] A502. Based on a multidimensional causal reasoning network, a decision-making strategy model is trained using a reinforcement learning framework. The optimal decision path set is generated with ecological restoration cost-benefit ratio and ecological stability gain as the dual objective reward function. The expression for the bi-objective reward function is: ;in, This represents the value of the bi-objective reward function. and Indicates the weighting coefficient. It represents the comprehensive benefits of ecological restoration (such as the quantitative value of ecosystem services like increased carbon sequestration and enhanced biodiversity). This represents the total cost of ecological restoration projects (including monetized costs such as manpower, materials, and time). Indicators representing ecological stability during the historical baseline period, Indicators representing current ecological stability (such as Shannon diversity index, vegetation cover, etc.); In step A502, it needs to be explained in detail that the reinforcement learning framework continuously tries and optimizes its strategy through the interaction between the agent and the environment to maximize cumulative rewards. In this scenario, the agent represents an ecosystem manager, the environment is a complex ecosystem, and the reward function is composed of the cost-benefit ratio of ecosystem restoration and the gain of ecosystem stability. When training the decision-making policy model, it is first necessary to set reasonable weight coefficients. and These two coefficients reflect the importance that ecological managers place on cost-effectiveness and ecological stability, and can be adjusted according to actual needs. Subsequently, the agent explores different decision paths in a multi-dimensional causal reasoning network, with each path representing a possible ecological restoration solution. For each path, the corresponding comprehensive ecological restoration benefit is calculated, including the quantified value of ecosystem services such as carbon sequestration increment and biodiversity enhancement, as well as the total cost of the ecological restoration project, encompassing monetized costs such as manpower, materials, and time. Simultaneously, the difference between the current ecological stability index and the historical baseline ecological stability index is calculated to measure the gain in ecological stability. Substituting these values ​​into the bi-objective reward function yields the reward value for that decision path. By continuously trying different decision paths and adjusting its strategy based on the reward value, the agent gradually converges to the optimal decision path set. These optimal paths not only consider the cost-effectiveness of ecological restoration but also take into account the improvement of ecological stability, providing ecological managers with a scientific and reasonable basis for decision-making. For example, when facing the problem of forest degradation, the optimal decision path set may include a variety of options such as afforestation, reducing human interference, and adjusting forestry management policies. Each option has its specific cost-effectiveness and ecological stability gains, and ecological managers can choose the most suitable option to implement based on the actual situation.

[0040] A503. Using natural language generation technology, the decision paths in the optimal decision path set are transformed into executable proposed solutions, including priority planning for ecological restoration projects, dynamic adjustment strategies for protected areas, and resource allocation schemes.

[0041] In step A503, it is necessary to explain in detail that natural language generation technology can transform complex technical data and model outputs into easily understandable and actionable textual descriptions. When converting decision paths from the optimal decision path set into executable proposed solutions, the system first conducts a detailed analysis of each decision path, extracting key information such as the specific type, scope, and expected effects of the ecological restoration project; the adjustment direction and scope of the protected area; and the types and quantities of resources required. Subsequently, the system uses natural language generation algorithms, combined with professional terminology and expression habits in the field of ecology, to integrate this key information into a clear and logically coherent proposed solution. For example, regarding the decision-making path of afforestation, the system might generate the following suggested solution: "It is recommended to implement an afforestation project in the XX area, prioritizing native suitable tree species. The estimated afforestation area is XX hectares, the estimated investment cost is XX million yuan, and it is expected to increase carbon sequestration by XX tons, while also helping to improve the vegetation cover and biodiversity of the area. During implementation, attention should be paid to reasonably arranging the construction time to avoid damage to the local ecological environment." Regarding the dynamic adjustment strategy for protected areas, the system might suggest: "Based on a comparative analysis of current ecological stability indicators and historical baselines, it is recommended to expand the XX protected area by XX kilometers in the XX direction to better protect the rare species and ecosystems of the area. At the same time, it is necessary to..." Regular monitoring and evaluation of the adjusted protected areas are necessary to ensure their ecological functions are effectively utilized. Regarding resource allocation plans, the system might suggest: "Based on the priority and actual needs of each ecological restoration project, it is recommended that limited resources (such as funds, manpower, and materials) be prioritized for allocation to Project XX and Project XX to ensure their smooth implementation and achievement of expected results. Simultaneously, a dynamic management mechanism for resource allocation needs to be established to adjust the resource allocation plan promptly based on project progress and actual needs." The actionable suggestions generated through natural language processing technology not only provide ecological managers with intuitive and easily understandable decision-making support but also help improve the scientific rigor and rationality of decision-making, promoting the smooth progress of ecological restoration work.

[0042] Example 2 An integrated air-space-ground ecological monitoring system based on multi-source data fusion includes: processor; Memory used to store processor-executable instructions; The processor is configured to implement an integrated air-space-ground ecological monitoring method based on multi-source data fusion when executing executable instructions.

[0043] It should be noted that the computer device includes a processor, a memory, and may also include one or more of a multimedia component, an input / output (I / O) interface, and a communication component.

[0044] The processor controls the overall operation of the computer device to complete all or part of the steps in the above-mentioned integrated air-space-ground ecological monitoring method based on multi-source data fusion.

[0045] Memory is used to store various types of data to support the operation of the computer device. This data may include, for example, instructions for any application or method used to operate on the computer device, as well as application-related data. Memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0046] The multimedia component may include a screen and an audio component, wherein the screen may be, for example, a touch screen, and the audio component is used to output and / or input audio signals; for example, the audio component may include a microphone for receiving external audio signals, the received audio signals may be further stored in memory or transmitted via a communication component; the audio component may also include at least one speaker for outputting audio signals.

[0047] I / O interfaces provide interfaces between the processor and other interface modules, such as keyboards, mice, buttons, etc.; these buttons can be virtual buttons or physical buttons.

[0048] The communication component is used for wired or wireless communication between the computer device and other devices; wireless communication, such as Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, 4G or 5G, or one or more combinations thereof, and the corresponding communication component may include: Wi-Fi module, Bluetooth module, NFC module, mobile communication module.

[0049] As a preferred embodiment, the computer device may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to execute the above-described integrated air-space-ground ecological monitoring method based on multi-source data fusion.

[0050] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A space-air-ground integrated ecological monitoring method based on multi-source data fusion, characterized in that, The method comprises: S1, collecting satellite raw data, aerial raw data and ground sensor data to obtain a multi-source heterogeneous data set; S2, spatio-temporal alignment of the multi-source heterogeneous data set to obtain spatio-temporal sequence aligned data; S3, multi-scale fusion of the spatio-temporal sequence aligned data to obtain a fusion feature vector; S4, dynamic deduction based on the fusion feature vector to generate an ecological change dynamic deduction result; S5, ecological monitoring and early warning and decision support based on the ecological change dynamic deduction result, wherein the ecological monitoring and early warning comprises real-time monitoring and early warning of abnormal changes in the ecological environment, and the decision support comprises providing targeted ecological protection and management suggestions according to the ecological change dynamic deduction result. 2.The space-ground integration ecological monitoring method based on multi-source data fusion of claim 1, wherein, In step S2, the multi-source heterogeneous data set is spatio-temporally aligned to obtain spatio-temporal sequence aligned data; comprising: S201, using a federated learning framework to clean the multi-source heterogeneous data set, and performing encrypted sample alignment, normalization and dimension reduction to obtain a cleaned and enhanced data subset; S202, constructing a distance matrix based on Euclidean distance to calculate the similarity of the cleaned and enhanced data subset to obtain a similarity matrix; S203, performing a spatio-temporal alignment operation based on the similarity matrix to match data of different sources and different times in the time and space dimensions to obtain spatio-temporal sequence aligned data. 3.The space-ground integration ecological monitoring method based on multi-source data fusion of claim 2, wherein, In step S201, the mathematical expression of the federated learning framework is: in, This represents the federal cleaning framework function. Indicates the federal batch normalization layer. This indicates the number of sources for a multi-source heterogeneous dataset. This represents a privacy set intersection protocol based on elliptic curve Diffie-Hellman. Indicates the first The original dataset of each client, Indicates the use of a public key Perform RSA encryption. Represents the normalized function. Indicates client Original dataset Specific numerical characteristics in Indicates client The average of the data. Indicates client standard deviation This indicates dimensionality reduction using kernel principal component analysis. Represents the kernel function. 4.The space-ground integration ecological monitoring method based on multi-source data fusion of claim 1, wherein, In step S3, the spatio-temporal sequence aligned data is multi-scale fused to obtain a fusion feature vector, comprising: S301, extracting satellite image texture features in the satellite raw data through a Gabor filter, extracting aviation hyperspectral data frequency domain features in the aerial raw data through wavelet transform, and analyzing the internal mode components of the ground sensor data through empirical mode decomposition to obtain a multi-source feature set; S302, evaluating the feature weights in the multi-source feature set using a ReliefF algorithm, and performing nonlinear dimension reduction on the multi-source feature set based on the feature weights using a t-SNE algorithm to obtain an optimized feature subset; S303, performing cross-modal feature fusion on the optimized feature subset through a multi-modal Transformer fusion model to obtain primary fusion features; S304, introducing ecological parameters as physical constraints, combining a random forest algorithm to establish an empirical model, and performing enhancement processing on the primary fusion features to obtain a fusion feature vector. 5.The space-ground integration ecological monitoring method based on multi-source data fusion of claim 4, wherein, The expression of the Transformer fusion model is: wherein, represents a primary fusion feature vector, represents a layer normalization, represents a satellite image texture feature vector, represents a feed-forward network, represents a number of attention heads, represents an attention mechanism, represents a satellite modality query vector, represents an aerial modality key vector, represents a ground modality value vector. 6.The space-ground integration ecological monitoring method based on multi-source data fusion of claim 4, wherein, In step S304, the introduction of ecological parameters as physical constraints, combined with a random forest algorithm to establish an empirical model, comprises: S3041, extracting ecological parameter benchmark values based on the primary fusion features, including leaf area index, net primary productivity and carbon storage, to construct a physical constraint parameter set; S3042, performing Z-score standardization processing on the primary fusion features based on the physical constraint parameter set to obtain a standardized feature set; S3043, training a random forest model through the standardized feature set and introducing a double-target loss function to fuse ecological constraints to obtain an empirical model. 7.The space-ground integration ecological monitoring method based on multi-source data fusion of claim 1, wherein, In step S4, dynamic deduction is performed based on the fusion feature vector to generate an ecological change dynamic deduction result, which includes: S401, constructing a double-model deduction architecture based on the fusion feature vector, the double-model deduction architecture including a random forest regression model and an LSTM time series prediction model; S402, performing static feature regression analysis on the fusion feature vector using the random forest regression model to obtain an ecological parameter prediction value, and simultaneously performing time series modeling on the fusion feature vector using the LSTM time series prediction model to capture long-term dependencies and generate an ecological dynamic trend sequence; S403, inputting the ecological parameter prediction value and the ecological dynamic trend sequence into an integrated fusion layer, combining model confidence through a weighted average algorithm to fuse the results and generate a preliminary deduction result; S404, quantifying and calibrating the uncertainty of the preliminary deduction result based on a Bayesian optimization framework, introducing ecological prior knowledge as a regularization constraint to generate a calibrated ecological change dynamic deduction result. 8.The space-ground integration ecological monitoring method based on multi-source data fusion of claim 1, wherein, In step S5, ecological monitoring and early warning based on the ecological change dynamic deduction result includes: S501, based on historical ecological change dynamic deduction results, constructing a probability distribution model of ecological parameters through kernel density estimation to generate a baseline threshold interval; S502, real-time comparison of the current ecological change dynamic deduction result with the baseline threshold interval, triggering a primary warning signal when the monitoring value exceeds the interval range; S503, calculating the gradient change rate of the monitoring value through a sliding window mechanism, and if the gradient change rate of three consecutive time windows exceeds a preset mutation threshold, triggering a secondary mutation warning signal; S504, fusing the primary warning signal and the secondary mutation warning signal, generating a comprehensive warning level through evidence theory synthesis rules, and outputting a warning report containing the spatio-temporal location, abnormal parameter type and warning level. 9.The space-ground integration ecological monitoring method based on multi-source data fusion of claim 1, wherein, In step S5, decision support based on the ecological change dynamic deduction result includes: A501, extracting key driving factors from the ecological change dynamic deduction result, associating entity relationships in the ecological domain ontology library, and constructing a multi-dimensional causal reasoning network; A502, training a decision strategy model based on the multi-dimensional causal reasoning network using a reinforcement learning framework, taking the ecological restoration cost-benefit ratio and ecological stability gain as a double-target reward function to generate an optimal decision path set; A503, converting the decision paths in the optimal decision path set into executable recommendation schemes through natural language generation technology, including ecological restoration engineering priority planning, protection area dynamic adjustment strategy and resource allocation scheme.

10. A space-air-ground integrated ecological monitoring system based on multi-source data fusion, characterized in that, The system includes: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to implement the space-ground-terrestrial integrated ecological monitoring method based on multi-source data fusion of any one of claims 1-9 when executing the executable instructions.

Images