An AI-based ecological green land dynamic monitoring method and system

Abstract

The present application relates to the technical field of wisdom management and maintenance, in particular to an ecological green land dynamic monitoring method and system based on AI, comprising the following steps: using a deep learning algorithm to analyze the nonlinear correlation between the growth state of vegetation and environmental stress factors, generating habitat health degree evaluation results, calculating the growth deviation vector of the current state characteristics and the standard habitat model, generating corresponding precise management and maintenance control instructions based on the deviation calculation results and feeding back the on-site execution terminal. In the present application, by establishing a deep fusion analysis mechanism of multi-modal data, intelligent perception and decision-making of the whole life cycle of pocket parks and three-dimensional green space are realized, breaking the limitations of traditional manual inspection in time and space coverage, solving the technical problems of fragmented green habitat data islandization and management and maintenance lag, using dynamic monitoring and real-time feedback closed loop to significantly improve the timeliness of pest warning and the accuracy of resource allocation, ensuring the biodiversity and functional stability of ecological space.

Description

Technical Field

[0001] This invention relates to the field of smart management technology, and in particular to an AI-based method and system for dynamic monitoring of ecological green spaces. Background Technology

[0002] The field of smart management technology involves the digital management and ecological function maintenance of urban green space resources. It mainly encompasses the use of the Internet of Things, remote sensing monitoring, and big data analytics to monitor and maintain urban ecological spaces throughout their entire lifecycle. Traditional dynamic monitoring methods for ecological green spaces rely on regular manual inspections or simple fixed-point sensors to collect temperature and humidity data. These data are then manually recorded and judged based on experience to assess the growth status of the green space and formulate maintenance plans. This approach is primarily used in routine surface greening management and basic irrigation control scenarios.

[0003] Existing technologies typically rely on regular manual inspections or distributed sensors for environmental monitoring. They lack unified integration and in-depth mining of multi-source heterogeneous data such as satellite remote sensing, ground-based IoT sensing, and video surveillance. This results in the inability to construct nonlinear correlation models between plant growth status and microenvironmental changes in real time. Furthermore, the lagging information feedback mechanism makes it difficult to trigger precise intervention measures in the early stages of pests, diseases, or droughts. Consequently, the allocation of green space management resources is uneven, failing to meet the high standards of biodiversity conservation and refined management required in complex habitat scenarios such as pocket parks and vertical greening. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an AI-based method and system for dynamic monitoring of ecological green spaces.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: an AI-based dynamic monitoring method for ecological green spaces, comprising the following steps:

[0006] S1: Collect multispectral remote sensing images, ground environmental parameters and vegetation morphology video streams through sensing devices, use edge computing nodes to synchronize and align data based on timestamps and geographic coordinates, establish correlation logic based on spectral reflectance and soil physicochemical properties, and construct the original ecological dataset.

[0007] S2: Use a convolutional neural network to perform feature extraction and dimensionality reduction on the original ecological dataset, analyze the nonlinear coupling relationship between vegetation growth status and microenvironment factors, construct a multimodal feature matrix, input it into a pre-trained habitat health assessment model, and output the habitat health index.

[0008] S3: Compare the habitat health index with the standard habitat threshold, calculate the growth deviation vector, use an anomaly detection algorithm to identify mutation features in the multimodal feature matrix, combine the growth deviation vector to determine the comprehensive disaster risk, and generate an anomaly warning signal;

[0009] S4: Based on the abnormal early warning signal analysis of maintenance needs, use the decision tree algorithm to match precise management strategies, integrate irrigation duration, pesticide concentration and inspection path parameters, construct equipment control commands, and send them to automated management equipment for execution.

[0010] As a further aspect of the present invention, step S1 specifically comprises:

[0011] S11: Activate the heterogeneous sensing devices deployed in the monitoring area, and synchronously acquire the multispectral remote sensing images, the ground environmental parameters, and the vegetation morphology video stream according to the preset high-frequency sampling cycle, and extract the acquisition timestamp and spatial latitude and longitude coordinate information from various data metadata.

[0012] S12: Use the edge computing node to perform spatiotemporal calibration on various heterogeneous data, remove invalid data frames with time deviations exceeding a preset threshold, and fill in the missing time-series data points through a linear interpolation algorithm to generate a spatiotemporally aligned basic data sequence.

[0013] S13: Call the pre-stored spectral inversion logic library, parse the band reflectance values ​​in the multispectral remote sensing image, establish the mapping relationship between them and the soil nitrogen, phosphorus and potassium content and water saturation in the ground environmental parameters, integrate the aligned image data, environmental values ​​and mapping relationship labels, and construct the original ecological dataset.

[0014] As a further aspect of the present invention, step S2 specifically comprises:

[0015] S21: Input the original ecological dataset into a multi-layer convolutional neural network, use convolutional kernels to extract texture features from the image data and perform max pooling operation, while normalizing and encoding the environmental parameter sequence to generate a high-dimensional hybrid feature vector.

[0016] S22: Analyze the dependency weights between elements in the high-dimensional hybrid feature vector through the attention mechanism, quantify the interaction strength between the vegetation growth status and the microenvironment factors such as light, temperature and humidity, analyze the nonlinear coupling characteristics, and construct the multimodal feature matrix;

[0017] S23: Load the habitat health assessment model, map the multimodal feature matrix to the preset health score space, calculate the ecological vitality value and stress resistance score of the current monitoring area through a fully connected layer, and output the habitat health index after weighted fusion.

[0018] As a further aspect of the present invention, step S3 specifically comprises:

[0019] S31: Obtain the standard habitat threshold corresponding to the current season, calculate the absolute value and rate of change of the difference between each indicator in the habitat health index and the standard habitat threshold, and generate the growth deviation vector;

[0020] S32: Apply the isolated forest anomaly detection algorithm to traverse and scan the multimodal feature matrix, identify abrupt changes that deviate from the center of the normal distribution, and mark the environmental factor type and abrupt change intensity level corresponding to the abrupt changes.

[0021] S33: Based on the mutation intensity level, the growth deviation vector is risk-weighted to comprehensively assess the probability of pest and disease outbreaks and the risk of water and fertilizer imbalance in the current green space, determine the level of the comprehensive disaster risk, and generate the abnormal early warning signal.

[0022] As a further aspect of the present invention, step S4 specifically comprises:

[0023] S41: Analyze the risk type code and disaster severity in the abnormal warning signal, retrieve the preset maintenance knowledge graph, determine the resource type and work urgency required to eliminate the current risk, and generate maintenance requirements;

[0024] S42: Input the maintenance requirements into the logical decision tree model, traverse the historical treatment cases and expert rules stored in the leaf nodes, match the optimal precision management strategy, and clarify the corresponding operation mode and resource allocation.

[0025] S43: Calculate the specific irrigation duration, pesticide concentration, and patrol path parameters of the unmanned equipment according to the precision management strategy, encapsulate them into a binary data packet conforming to the industrial control protocol, construct the equipment control command, and issue it.

[0026] As a further aspect of the present invention, the calculation process of the habitat health index specifically includes:

[0027] Based on the internal parameters of the habitat health assessment model, the habitat health index is calculated according to the following formula:

[0028] ;

[0029] in, This represents the calculated habitat health index. Represents the total number of feature dimensions. Representing the The importance weight coefficients of each feature dimension Represents the first in the multimodal feature matrix Normalized values ​​of biomass characteristics in dimensionality. Stability score representing soil physicochemical properties. and These represent the model hyperparameters for biomass characteristics and environmental stability, respectively. This represents a factor that regulates environmental sensitivity.

[0030] As a further aspect of the present invention, the process for determining the comprehensive disaster risk specifically includes:

[0031] Using the aforementioned growth deviation vector and mutation characteristics, the disaster risk value is calculated and the level is determined according to the following logic:

[0032] ;

[0033] in, The quantitative value representing the comprehensive disaster risk, This represents the growth deviation vector. This represents the transpose of the growth deviation vector. The inverse matrix representing the covariance matrix of historical disaster data. The weights of the influence factors representing the characteristics of mutations This represents the set of anomalous amplitudes for each feature point in the mutation feature set. This represents the maximum abnormal amplitude in the set of abnormal amplitudes;

[0034] The system according to The final risk level is determined by the preset range in which the value falls.

[0035] As a further aspect of the present invention, the construction process of the spectral inversion logic library specifically includes:

[0036] Hyperspectral images of the monitoring area and synchronously measured soil physicochemical data were acquired within a historical time period. The correlation coefficient between spectral bands and soil parameters was analyzed using partial least squares regression algorithm.

[0037] Feature bands with correlation coefficients greater than the preset confidence level were selected, and a multiple linear regression equation was established between the reflectance of the feature bands and soil moisture content and organic matter content.

[0038] The prediction accuracy of the multiple linear regression equation is verified using an independent validation dataset. Equations with a standard coefficient of determination are solidified into inversion rules in the spectral inversion logic library for subsequent real-time data analysis.

[0039] As a further aspect of the present invention, the training process of the logical decision tree model specifically includes:

[0040] Collect disaster event records, environmental parameters at the time, response measures taken, and post-response recovery data from the green space management historical database to construct a training sample set;

[0041] Using disaster type and environmental characteristics as input nodes and the effectiveness of response measures as the splitting criterion, the information gain algorithm is used to construct the hierarchical structure of the decision tree, and overfitting branches are pruned and optimized.

[0042] The optimized tree structure is transformed into an executable rule engine, and the weight parameters of the leaf nodes are dynamically updated based on the new management feedback data, thus completing the deployment of the logical decision tree model.

[0043] An AI-based dynamic monitoring system for ecological green spaces, the system being used to implement the aforementioned AI-based dynamic monitoring method for ecological green spaces, the system comprising:

[0044] The ecological data perception module is used to collect multispectral remote sensing images, ground environmental parameters and vegetation morphology video streams through sensing devices. It uses edge computing nodes to synchronize and align data based on timestamps and geographic coordinates, and establishes correlation logic based on spectral reflectance and soil physicochemical properties to construct the original ecological dataset.

[0045] The habitat health assessment module is used to perform feature extraction and dimensionality reduction on the original ecological dataset using a convolutional neural network, analyze the nonlinear coupling relationship between vegetation growth status and microenvironmental factors, construct a multimodal feature matrix, input it into a pre-trained habitat health assessment model, and output a habitat health index.

[0046] The disaster risk early warning module is used to compare the habitat health index with the standard habitat threshold, calculate the growth deviation vector, use an anomaly detection algorithm to identify the mutation features in the multimodal feature matrix, combine the growth deviation vector to determine the comprehensive disaster risk, and generate an anomaly early warning signal.

[0047] The automated maintenance execution module is used to analyze maintenance needs based on the abnormal early warning signal, match precise maintenance strategies using a decision tree algorithm, integrate irrigation duration, pesticide concentration and inspection path parameters, construct equipment control commands, and send them to the automated maintenance equipment for execution.

[0048] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0049] This invention establishes a deep fusion and analysis mechanism for multi-source heterogeneous data, enabling full-dimensional perception and dynamic reconstruction of the habitat status of ecological green spaces. It utilizes intelligent algorithms to calculate the optimal management strategy in real time, breaking the limitations of monitoring single environmental factors and solving the problems of fragmented data collection and delayed decision response in traditional methods. By using the calculation logic of growth deviation vector, it significantly improves the accuracy and response speed of pest and disease early warning, ensuring the dynamic balance and precise maintenance of key ecological functional areas, and providing a standardized and intelligent construction path for complex urban green space scenarios. Attached Figure Description

[0050] Figure 1 This is a flowchart of the AI-based dynamic monitoring method for ecological green spaces according to the present invention;

[0051] Figure 2 Flowchart for constructing the original ecological dataset for this invention;

[0052] Figure 3 This is a flowchart of the habitat health index calculation process of the present invention;

[0053] Figure 4 This is a flowchart illustrating the comprehensive disaster risk assessment process of this invention.

[0054] Figure 5 This is a flowchart illustrating the precise maintenance strategy matching and execution process of this invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this invention clearer, the software-based technical solution is described in detail below with reference to system architecture diagrams and embodiments. It should be understood that the specific embodiments described herein are only for explaining the technical solutions of this invention and do not constitute a limitation on the scope of protection.

[0056] In the description of this invention, the system architecture relationships or data processing flows indicated by terms such as "layer," "module," "interface," "data flow," "client," and "server" are all defined based on the architecture diagram or flowchart corresponding to the embodiments. This way of describing is only used to clearly illustrate the logical relationships between the elements in the technical solution, and not to limit the physical deployment form. The term "multiple" includes two or more technical units, including but not limited to multiple data nodes, processing threads, service instances, or functional components and other scalable elements. The specific number is determined according to the actual business scenario and needs to be specifically specified.

[0057] Please see Figure 1 and Figure 2 This invention provides a technical solution: an AI-based dynamic monitoring method for ecological green spaces, comprising the following steps:

[0058] S1: Collect multispectral remote sensing images, ground environmental parameters, and vegetation morphology video streams through sensing devices. Use edge computing nodes to synchronize and align data based on timestamps and geographic coordinates. Establish correlation logic based on spectral reflectance and soil physicochemical properties to construct the original ecological dataset.

[0059] S11: Activate the heterogeneous sensing devices deployed in the monitoring area, and synchronously acquire multispectral remote sensing images, ground environmental parameters, and vegetation morphology video streams according to the preset high-frequency sampling cycle, and extract the acquisition timestamps and spatial latitude and longitude coordinates from various data metadata.

[0060] S12: Utilize edge computing nodes to perform spatiotemporal calibration on various heterogeneous data, remove invalid data frames with time deviations exceeding a preset threshold, and fill in missing time-series data points through linear interpolation algorithms to generate a spatiotemporally aligned basic data sequence.

[0061] S13: Call the pre-stored spectral inversion logic library, parse the band reflectance values ​​in the multispectral remote sensing image, establish the mapping relationship between them and the soil nitrogen, phosphorus and potassium content and water saturation in the ground environmental parameters, integrate the aligned image data, environmental values ​​and mapping relationship labels, and construct the original ecological dataset.

[0062] The construction process of the spectral inversion logic library specifically includes:

[0063] Hyperspectral images of the monitoring area and synchronously measured soil physicochemical data were acquired within a historical time period. The correlation coefficient between spectral bands and soil parameters was analyzed using partial least squares regression algorithm.

[0064] Feature bands with correlation coefficients greater than the preset confidence level were selected, and a multiple linear regression equation was established between the reflectance of the feature bands and soil moisture content and organic matter content.

[0065] The prediction accuracy of the multiple linear regression equation is verified using an independent validation dataset. Equations with a standard coefficient of determination are solidified into inversion rules in the spectral inversion logic library for subsequent real-time data analysis.

[0066] A heterogeneous sensing network built on the ecological green space was initiated and initialized. This network includes a pan-tilt multispectral camera mounted 10 meters above the ground, a distributed soil sensor array distributed on the surface of the green space and 20 centimeters below ground, and fixed-focus high-definition monitoring cameras covering the entire area. A unified Network Time Protocol (NTP) synchronization command was issued through the edge gateway to calibrate the system time error of all sensing devices to within 1 millisecond. The acquisition band of the multispectral remote sensing images was set to cover the visible to near-infrared light region (450 nm to 950 nm), and the acquisition channels for ground environmental parameters, including soil temperature, soil moisture, conductivity, and pH value, were configured. After the data transmission link was established, the raw data packets returned by each sensing terminal were received in real time. The received data streams were unpacked and cleaned using the edge computing nodes, and the data sources from different locations were spatially gridded and categorized according to the preset geographic partitioning code. Based on the correlation logic between spectral reflectance and soil physicochemical properties (such as total nitrogen, available phosphorus, and organic matter) that has been pre-determined and established in the laboratory, a basic feature database is loaded to lay the groundwork for the subsequent construction of a raw ecological dataset containing spatiotemporal dimensions.

[0067] The aforementioned heterogeneous sensing network refers to a comprehensive monitoring system composed of different types of sensors (such as optical cameras, electrochemical sensors, environmental monitoring stations, etc.). The sensors differ in data format, sampling frequency, and communication protocol, and need to be accessed and managed uniformly through middleware technology.

[0068] Activate the heterogeneous sensing device cluster deployed within the monitoring area to execute high-frequency concurrent data acquisition tasks. Send trigger commands to the multispectral imaging device to capture remote sensing images containing five bands—red, green, blue, red-edge, and near-infrared—at a sampling frequency of once every 10 minutes, with the image resolution set to 1920×1080 pixels. Simultaneously, instruct the ground sensor network to continuously read soil temperature, humidity, and nitrogen, phosphorus, and potassium electrochemical signals at a frequency of 1 Hz; instruct the high-definition camera to record a video stream of vegetation morphology under wind or light changes at a frame rate of 30 frames per second. Parse the header of each frame of the raw data packet to extract its metadata segment. Specifically, read the "DateTimeOriginal" field from the Exchangeable Image File Format (EXIF) information or JSON message header as the acquisition timestamp (accurate to milliseconds), and read the "Latitude" and "Longitude" fields written by the Global Positioning System (GPS) module as spatial latitude and longitude coordinate information (accurate to 6 decimal places). For example, the timestamp "2025-06-15T10:30:00.050Z" and coordinates "31.230416,121.473701" are extracted from a multispectral image frame, and this spatiotemporal label is used as the unique index key value of the data frame and stored in a temporary cache queue for subsequent processing.

[0069] High-performance processors on edge computing nodes are used to read various heterogeneous data from temporary cache queues. The acquisition time of multispectral imagery is used as the baseline timeline, with a 500-millisecond tolerance threshold for time synchronization. The ground environment parameter sequence and keyframes of the video stream are traversed, and the time difference between their timestamps and the baseline timeline is calculated. If the absolute value of the timestamp deviation for a soil data point exceeds 500 milliseconds, the data is considered invalid asynchronous data and is directly removed from the sequence to eliminate timing misalignments caused by network latency. For any timing gaps left after cleaning, a linear interpolation algorithm is used to fill them. Specifically, if at a certain time point... At 10:00:00, the soil moisture was measured to be 25.5%. The result measured at (10:00:02) was 25.9%, but the intermediate time point was missing. For data at (10:00:01), calculate the completion value. After the above calibration and completion operations, a basic data sequence with each data dimension strictly aligned on the time axis is generated, ensuring that the image pixels at the same moment can accurately correspond to the soil environmental values ​​at that time.

[0070] The aforementioned linear interpolation algorithm is a mathematical method that uses the linear relationship between two known data points to estimate the value of a missing intermediate point. It assumes that the data changes linearly in a short period of time and calculates the corresponding value at the target time by using the slope.

[0071] The pre-built spectral inversion logic library in memory is invoked to perform pixel-by-pixel analysis on the spatiotemporally aligned multispectral image. Gray values ​​of the red band (660 nm) and near-infrared band (850 nm) in the image are extracted and converted into surface reflectance values. Based on the multiple linear regression equation stored in the logic library, the reflectance of the characteristic bands is input to invert and calculate the estimated values ​​of soil moisture saturation and nitrogen, phosphorus, and potassium content corresponding to the current pixel. Regarding the construction process of the spectral inversion logic library, historical hyperspectral data (more than 100 bands) from the past three years and synchronously measured soil physicochemical data (such as moisture content measured by oven drying) of the monitoring area are first collected. Partial least squares regression is used, with the full-band reflectance of the spectrum as the independent variable matrix and the measured soil parameters as the dependent variable matrix, for regression analysis. Projection vectors and loading vectors are calculated, and the variation weights of the explanatory power of each band on soil parameters are analyzed. Bands with variation weight values ​​greater than 1.2 are selected as characteristic bands; for example, 1450 nm and 1900 nm are selected as moisture-sensitive bands. Subsequently, a multiple linear regression model was established using the selected characteristic band reflectance as input. An independent validation dataset (200 samples) was input into the model to calculate predicted and measured values, and the root mean square error (RMSE) and coefficient of determination were statistically analyzed. If the coefficient of determination for a given equation is greater than 0.85 and the RMSE is less than 5%, the equation is deemed to have met the accuracy standard, and its coefficient matrix is ​​stored in a logic library. In real-time processing, the inverted soil parameter layer is superimposed on the original image to form a multidimensional data cube containing "spectral-video-physicochemical parameters," and corresponding label information (such as "vegetation cover: high," "water stress: none") is added, completing the construction of the original ecological dataset.

[0072] Table 1. Correlation analysis of characteristic bands and soil parameters.

[0073] As shown in Table 1, historical data analysis revealed a high correlation between different wavelength bands and specific soil parameters, providing a basis for variable selection in the regression equations of the inversion logic library. In practical applications, the system directly calls the combination of wavelength bands with the highest correlation coefficients and substitutes them into the regression equation for calculation. For example, substituting the reflectance of the 1450nm band (0.35) and the reflectance of the 850nm band (0.60) collected at the current time into the moisture inversion equation, the current soil moisture saturation is calculated to be 42%.

[0074] The aforementioned partial least squares regression algorithm is a statistical method that combines the advantages of principal component analysis, canonical correlation analysis, and multiple linear regression analysis. It is particularly suitable for data analysis scenarios where variables exhibit severe multicollinearity and the sample size is smaller than the number of variables.

[0075] Please see Figure 1 and Figure 3S2: Use convolutional neural networks to perform feature extraction and dimensionality reduction on the original ecological dataset, analyze the nonlinear coupling relationship between vegetation growth status and microenvironment factors, construct a multimodal feature matrix, input it into a pre-trained habitat health assessment model, and output the habitat health index.

[0076] S21: Input the original ecological dataset into a multi-layer convolutional neural network, use convolutional kernels to extract texture features from the image data and perform max pooling operation, while normalizing and encoding the environmental parameter sequence to generate a high-dimensional hybrid feature vector.

[0077] S22: By analyzing the dependency weights between elements in a high-dimensional mixed feature vector through an attention mechanism, the interaction strength between vegetation growth status and microenvironmental factors such as light, temperature and humidity is quantified, nonlinear coupling characteristics are analyzed, and a multimodal feature matrix is ​​constructed.

[0078] S23: Load the habitat health assessment model, map the multimodal feature matrix to the preset health score space, calculate the ecological vitality value and stress resistance score of the current monitoring area through the fully connected layer, and output the habitat health index after weighted fusion.

[0079] The calculation process for the habitat health index specifically includes:

[0080] Based on the internal parameters of the habitat health assessment model, the habitat health index is calculated using the following formula:

[0081] ;

[0082] in, This represents the calculated habitat health index. Represents the total number of feature dimensions. Representing the The importance weight coefficients of each feature dimension Represents the first in the multimodal feature matrix Normalized values ​​of biomass characteristics in dimensionality. Stability score representing soil physicochemical properties. and These represent the model hyperparameters for biomass characteristics and environmental stability, respectively. This represents a factor that regulates environmental sensitivity.

[0083] A deep learning computing framework was used to load the original ecological dataset and perform deep feature mining. First, the input data was standardized and preprocessed, mapping pixel values ​​and environmental parameters to a uniform range of 0 to 1. A two-stream convolutional neural network architecture was constructed and run, with one branch specifically processing multispectral image data and the other processing environmental parameter sequence data. Deep within the network, the feature maps from the two branches were concatenated and fused, and fully connected layers were used to analyze the nonlinear coupling relationship between vegetation growth status (e.g., yellowing leaves, sparse canopy) and microenvironmental factors (e.g., localized drought, low nitrogen). The output was connected to a pre-trained habitat health assessment model, which outputs a continuous value between 0 and 100 as a habitat health index, quantitatively representing the overall ecological quality of the current green space.

[0084] The aforementioned dual-stream convolutional neural network architecture refers to a deep learning model structure that includes two independent input channels. The two channels extract data features from different modalities (such as images and numerical sequences), and finally, the features are merged in the fusion layer to achieve multimodal joint analysis.

[0085] The original ecological dataset was split into two parts: an image tensor and an environment vector. A 256×256×5 multispectral image tensor was input into the image branch of a multi-layer convolutional neural network (CNN). The first convolutional layer used 32 3×3 kernels with a stride of 1 to extract texture features while preserving edge information. A Rectified Linear Unified Activation (ReLU) function was used to increase non-linearity. Following this, a 2×2 max pooling operation was performed to reduce the feature map size to 128×128, reducing computation and extracting key features. This "convolution-activation-pooling" process was repeated four times, ultimately generating a 16×16×256 high-dimensional image feature map. Simultaneously, a normalized sequence of environmental parameters (e.g., 0.45, 0.60, 0.78, 0.30 representing temperature, humidity, nitrogen, and light, respectively) was input into a one-dimensional convolutional network or fully connected layer for encoding, mapping it to a feature vector of length 256. The flattened image feature vector is concatenated with the environmental feature vector to generate a high-dimensional hybrid feature vector with a length of 65792, thus completing the initial fusion of multi-source information.

[0086] A self-attention mechanism is introduced to process high-dimensional mixed feature vectors. First, the mixed feature vectors are multiplied by three learnable weight matrices. , , Generate query vectors Key vector Sum value vector .calculate and The dot product of the vectors is divided by a scaling factor (the square root of the vector dimension, e.g., 16), and then normalized using the Softmax function to obtain the attention weight matrix. This weight matrix quantifies the dependencies between elements within the feature vector. For example, it identifies a high-weight correlation (0.85) between the environmental feature of "low soil moisture" and the image feature of "abnormal leaf reflectance," thus mathematically capturing the physical coupling mechanism of "drought leading to vegetation water loss." The attention weight matrix is ​​then used to adjust the value vectors... Weighted summation is performed to output the context vector. A multi-head attention mechanism (with 8 heads) is used to capture the dependencies between different subspaces in parallel, ultimately constructing a multimodal feature matrix that integrates the interaction intensity between vegetation growth status and microenvironmental factors such as light, temperature, and humidity. This matrix has a dimension of 64×256.

[0087] The self-attention mechanism mentioned above refers to a technique that dynamically calculates the correlation weights between elements in an input sequence in a neural network. It allows the model to focus on information from other relevant positions in the sequence while processing information at the current position, thereby capturing long-distance dependencies.

[0088] A pre-trained habitat health assessment model is loaded, and the multimodal feature matrix generated by S22 is input into the fully connected layer of the model. The fully connected layer contains two hidden layers with 128 and 64 nodes respectively, and the final output is key parameters used to calculate the health index. Model internal parameters (Biomass weight) is preset to 0.6. (Environmental stability weight) is preset to 0.4. The environmental sensitivity adjustment factor is preset to 1.2. These hyperparameters were optimized through comparative experiments using historical healthy green space and degraded green space data. Based on the internal parameters of the habitat health assessment model, the habitat health index is calculated:

[0089] ;

[0090] in, This represents the calculated habitat health index; The total number of feature dimensions (3 in this example); Representing the The importance weight coefficients of each feature dimension; Represents the first in the multimodal feature matrix Normalized values ​​of biomass characteristics in dimensionality; Stability score representing soil physicochemical properties (range 0-5). and These represent the model hyperparameters for biomass characteristics and environmental stability, respectively. This represents the environmental sensitivity moderating factor. Assume the normalized values ​​of the three extracted feature dimensions... The corresponding weighting coefficients are 0.8, 0.7, and 0.75, respectively. The values ​​are 0.4, 0.3, and 0.3. First, calculate the biomass portion: The model evaluation output The value is 2.5. The calculation of environmental stability is as follows: .

[0091] Final calculation value: The results indicate that the habitat health index of the currently monitored area is approximately 0.83, indicating a healthy state.

[0092] Please see Figure 1 and Figure 4 S3: Compare the habitat health index with the standard habitat threshold, calculate the growth deviation vector, use the anomaly detection algorithm to identify the mutation features in the multimodal feature matrix, combine the growth deviation vector to determine the comprehensive disaster risk, and generate anomaly early warning signals.

[0093] S31: Obtain the standard habitat threshold corresponding to the current season, calculate the absolute value and rate of change of the difference between each indicator in the habitat health index and the standard habitat threshold, and generate a growth deviation vector.

[0094] S32: The isolated forest anomaly detection algorithm is applied to scan the multimodal feature matrix to identify abrupt changes that deviate from the center of the normal distribution, and the environmental factor type and abrupt change intensity level corresponding to the abrupt changes are marked.

[0095] S33: Risk-weighted analysis of the growth deviation vector based on the mutation intensity level, comprehensive assessment of the current probability of pest and disease outbreaks and the risk of water and fertilizer imbalance in green spaces, determination of the level of comprehensive disaster risk, and generation of abnormal early warning signals.

[0096] The process of determining comprehensive disaster risk specifically includes:

[0097] Using the growth deviation vector and mutation characteristics, the disaster risk value is calculated and the level is determined according to the following logic:

[0098] ;

[0099] in, A quantitative value representing comprehensive disaster risk. Represents the growth deviation vector. This represents the transpose of the growth deviation vector. The inverse matrix representing the covariance matrix of historical disaster data. The weights of the influence factors representing the characteristics of mutations This represents the set of anomalous amplitudes for each feature point in the mutation feature set. Represents the maximum abnormal amplitude in the set of abnormal amplitudes;

[0100] The system according to The final risk level is determined by the preset range in which the value falls.

[0101] The calculated habitat health index of 0.833 is compared with the pre-stored standard habitat threshold in the system database. If the current season is summer, the corresponding standard health threshold is 0.85, and the growth deviation is calculated. Then, the anomaly detection program is initiated, calling the isolated forest algorithm to perform an unsupervised scan of the multimodal feature matrix. The algorithm constructs multiple binary trees, isolates data points by randomly selecting features and split points, and judges anomalies based on path length. If the path length of a feature point is found to be significantly shorter than the average path length, it is determined to be a mutation feature (such as a sudden lesion feature). Combining the growth deviation vector with the identified mutation features, the current disaster risk level of the green space is comprehensively determined, for example, determined to be "moderate drought risk," and an anomaly warning signal containing the risk type and level is generated.

[0102] The seasonal database was queried to obtain the set of standard habitat thresholds for green spaces corresponding to the current month (e.g., June), including a standard health index threshold of 0.85, a standard leaf area index of 2.5, and a standard soil moisture range [30%, 45%]. The difference between the calculated real-time habitat health index of 0.833 and the standard threshold of 0.85 was calculated, yielding an absolute deviation of 0.017. Simultaneously, the rate of change for each individual indicator was calculated; for example, the rate of change of the current soil moisture of 25.7% relative to the standard lower limit of 30% was 14.3%. The health index deviation, the rate of change of key environmental factors, and other components were combined to generate a growth deviation vector. (These correspond to the total index deviation, water deviation, and nitrogen deviation, respectively). This vector intuitively reflects the degree and direction in which the current ecological situation deviates from the ideal state.

[0103] An isolated forest anomaly detection algorithm is applied to perform a full scan of the multimodal feature matrix. The algorithm is initialized with 100 trees, each with a subsampling size of 256. For each input data point, the algorithm randomly selects a feature dimension and then randomly selects a split point between the maximum and minimum values ​​of that dimension, recursively dividing the data space into subspaces until the data point is isolated. The average path length of the data point across the 100 trees is calculated. An anomaly score threshold of 0.6 is set. If the average path length of a feature vector (corresponding to a yellowed area in the image) is short, and the calculated anomaly score reaches 0.75 (greater than 0.6), it is marked as a mutation feature. The system backtracks the original environmental factor type corresponding to this feature, identifying that the mutation is mainly caused by a "blue shift at the red edge of the spectrum," and determines its mutation intensity level as "Level 2" (medium intensity) based on the amplitude.

[0104] The aforementioned isolated forest anomaly detection algorithm refers to an unsupervised anomaly detection method based on ensemble learning. It utilizes the characteristic that anomalous data is easily isolated (i.e., the path is relatively short in the decision tree) and quickly identifies outliers in the dataset that deviate from the normal distribution by constructing multiple random trees.

[0105] Based on the extracted growth deviation vector Based on the mutation characteristic information and the growth deviation vector, the disaster risk value is calculated and the level is determined according to the following logic:

[0106] ;

[0107] in, A quantitative value representing comprehensive disaster risk; Represents the growth deviation vector; Represents the transpose of the growth deviation vector; The inverse matrix representing the covariance matrix of historical disaster data; Weights of influence factors representing mutation characteristics; The set of anomalous amplitudes representing each feature point in the mutation feature set; This represents the maximum abnormal amplitude in the set of abnormal amplitudes. The inverse matrix of the covariance matrix of historical disaster data is defined. A 3×3 matrix was used to quantify the correlation and inverse weights among fluctuations in three variables—health index, water content, and nitrogen—in historical data. Weights were also assigned to factors influencing mutation characteristics. It is 1.5. Take the value obtained from S31. After matrix operations, the quadratic form terms The result was 0.18 (representing the underlying risk due to cumulative deviation). This is the largest anomalous amplitude among the detected mutation feature set. The value is 0.25. Substitute this value into the formula to calculate. The system's preset risk level ranges are: 0 to 0.3 for low risk, 0.3 to 0.6 for medium risk, and 0.6 to 1.0 for high risk. Since the calculated result of 0.555 falls within the medium risk range, the system determines the current comprehensive disaster risk level to be "medium risk," with the main risk type tending towards early vegetation degradation caused by water and fertilizer imbalance. It then generates an abnormal warning signal containing "risk value 0.555, level: medium, type: water and fertilizer imbalance."

[0108] Please see Figure 1 and Figure 5 S4: Based on the analysis of abnormal early warning signals, maintenance needs are identified, and a decision tree algorithm is used to match precise management strategies. The parameters of irrigation duration, pesticide concentration, and inspection path are integrated to construct equipment control commands, which are then sent to automated management equipment for execution.

[0109] S41: Analyze the risk type code and disaster severity in the abnormal warning signal, retrieve the preset maintenance knowledge graph, determine the type of resources and the urgency of the operation required to eliminate the current risk, and generate maintenance requirements.

[0110] S42: Input maintenance requirements into the logical decision tree model, traverse the historical treatment cases and expert rules stored in the leaf nodes, match the optimal precision maintenance strategy, and clarify the corresponding operation mode and resource allocation.

[0111] S43: Calculate the specific irrigation duration, pesticide concentration, and unmanned equipment inspection path parameters based on the precision management strategy, encapsulate them into binary data packets that conform to the industrial control protocol, construct equipment control commands, and issue them.

[0112] The training process of a logical decision tree model specifically includes:

[0113] Collect disaster event records, environmental parameters at the time, response measures taken, and post-response recovery data from the green space management historical database to construct a training sample set;

[0114] Using disaster type and environmental characteristics as input nodes and the effectiveness of response measures as the splitting criterion, the information gain algorithm is used to construct the hierarchical structure of the decision tree, and overfitting branches are pruned and optimized.

[0115] The optimized tree structure is transformed into an executable rule engine, and the weight parameters of the leaf nodes are dynamically updated based on the new management feedback data to complete the deployment of the logical decision tree model.

[0116] Upon receiving the abnormal warning signal generated by S3, the system enters the decision-making and response phase. First, the signal content is analyzed, identifying the core problem as "insufficient soil moisture accompanied by mild nitrogen deficiency." Based on this analysis, a decision tree algorithm is used to search and match within the automated management strategy library. The decision tree derives the optimal management strategy combination based on the current wind speed (2.5 m / s), light intensity (strong), and risk level (medium). Specific strategies include: activating the intelligent sprinkler system and setting the irrigation duration to 25 minutes; simultaneously, instructing a drone equipped with a foliar fertilizer spraying device, setting the application concentration to 1:500, and planning a "U"-shaped patrol and operation path covering the abnormal area. These parameters are encapsulated into control commands in Modbus TCP or MQTT protocol format and sent via an IoT gateway to the on-site solenoid valve controller and drone base station, driving the hardware to perform physical operations.

[0117] The system analyzes the risk type codes in abnormal early warning signals. For example, code "E-02" corresponds to "soil drought," and code "N-01" corresponds to "slight nitrogen deficiency." Based on a risk value of 0.555, the severity of the disaster is determined to be "moderately urgent," requiring intervention within 24 hours. The system retrieves a pre-set maintenance knowledge graph, which stores knowledge in triplets, such as <soil drought, mitigation method, sprinkler irrigation> and <nitrogen deficiency, mitigation method, urea application>. Combining the urgency level, a specific maintenance requirement list is generated: approximately 15 tons of water resources are needed (based on area estimation), 5 kg of quick-acting nitrogen fertilizer is needed, and the operation must be carried out during windless or lightly windy periods.

[0118] The maintenance needs and current real-time environmental characteristics (temperature 28℃, wind speed 2.5m / s, no rain expected in the next 3 hours) are input into the logical decision tree model. The decision tree traverses from the root node, which determines the "risk type," and the branch points to "integrated water and fertilizer management." The next node determines the "wind speed," and since the wind speed of 2.5m / s is less than the threshold of 3.0m / s, it enters the "suitable spraying" branch. The next node determines the "light intensity," and given the current strong light, to avoid the lensing effect of water droplets scorching the leaves, the logical branch points to either "delay execution until evening" or "adopt root drip irrigation + low-altitude fogging mode." The training process of the logical decision tree model involves collecting a historical database of green space management data from the past 5 years, containing 5000 disaster event records. "Disaster type" (e.g., pests, drought) and "environmental characteristics" (e.g., high temperature, heavy rainfall) are used as input features, and "effectiveness of treatment measures" (e.g., whether the normalized vegetation index recovery rate exceeds 5% within one week after the measures are implemented) is used as a label. The decision tree is constructed using an information gain algorithm. The information entropy of the dataset is calculated, all features are traversed, and the information gain of each feature after splitting the dataset is calculated. The feature with the largest gain is selected as the splitting node. For example, if the calculation shows that the "wind speed" feature has the largest information gain for "spraying operation effect", it is selected as the upper-level node. After generating the tree structure, branches with fewer than 10 samples are pruned to prevent overfitting. Finally, the optimal precision management strategy matched by the model is: using a combined operation mode of drone low-altitude mist fertilization and ground drip irrigation, with the resource allocation being an integrated water and fertilizer solution.

[0119] The aforementioned information gain algorithm refers to a method that measures the contribution of a feature to the classification result by calculating the difference in information entropy before and after splitting the dataset during the decision tree construction process. The larger the gain, the stronger the classification ability of the feature.

[0120] Based on the established precise management strategy, specific execution parameters are calculated numerically. For irrigation duration, the formula is derived using a physical model: .

[0121] Where T represents irrigation duration (hours); A represents the area of ​​operation; Represents the target moisture content; D represents the current moisture content; Q represents the soil depth; and Q represents the equipment flow rate. This represents the utilization factor. The obtained parameter A (value is 1000 square meters)... (Value is 0.40) (Value is 0.257), D (value is 0.2 meters), Q (value is 5 cubic meters / hour) and Substituting the value (0.9) into the above formula, the final result T is calculated to be 6.35 hours. The system automatically divides it into 3 time periods for execution to prevent runoff. The pesticide concentration is set to a standard 1.5% urea solution. For the unmanned equipment patrol path, parallel search lines with a spacing of 5 meters are planned based on the area boundary coordinates, generating a path file containing a series of GPS waypoints. Finally, parameters such as "open valves V1-V4, duration 127 minutes" and "UAV take off, load path file Path_01, spraying flow rate 2L / min" are encapsulated into a data packet conforming to industrial control protocols (such as hexadecimal Hex format binary code stream), constructing the final equipment control command and sending it to the execution terminal.

[0122] Table 2 Examples of Automated Maintenance Strategy Matching

[0123] As shown in Table 2, the system matches differentiated strategies based on different input conditions. In this embodiment, the strategy parameters ultimately selected through decision tree logic are transformed into specific equipment instructions, ensuring the accuracy and effectiveness of maintenance operations. Experimental data shows that after applying this automated maintenance strategy, the water and fertilizer utilization rate of green space maintenance increased by 25%, and the cost of manual inspection decreased by 60%.

[0124] An AI-based dynamic monitoring system for ecological green spaces, used to execute the aforementioned AI-based dynamic monitoring method for ecological green spaces, the system comprising:

[0125] The ecological data perception module is used to collect multispectral remote sensing images, ground environmental parameters and vegetation morphology video streams through sensing devices. It uses edge computing nodes to synchronize and align data based on timestamps and geographic coordinates, and establishes correlation logic based on spectral reflectance and soil physicochemical properties to construct the original ecological dataset.

[0126] The habitat health assessment module is used to perform feature extraction and dimensionality reduction on the original ecological dataset using a convolutional neural network, analyze the nonlinear coupling relationship between vegetation growth status and microenvironment factors, construct a multimodal feature matrix, input it into a pre-trained habitat health assessment model, and output a habitat health index.

[0127] The disaster risk early warning module is used to compare the habitat health index with the standard habitat threshold, calculate the growth deviation vector, use the anomaly detection algorithm to identify the mutation features in the multimodal feature matrix, combine the growth deviation vector to determine the comprehensive disaster risk, and generate anomaly early warning signals.

[0128] The automated maintenance execution module is used to analyze maintenance needs based on abnormal early warning signals, match precise maintenance strategies using decision tree algorithms, integrate irrigation duration, pesticide concentration, and inspection path parameters, construct equipment control commands, and send them to the automated maintenance equipment for execution.

[0129] The above embodiments illustrate preferred embodiments of the present invention. Any equivalent adjustments to the technical solution based on software engineering methods are within the scope of protection, including but not limited to: implementing algorithm logic using different programming languages, refactoring functional modules into services, adjusting data interaction protocols, and optimizing resource scheduling strategies. Any implementation scheme derived from reasonable modifications to the data processing flow, service call chain, or system architecture layer without departing from the core technology of the present invention should be considered within the protection scope defined by the technical solution of the present invention.

Claims

1. An AI-based dynamic monitoring method for ecological green spaces, characterized in that, Includes the following steps: S1: Collect multispectral remote sensing images, ground environmental parameters and vegetation morphology video streams through sensing devices, use edge computing nodes to synchronize and align data based on timestamps and geographic coordinates, establish correlation logic based on spectral reflectance and soil physicochemical properties, and construct the original ecological dataset. S2: Use a convolutional neural network to perform feature extraction and dimensionality reduction on the original ecological dataset, analyze the nonlinear coupling relationship between vegetation growth status and microenvironment factors, construct a multimodal feature matrix, input it into a pre-trained habitat health assessment model, and output the habitat health index. S3: Compare the habitat health index with the standard habitat threshold, calculate the growth deviation vector, use an anomaly detection algorithm to identify mutation features in the multimodal feature matrix, combine the growth deviation vector to determine the comprehensive disaster risk, and generate an anomaly warning signal; S4: Based on the abnormal early warning signal analysis of maintenance needs, use the decision tree algorithm to match precise management strategies, integrate irrigation duration, pesticide concentration and inspection path parameters, construct equipment control commands, and send them to automated management equipment for execution.

2. The AI-based dynamic monitoring method for ecological green spaces according to claim 1, characterized in that, The specific steps of S1 are as follows: S11: Activate the heterogeneous sensing devices deployed in the monitoring area, and synchronously acquire the multispectral remote sensing images, the ground environmental parameters, and the vegetation morphology video stream according to the preset high-frequency sampling cycle, and extract the acquisition timestamp and spatial latitude and longitude coordinate information from various data metadata. S12: Use the edge computing node to perform spatiotemporal calibration on various heterogeneous data, remove invalid data frames with time deviations exceeding a preset threshold, and fill in the missing time-series data points through a linear interpolation algorithm to generate a spatiotemporally aligned basic data sequence. S13: Call the pre-stored spectral inversion logic library, parse the band reflectance values ​​in the multispectral remote sensing image, establish the mapping relationship between them and the soil nitrogen, phosphorus and potassium content and water saturation in the ground environmental parameters, integrate the aligned image data, environmental values ​​and mapping relationship labels, and construct the original ecological dataset.

3. The AI-based dynamic monitoring method for ecological green spaces according to claim 1, characterized in that, The specific steps of S2 are as follows: S21: Input the original ecological dataset into a multi-layer convolutional neural network, use convolutional kernels to extract texture features from the image data and perform max pooling operation, while normalizing and encoding the environmental parameter sequence to generate a high-dimensional hybrid feature vector. S22: Analyze the dependency weights between elements in the high-dimensional hybrid feature vector through the attention mechanism, quantify the interaction strength between the vegetation growth status and the microenvironment factors such as light, temperature and humidity, analyze the nonlinear coupling characteristics, and construct the multimodal feature matrix; S23: Load the habitat health assessment model, map the multimodal feature matrix to the preset health score space, calculate the ecological vitality value and stress resistance score of the current monitoring area through a fully connected layer, and output the habitat health index after weighted fusion.

4. The AI-based dynamic monitoring method for ecological green spaces according to claim 1, characterized in that, The specific steps of S3 are as follows: S31: Obtain the standard habitat threshold corresponding to the current season, calculate the absolute value and rate of change of the difference between each indicator in the habitat health index and the standard habitat threshold, and generate the growth deviation vector; S32: Apply the isolated forest anomaly detection algorithm to traverse and scan the multimodal feature matrix, identify abrupt changes that deviate from the center of the normal distribution, and mark the environmental factor type and abrupt change intensity level corresponding to the abrupt changes. S33: Based on the mutation intensity level, the growth deviation vector is risk-weighted to comprehensively assess the probability of pest and disease outbreaks and the risk of water and fertilizer imbalance in the current green space, determine the level of the comprehensive disaster risk, and generate the abnormal early warning signal.

5. The AI-based dynamic monitoring method for ecological green spaces according to claim 1, characterized in that, The specific steps of S4 are as follows: S41: Analyze the risk type code and disaster severity in the abnormal warning signal, retrieve the preset maintenance knowledge graph, determine the resource type and work urgency required to eliminate the current risk, and generate maintenance requirements. S42: Input the maintenance requirements into the logical decision tree model, traverse the historical treatment cases and expert rules stored in the leaf nodes, match the optimal precision management strategy, and clarify the corresponding operation mode and resource allocation. S43: Calculate the specific irrigation duration, pesticide concentration, and patrol path parameters of the unmanned equipment according to the precision management strategy, encapsulate them into a binary data packet conforming to the industrial control protocol, construct the equipment control command, and issue it.

6. The AI-based dynamic monitoring method for ecological green spaces according to claim 3, characterized in that, The calculation process of the habitat health index specifically includes: Based on the internal parameters of the habitat health assessment model, the habitat health index is calculated according to the following formula: ； in, This represents the calculated habitat health index. Represents the total number of feature dimensions. Representing the The importance weight coefficients of each feature dimension Represents the first in the multimodal feature matrix Normalized values ​​of biomass characteristics in dimensionality. Stability score representing soil physicochemical properties. and These represent the model hyperparameters for biomass characteristics and environmental stability, respectively. This represents a factor that regulates environmental sensitivity.

7. The AI-based dynamic monitoring method for ecological green spaces according to claim 4, characterized in that, The process for determining the comprehensive disaster risk specifically includes: Using the aforementioned growth deviation vector and mutation characteristics, the disaster risk value is calculated and the level is determined according to the following logic: ； in, The quantitative value representing the comprehensive disaster risk, This represents the growth deviation vector. This represents the transpose of the growth deviation vector. The inverse matrix representing the covariance matrix of historical disaster data. The weights of the influence factors representing the characteristics of mutations This represents the set of anomalous amplitudes for each feature point in the mutation feature set. This represents the maximum abnormal amplitude in the set of abnormal amplitudes; The system according to The final risk level is determined by the preset range in which the value falls.

8. The AI-based dynamic monitoring method for ecological green spaces according to claim 2, characterized in that, The construction process of the spectral inversion logic library specifically includes: Hyperspectral images of the monitoring area and synchronously measured soil physicochemical data were acquired within a historical time period. The correlation coefficient between spectral bands and soil parameters was analyzed using partial least squares regression algorithm. Feature bands with correlation coefficients greater than the preset confidence level were selected, and a multiple linear regression equation was established between the reflectance of the feature bands and soil moisture content and organic matter content. The prediction accuracy of the multiple linear regression equation is verified using an independent validation dataset. Equations with a standard coefficient of determination are solidified into inversion rules in the spectral inversion logic library for subsequent real-time data analysis.

9. The AI-based dynamic monitoring method for ecological green spaces according to claim 5, characterized in that, The training process of the logical decision tree model specifically includes: Collect disaster event records, environmental parameters at the time, response measures taken, and post-response recovery data from the green space management historical database to construct a training sample set; Using disaster type and environmental characteristics as input nodes and the effectiveness of response measures as the splitting criterion, the information gain algorithm is used to construct the hierarchical structure of the decision tree, and overfitting branches are pruned and optimized. The optimized tree structure is transformed into an executable rule engine, and the weight parameters of the leaf nodes are dynamically updated based on the new management feedback data, thus completing the deployment of the logical decision tree model.

10. An AI-based dynamic monitoring system for ecological green spaces, characterized in that, The system is used to implement the AI-based dynamic monitoring method for ecological green spaces as described in any one of claims 1-9, and the system comprises: The ecological data perception module is used to collect multispectral remote sensing images, ground environmental parameters and vegetation morphology video streams through sensing devices. It uses edge computing nodes to synchronize and align data based on timestamps and geographic coordinates, and establishes correlation logic based on spectral reflectance and soil physicochemical properties to construct the original ecological dataset. The habitat health assessment module is used to perform feature extraction and dimensionality reduction on the original ecological dataset using a convolutional neural network, analyze the nonlinear coupling relationship between vegetation growth status and microenvironment factors, construct a multimodal feature matrix, input it into a pre-trained habitat health assessment model, and output a habitat health index. The disaster risk early warning module is used to compare the habitat health index with the standard habitat threshold, calculate the growth deviation vector, use anomaly detection algorithm to identify mutation features in the multimodal feature matrix, combine the growth deviation vector to determine the comprehensive disaster risk, and generate anomaly early warning signals. The automated maintenance execution module is used to analyze maintenance needs based on the abnormal early warning signal, match precise maintenance strategies using a decision tree algorithm, integrate irrigation duration, pesticide concentration and inspection path parameters, construct equipment control commands, and send them to the automated maintenance equipment for execution.

Images