Forest ecosystem long-term monitoring system and method based on machine learning

Abstract

The invention relates to a forest ecosystem long-term monitoring system and method based on machine learning. The method comprises the following steps: acquiring multi-source detection data of the forest detection area; the multi-source detection data comprises visible image data, infrared image data, three-dimensional scanning data and meteorological parameter data; establishing a distributed model training mechanism through a federated learning framework, and performing cleaning fusion on the multi-source detection data to obtain an ecological feature vector; performing multi-branch analysis on the ecological feature vector by adopting a multi-task deep neural network model to obtain a biodiversity recognition result, a vegetation index inversion result and an environment anomaly detection result; and according to the ecological feature vector, the biodiversity recognition result, the vegetation index inversion result and the environmental anomaly detection result, generating an ecological evolution digital twinborn body. By adopting the method, data of a plurality of important aspects of forest resource supervision can be integrated, analyzed and visually displayed.

Description

Technical Field

[0001] The present invention belongs to the field of forest resource monitoring, and in particular relates to a long-term forest ecosystem monitoring system and method based on machine learning. Background Art

[0002] With the development of forest resource monitoring technology, a technology based on multi-source sensors to detect forest ecology has emerged. This technology analyzes and detects forest fire risks and biodiversity by collecting data from sensors installed in the forest and images taken by drones.

[0003] The existing forest ecological detection system integrates temperature sensors preset in the forest, combines them with regular drone aerial images, extracts sensor and aerial image features through convolutional neural networks, and fuses them to obtain fused data from multiple sources of sensors. Based on this fused data, forest fire risks and biodiversity are analyzed.

[0004] However, the current forest ecological monitoring system lacks more analytical perspectives and is insufficiently developed for the integration of multi-source sensors. It only analyzes and predicts one aspect of forest resource management through two sensor data points, and lacks data integration, analysis, and visualization of multiple important aspects of forest resource management. Summary of the Invention

[0005] Based on this, it is necessary to address the above technical issues and provide a long-term forest ecosystem monitoring system and method based on machine learning that can integrate data on multiple important aspects of forest resource supervision, and perform analysis and visualization.

[0006] In a first aspect, the present application provides a long-term forest ecosystem monitoring system based on machine learning, comprising:

[0007] Sensors for acquiring sensing data of forest detection areas;

[0008] The control terminal is connected to the sensor to analyze and process the sensor data, obtain the monitoring results, and generate the ecological evolution digital twin based on the monitoring results;

[0009] A server, connected to the control terminal, for providing meteorological information of the forest detection area;

[0010] Visualization hardware, connected to the control terminal, is used to display monitoring results based on the ecological evolution digital twin;

[0011] Among them, analyzing and processing sensor data to obtain monitoring results includes:

[0012] Acquire multi-source detection data of the forest detection area; the multi-source detection data includes visible image data, infrared image data, 3D scanning data and meteorological parameter data;

[0013] A distributed model training mechanism is established through a federated learning framework to clean and fuse multi-source detection data to obtain ecological feature vectors;

[0014] A multi-task deep neural network model is used to analyze ecological feature vectors in multiple branches to obtain biodiversity identification results, vegetation index inversion results, and environmental anomaly detection results;

[0015] Based on the ecological characteristic vectors, biodiversity identification results, vegetation index inversion results and environmental anomaly detection results, an ecological evolution digital twin is generated; the ecological evolution digital twin is used to dynamically display the trend map of the detection area and generate ecological early warning signals.

[0016] Furthermore, the system also includes:

[0017] The mobile unit is connected to the sensor and is used to drive the sensor to move and collect sensor data;

[0018] The energy supply unit is connected to sensors, control terminals, servers, visualization hardware and mobile units to provide energy.

[0019] In a second aspect, the present application further provides a method for long-term monitoring of forest ecosystems based on machine learning, which is applied to a control terminal of a long-term monitoring system for forest ecosystems based on machine learning described in the first aspect of the present application, comprising:

[0020] Acquire multi-source detection data of the forest detection area; the multi-source detection data includes visible image data, infrared image data, 3D scanning data and meteorological parameter data;

[0021] A distributed model training mechanism is established through a federated learning framework to clean and fuse multi-source detection data to obtain ecological feature vectors;

[0022] A multi-task deep neural network model is used to analyze ecological feature vectors in multiple branches to obtain biodiversity identification results, vegetation index inversion results, and environmental anomaly detection results;

[0023] Based on the ecological characteristic vectors, biodiversity identification results, vegetation index inversion results and environmental anomaly detection results, an ecological evolution digital twin is generated; the ecological evolution digital twin is used to dynamically display the trend map of the detection area and generate ecological early warning signals.

[0024] Furthermore, a multi-task deep neural network model was used to analyze ecological feature vectors in multiple branches to obtain biodiversity identification results, vegetation index inversion results, and environmental anomaly detection results, including:

[0025] Based on a multi-task neural network model, the ecological feature vectors were analyzed to obtain preliminary biodiversity identification results, preliminary vegetation index inversion results, preliminary environmental anomaly detection results, and fusion shared features;

[0026] The Monte Carlo Dropout confidence algorithm is used to adjust the preliminary biodiversity identification results, preliminary vegetation index inversion results and preliminary environmental anomaly detection results according to the fusion shared features to obtain the biodiversity identification results, vegetation index inversion results and environmental anomaly detection results.

[0027] Furthermore, based on the multi-task neural network model, the ecological feature vectors were analyzed to obtain preliminary biodiversity identification results, preliminary vegetation index inversion results, preliminary environmental anomaly detection results and fusion shared features, including:

[0028] The following formula is used to extract cross-task features of ecological feature vectors through a fully connected layer and a dynamic feature routing mechanism to obtain fused shared features:

[0029]

[0030] Among them, h s is the fusion shared feature, F g is the ecological characteristic vector, c ij is the dynamic routing weight, W ij is the capsule weight from input capsule i to output capsule j, b ij is the number of routing pairs from input capsule i to output capsule j;

[0031] Among them, the iterative update formula of dynamic routing weight is:

[0032]

[0033] Among them, c ij is the dynamic routing weight, b ij is the number of routing pairs from input capsule i to output capsule j, b ik is the number of routing pairs from input capsule i to output capsule k;

[0034] Multi-category classification is performed on the fused shared features to obtain preliminary biodiversity identification results, preliminary vegetation index inversion results, and preliminary environmental anomaly detection results;

[0035] Among them, the following formula is used to balance the multi-class classification weights using uncertainty weighting:

[0036]

[0037] in, is the total loss for multi-class classification, is the loss function of the i-th task, is the noise variance of the ith task.

[0038] Furthermore, the Monte Carlo Dropout confidence algorithm is used to adjust the preliminary biodiversity identification results, preliminary vegetation index inversion results, and preliminary environmental anomaly detection results based on the fusion shared features. The biodiversity identification results, vegetation index inversion results, and environmental anomaly detection results are obtained, including:

[0039] Based on the Dropout sampling inference algorithm, the fused shared features are forward inferred to obtain a confident prediction result;

[0040] Use the following formula to calculate the prediction confidence variance corresponding to the confidence prediction result:

[0041]

[0042] in, is the prediction confidence variance, T is the number of samples, y t is the confidence prediction result of the t-th sampling;

[0043] The prediction confidence variance is combined with the preliminary biodiversity identification results, preliminary vegetation index inversion results, and preliminary environmental anomaly detection results to obtain the corresponding biodiversity identification results, vegetation index inversion results, and environmental anomaly detection results.

[0044] Furthermore, a distributed model training mechanism is established through the federated learning framework to clean and integrate multi-source detection data to obtain ecological feature vectors, including:

[0045] Using isolation forest technology, data cleaning is performed on visible image data, infrared image data, 3D scanning data, and meteorological parameter data to obtain corresponding ideal visible image data, ideal infrared image data, ideal 3D scanning data, and ideal meteorological parameter data;

[0046] Performing linear projection and feature extraction on ideal visible image data, ideal infrared image data, ideal three-dimensional scanning data, and ideal meteorological parameter data to obtain corresponding visible image features, infrared image features, three-dimensional scanning features, and meteorological parameter features;

[0047] The fusion weight is calculated based on the attention mechanism using the following formula:

[0048]

[0049] Among them, h i is the fusion weight of the i-th head, softmax is the classification function, Q i is the query matrix of the i-th head, is the key matrix of the i-th head, d is the feature dimension, H is the total number of attention heads, P is the position encoding matrix, V i is the value matrix of the i-th head;

[0050] A gating mechanism is introduced to dynamically adjust the fusion weights, and the visible image features, infrared image features, 3D scanning features and meteorological parameter features are dynamically fused to obtain the primary fusion feature vector.

[0051] The principal component analysis technique is used to compress the primary fusion feature vector to obtain the fusion feature vector;

[0052] Based on the federated learning technology, the fused feature vectors are globally integrated to obtain the ecological feature vectors.

[0053] Furthermore, based on federated learning technology, the fusion feature vectors are globally integrated to obtain the ecological feature vectors, including:

[0054] According to the fusion feature vector, based on the federated learning technology, the ecological feature vector is calculated using the following formula:

[0055]

[0056] Among them, F g is the ecological feature vector, N is the total number of edge points participating in federated learning, is the fusion feature vector of the i-th edge point, is the alignment matrix of the i-th edge point.

[0057] Furthermore, based on the ecological feature vectors, biodiversity identification results, vegetation index inversion results, and environmental anomaly detection results, an ecological evolution digital twin is generated, including:

[0058] According to the ecological feature vector, biodiversity identification results, vegetation index inversion results and environmental anomaly detection results, the distribution feature contribution is calculated based on cooperative game theory to obtain the corresponding biodiversity contribution, vegetation index contribution and environmental anomaly contribution;

[0059] Project the biodiversity contribution, vegetation index contribution, and environmental anomaly contribution into the preset detection area to obtain a feature importance heat map;

[0060] Based on the preset three-dimensional scanning model of the detection area, a detection area network model is constructed;

[0061] The biodiversity identification results, vegetation index inversion results, and environmental anomaly detection results are vertex-associated with the detection region network model to obtain a fusion region network model.

[0062] The feature importance heat map is mapped to the texture and transparency of the fusion regional network model to obtain the ecological evolution digital twin.

[0063] Furthermore, the method further comprises:

[0064] The ecological feature vectors, biodiversity identification results, vegetation index inversion results and environmental anomaly detection results are divided into blocks according to the spatiotemporal windows to obtain standardized data blocks;

[0065] Use the SHA-256 algorithm to generate a unique identifier corresponding to the standardized data block and obtain the data hash;

[0066] Based on data hash and standardized data blocks, blockchain technology is used for storage to obtain analysis results and store information.

[0067] The present application provides a long-term forest ecosystem monitoring system and method based on machine learning, which obtains multi-source detection data of the forest detection area; the multi-source detection data includes visible image data, infrared image data, three-dimensional scanning data and meteorological parameter data; a distributed model training mechanism is established through a federated learning framework, and the multi-source detection data is cleaned and integrated to obtain ecological feature vectors; a multi-task deep neural network model is used to perform multi-branch analysis of the ecological feature vectors to obtain biodiversity identification results, vegetation index inversion results and environmental anomaly detection results; an ecological evolution digital twin is generated based on the ecological feature vectors, biodiversity identification results, vegetation index inversion results and environmental anomaly detection results; the ecological evolution digital twin is used to dynamically display trend maps of the detection area and generate ecological early warning signals, realize the integration of visible image data, infrared image data, three-dimensional scanning data and meteorological parameter data, conduct comprehensive analysis and visual display of multiple forest supervision aspects, and improve the efficiency of forest supervision. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0069] Figure 1 A schematic diagram of the structure of a long-term forest ecosystem monitoring system based on machine learning provided in one embodiment of the present application;

[0070] Figure 2 A schematic structural diagram of a long-term forest ecosystem monitoring system based on machine learning provided in another embodiment of the present application;

[0071] Figure 3 A flowchart of a long-term monitoring method of forest ecosystems based on machine learning is provided as an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0072] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0073] In one embodiment, Figure 1 As shown, a long-term forest ecosystem monitoring system 100 based on machine learning is provided, comprising:

[0074] Sensor 101 is used to obtain sensor data of the forest detection area.

[0075] Specifically, the sensor 101 is used to obtain sensor data of the forest area to be inspected. Schematically, the sensor data includes but is not limited to optical imaging data, infrared imaging data, and LiDAR three-dimensional scanning data captured in the inspection area.

[0076] The control terminal 102 is connected to the sensor, and is used to analyze and process the sensor data, obtain monitoring results, and generate an ecological evolution digital twin based on the monitoring results.

[0077] Specifically, control terminal 102 is connected to sensor 101 and is used to organize and analyze standard photos, infrared images, and 3D regional scan data acquired by sensor 101 to obtain monitoring results for the forest area. Based on these monitoring results, a visual digital twin of the ecological evolution is generated. Illustratively, the information displayed by the digital twin includes, but is not limited to, 3D ecological map rendering, dynamic trend graph generation, and ecological early warning signals.

[0078] The server 103 is connected to the control terminal and is used to provide meteorological information of the forest detection area.

[0079] Specifically, the server 103 is connected to the control terminal 102 to obtain meteorological information of the forest area and transmit it to the control terminal 102 for comprehensive analysis.

[0080] The visualization hardware 104 is connected to the control terminal and is used to display the monitoring results based on the ecological evolution digital twin.

[0081] Specifically, the visualization hardware 104 is connected to the control terminal 102 and is used to visualize the ecological evolution digital twin generated by the control terminal 102 and display the monitoring results.

[0082] Schematically, the sensor 101, the control terminal 102, the server 103 and the visualization hardware 104 may be connected via wires or wired connections or wireless connections. Figure 1 The example of connection through wires is shown with only one thick line, but this does not mean that there is only one wire or one type of wire.

[0083] Among them, analyzing and processing sensor data to obtain monitoring results includes:

[0084] Acquire multi-source detection data of the forest detection area; the multi-source detection data includes visible image data, infrared image data, 3D scanning data and meteorological parameter data;

[0085] A distributed model training mechanism is established through a federated learning framework to clean and fuse multi-source detection data to obtain ecological feature vectors;

[0086] A multi-task deep neural network model is used to analyze ecological feature vectors in multiple branches to obtain biodiversity identification results, vegetation index inversion results, and environmental anomaly detection results;

[0087] Based on the ecological characteristic vectors, biodiversity identification results, vegetation index inversion results and environmental anomaly detection results, an ecological evolution digital twin is generated; the ecological evolution digital twin is used to dynamically display the trend map of the detection area and generate ecological early warning signals.

[0088] Specifically, control terminal 102 preprocesses and cleans the visible image data, infrared image data, and three-dimensional scanning data obtained from sensor 101, and meteorological parameter data obtained from server 103. It then performs multi-task branch prediction using a multi-task deep neural network model to obtain biodiversity identification results, vegetation index inversion results, and environmental anomaly detection results. Based on these biodiversity identification results, vegetation index inversion results, and environmental anomaly detection results, it generates an ecological evolution digital twin, visually displays a trend map of the detection area, and generates an ecological warning signal. Schematically, this visual display includes, but is not limited to, a trend map of the detection area and the generation of an ecological warning signal.

[0089] This embodiment sets up sensors to collect optical imaging data, infrared imaging data, and LiDAR three-dimensional scanning data of the detected forest area, and transmits them to the control terminal. It combines the meteorological parameter information provided by the server to make predictions in three aspects: biodiversity, vegetation index, and environmental anomaly detection, and generates an ecological evolution digital twin to visualize the monitoring and prediction data. It realizes comprehensive analysis and visualization of multiple aspects of forest supervision, thereby improving the efficiency of forest supervision.

[0090] In one embodiment, Figure 2 As shown, the system also includes:

[0091] The mobile unit 105 is connected to the sensor and is used to drive the sensor to move and collect sensor data.

[0092] Specifically, the system also includes a mobile unit 105, connected to the sensor 101, for moving the sensor to collect sensor data. Illustratively, mobile unit 105 can be a drone, an unmanned vehicle, or a robot dog. Optionally, mobile unit 105 is also connected to the control terminal 102, for receiving collection instructions from the control terminal 102 and taking action.

[0093] Preferably, the mobile unit 105 also includes a control chip equipped with a terrain analysis algorithm for receiving a collection instruction from the control terminal 102 to determine the location information of the detection area, and autonomously planning the optimal travel path and the optimal monitoring point layout of the detection area.

[0094] The energy supply unit 106 is connected to the sensor, the control terminal, the server, the visualization hardware and the mobile unit to provide energy.

[0095] Specifically, the power supply 106 is connected to the sensors, control terminals, servers, visualization hardware, and mobile units to provide energy to other units or modules in the system to enable them to operate normally. Illustratively, the power supply 106 also integrates photovoltaic power generation and energy collection modules to support long-term field operation.

[0096] Schematically, the sensor 101, the control terminal 102, the server 103, the visualization hardware 104, the mobile unit 105 and the energy supply unit 106 may be connected by wires or wired connections or wireless connections. Figure 2 The example of connection through wires is shown with only one thick line, but this does not mean that there is only one wire or one type of wire.

[0097] This embodiment sets up a mobile unit to ensure that the sensor moves to the forest area that needs to be detected and obtains sensor data, and sets up an energy supply unit to provide energy for the system to ensure the normal operation of the system, thereby realizing comprehensive analysis and visual display of multiple forest supervision aspects and improving the reliability of the system.

[0098] The embodiment of the present application also provides a method for long-term monitoring of forest ecosystems based on machine learning, which is applied to the control terminal 102 of the above-mentioned long-term monitoring system of forest ecosystems based on machine learning, such as Figure 3 Shown, including:

[0099] S301, acquiring multi-source detection data of a forest detection area; the multi-source detection data includes visible image data, infrared image data, three-dimensional scanning data, and meteorological parameter data.

[0100] Specifically, the control terminal obtains visible image data, infrared image data, and 3D scanning data of the forest detection area from the sensor, and obtains meteorological parameter data of the detection area from the server. The visible image data, infrared image data, 3D scanning data, and meteorological parameter data are collectively referred to as multi-source detection data. Schematically, the 3D scanning data is scanned data obtained through LiDAR 3D scanning.

[0101] S302: Establish a distributed model training mechanism through the federated learning framework to clean and integrate multi-source detection data to obtain ecological feature vectors.

[0102] Specifically, the control terminal cleans and integrates multi-source detection data to obtain initial ecological characteristic data for each edge monitoring point. Using the federated learning framework, a distributed model training mechanism is established to globally integrate the initial ecological characteristic data from each edge point to obtain a global ecological characteristic vector. Schematically, the number of edge monitoring points participating in federated learning can be set based on actual forest monitoring needs and budget requirements.

[0103] S303 uses a multi-task deep neural network model and multi-branch analysis of ecological feature vectors to obtain biodiversity identification results, vegetation index inversion results, and environmental anomaly detection results.

[0104] Specifically, the control terminal uses a multi-task deep neural network model to perform multi-branch, multi-task analysis on ecological feature vectors, generating results for biodiversity identification, vegetation index inversion, and environmental anomaly detection. Illustratively, the multi-task deep neural network model uses the input fusion vector and appropriately sets task-specific headers to generate analysis results for the corresponding tasks. The multi-task branches include, but are not limited to, biodiversity identification, vegetation index inversion, and environmental anomaly detection.

[0105] S304: Generate an ecological evolution digital twin based on the ecological characteristic vector, biodiversity identification results, vegetation index inversion results, and environmental anomaly detection results. The ecological evolution digital twin is used to dynamically display trend maps of the detection area and generate ecological early warning signals.

[0106] Specifically, the control terminal calculates the characteristic components that contribute to biodiversity identification, vegetation index inversion, and environmental anomaly detection based on ecological eigenvectors. Based on these components and a preset geographic space, it generates an ecological evolution digital twin to dynamically display trend maps of the detected area and generate ecological warning signals. Illustratively, the ecological evolution digital twin display includes, but is not limited to, 3D ecological map rendering, dynamic trend map generation, and ecological warning signals.

[0107] This embodiment obtains visible image data, infrared image data, three-dimensional scanning data and meteorological parameter data of the detection area, integrates the data of each edge monitoring point through federated learning, and adopts a multi-task deep neural network model to obtain the corresponding biodiversity identification results, vegetation index inversion results and environmental anomaly detection results. Based on the biodiversity identification results, vegetation index inversion results and environmental anomaly detection results, an ecological evolution digital twin is generated, which realizes comprehensive analysis and visualization of multiple aspects of forest supervision, thereby improving the efficiency of forest supervision.

[0108] In one embodiment, a multi-task deep neural network model is used to analyze ecological feature vectors in multiple branches to obtain biodiversity identification results, vegetation index inversion results, and environmental anomaly detection results, including:

[0109] S401, based on the multi-task neural network model, analyzes the ecological feature vectors to obtain preliminary biodiversity identification results, preliminary vegetation index inversion results, preliminary environmental anomaly detection results and fusion shared features.

[0110] Specifically, a multi-task neural network model is employed, along with a multi-task-specific head. This performs a multi-branch, multi-task analysis on ecological feature vectors, yielding preliminary biodiversity identification results, preliminary vegetation index inversion results, and preliminary environmental anomaly detection results. Schematically, the fused shared features are shared vectors generated by the multi-task neural network model, which are used by the multi-task-specific head to calculate probabilities for different prediction features.

[0111] S402 , using a Monte Carlo Dropout confidence algorithm, adjusts the preliminary biodiversity identification results, preliminary vegetation index inversion results, and preliminary environmental anomaly detection results based on the fused shared features to obtain biodiversity identification results, vegetation index inversion results, and environmental anomaly detection results.

[0112] Specifically, a Monte Carlo Dropout confidence algorithm is used to calculate the confidence levels of the preliminary biodiversity identification results, preliminary vegetation index inversion results, and preliminary environmental anomaly detection results. Response levels are then assigned to each of these results based on pre-defined response level rules. These response levels are then combined with the preliminary biodiversity identification results, preliminary vegetation index inversion results, and preliminary environmental anomaly detection results to produce the biodiversity identification results, vegetation index inversion results, and environmental anomaly detection results. Schematically, the response level rules can be set based on the accuracy requirements for prediction results in actual practice.

[0113] This embodiment sets a multi-task neural network model and analyzes ecological feature vectors to obtain preliminary biodiversity identification results, preliminary vegetation index inversion results, preliminary environmental anomaly detection results, and fused shared features. Confidence calculation is performed on the obtained results based on the Monte Carlo Dropout confidence algorithm. The obtained confidence is combined with the preset response level rule and the preliminary biodiversity identification results, preliminary vegetation index inversion results, and preliminary environmental anomaly detection results to obtain biodiversity identification results, vegetation index inversion results, and environmental anomaly detection results. The confidence algorithm is used to verify the accuracy of the prediction results of the multi-task neural network model, thereby improving the reliability of the biodiversity identification results, vegetation index inversion results, and environmental anomaly detection results.

[0114] In one embodiment, based on a multi-task neural network model, ecological feature vectors are analyzed to obtain preliminary biodiversity identification results, preliminary vegetation index inversion results, preliminary environmental anomaly detection results, and fused shared features, including:

[0115] In S501, the following formula is used to extract cross-task features of the ecological feature vector through the fully connected layer and the dynamic feature routing mechanism to obtain the fused shared features:

[0116]

[0117] Among them, h s is the fusion shared feature, F g is the ecological characteristic vector, c ij is the dynamic routing weight, W ij is the capsule weight from input capsule i to output capsule j, b ij is the number of routing pairs from input capsule i to output capsule j;

[0118] Among them, the iterative update formula of dynamic routing weight is:

[0119]

[0120] Among them, c ij is the dynamic routing weight, b ij is the number of routing pairs from input capsule i to output capsule j, b ik is the number of routing pairs from input capsule i to output capsule k.

[0121] Specifically, a dynamic feature routing mechanism is adopted, and a fully connected layer is set based on the capsule concept to aggregate cross-task features of the ecological feature vector to obtain fused shared features. Schematically, three input capsules can be set according to the needs of three tasks: biodiversity identification, vegetation index inversion, and environmental anomaly detection. The number of input capsules can also be increased or decreased according to the settings of more prediction tasks in actual work. Optionally, the fused shared feature h s is a shared feature vector used by the multi-task head for multi-task prediction. Schematically, the dynamic routing weight c ij It is used to represent the probability distribution of the capsule from input capsule i to output capsule j. Optionally, the capsule weight W from input capsule i to output capsule j ij It is used to represent the spatial relationship between input capsule i and output capsule j. Schematically, the number of routing pairs b from input capsule i to output capsule j is ij represents the number of routing pairs from input capsule i to output capsule j, which is used to dynamically adjust the connection strength from input capsule i to output capsule j. Optionally, the number of routing pairs b from input capsule i to output capsule k is ik Used to represent the number of routing pairs from input capsule i to output capsule k, in calculating the dynamic routing weight c ij is used for normalization.

[0122] S502, performing multi-category classification on the fused shared features to obtain preliminary biodiversity identification results, preliminary vegetation index inversion results, and preliminary environmental anomaly detection results;

[0123] Among them, the following formula is used to balance the multi-class classification weights using uncertainty weighting:

[0124]

[0125] in, is the total loss for multi-class classification, is the loss function of the i-th task, is the noise variance of the ith task.

[0126] Specifically, a multi-task neural network model is used to perform multi-category classification on the fused shared features to obtain preliminary biodiversity identification results, preliminary vegetation index inversion results, and preliminary environmental anomaly detection results. The total loss of the three task branches of the multi-task neural network model is calculated by the formula to optimize the model. Schematically, the total loss of multi-category classification is Represents the comprehensive loss of the multi-task neural network model on multiple tasks, which is used to optimize model parameters and balance the losses of multiple tasks. Optionally, the loss function of the i-th task It can be the loss function of biodiversity identification, the loss function of vegetation index inversion result prediction and the loss function of environmental anomaly detection. Schematically, the noise variance of the i-th task is It can be the noise variance of biodiversity identification, the noise variance of vegetation index inversion result prediction, and the noise variance of environmental anomaly detection.

[0127] This embodiment adopts a dynamic feature routing mechanism to extract cross-task features of ecological feature vectors to obtain fused shared features. Based on the multi-task neural network model, the fused shared features are predicted for three tasks respectively to obtain preliminary biodiversity identification results, preliminary vegetation index inversion results and preliminary environmental anomaly detection results. A total loss function is set to optimize the parameters of the multi-task neural network model, thereby improving the accuracy of the results predicted by the multi-task neural network model.

[0128] In one embodiment, a Monte Carlo Dropout confidence algorithm is used to adjust preliminary biodiversity identification results, preliminary vegetation index inversion results, and preliminary environmental anomaly detection results based on fused shared features to obtain biodiversity identification results, vegetation index inversion results, and environmental anomaly detection results, including:

[0129] S601, based on the Dropout sampling inference algorithm, forward inference is performed on the fused shared features to obtain a confident prediction result.

[0130] Specifically, based on the Dropout sampling inference algorithm, forward inference is performed on the fused shared features to simulate the posterior distribution of the model parameters and obtain a confident prediction result. For example, the number of forward inferences is set to 50 by default, but can be set to a higher value based on the accuracy requirements of the prediction results in actual work.

[0131] S602: Calculate the prediction confidence variance corresponding to the confidence prediction result using the following formula:

[0132]

[0133] in, is the prediction confidence variance, T is the number of samples, y t is the confidence prediction result of the t-th sampling.

[0134] Specifically, the confidence variance is calculated based on the confidence prediction result to obtain the prediction confidence variance, which is used to reflect the confidence of the confidence prediction result. It is used to reflect the confidence of the confidence prediction result. The larger the prediction confidence variance is, the lower the confidence of the confidence prediction result is. Optionally, the sampling number Y is the number of forward reasonings set in S601. In principle, the confidence prediction result y of the tth sampling is tis the confidence prediction result obtained in S601.

[0135] S603 , combining the prediction confidence variance with the preliminary biodiversity identification result, the preliminary vegetation index inversion result, and the preliminary environmental anomaly detection result to obtain corresponding biodiversity identification result, vegetation index inversion result, and environmental anomaly detection result.

[0136] Specifically, based on the obtained prediction confidence variance, the corresponding response level is determined according to the preset response level rules. The response level is then combined with the preliminary biodiversity identification results, preliminary vegetation index inversion results, and preliminary environmental anomaly detection results to obtain the biodiversity identification results, vegetation index inversion results, and environmental anomaly detection results. For example, the preset response level rules can refer to Table 1, or they can be set based on the accuracy requirements of the prediction results in actual work:

[0137] Table 1: Example table of response level rules

[0138] Response Level Prediction confidence variance Low <![CDATA[σ c ≤0.1]]> middle <![CDATA[0.1<σ c <0.3]]> high <![CDATA[σ c ≥0.3]]>

[0139] This embodiment uses the Monte Carlo Dropout confidence algorithm to perform forward reasoning on the fused shared features to obtain the corresponding prediction confidence, and judges the prediction confidence according to preset rules to obtain a response level. It then combines the preliminary biodiversity identification results, preliminary vegetation index inversion results, and preliminary environmental anomaly detection results to obtain biodiversity identification results, vegetation index inversion results, and environmental anomaly detection results. The confidence calculation is used to judge the prediction accuracy of the multi-task deep neural network model, demonstrate the credibility of the prediction results, and provide a reliable data basis for the display of detection data.

[0140] In one embodiment, a distributed model training mechanism is established through a federated learning framework to clean and fuse multi-source detection data to obtain ecological feature vectors, including:

[0141] S701, using isolation forest technology, cleans visible image data, infrared image data, three-dimensional scanning data and meteorological parameter data to obtain corresponding ideal visible image data, ideal infrared image data, ideal three-dimensional scanning data and ideal meteorological parameter data.

[0142] Specifically, based on the isolation forest technique, a binary decision tree is constructed based on the visible image data, infrared image data, 3D scan data, and meteorological parameter data. The path length of each sample in the tree is calculated and compared with a preset length threshold. Samples with path lengths greater than the threshold are identified as abnormal data. This abnormal data is then interpolated and supplemented to obtain the corresponding ideal visible image data, ideal infrared image data, ideal 3D scan data, and ideal meteorological parameter data. Illustratively, interpolation and supplementation can be based on a completion algorithm in the spatiotemporal k-domain.

[0143] S702 , linear projection and feature extraction are performed on the ideal visible image data, ideal infrared image data, ideal three-dimensional scanning data, and ideal meteorological parameter data to obtain corresponding visible image features, infrared image features, three-dimensional scanning features, and meteorological parameter features.

[0144] Specifically, linear projection and feature extraction are performed on the ideal visible image data, ideal infrared image data, ideal 3D scan data, and ideal meteorological parameter data, and then mapped to a unified feature space to obtain corresponding visible image features, infrared image features, 3D scan features, and meteorological parameter features. Illustratively, a convolutional neural network (CNN) based on machine learning can be used to extract and map the features of the ideal visible image data, ideal infrared image data, ideal 3D scan data, and ideal meteorological parameter features.

[0145] S703, using the following formula, the fusion weight is calculated based on the attention mechanism:

[0146]

[0147] Among them, h i is the fusion weight of the i-th head, softmax is the classification function, Q i is the query matrix of the i-th head, is the key matrix of the i-th head, d is the feature dimension, H is the total number of attention heads, P is the position encoding matrix, V i is the value matrix of the i-th head.

[0148] Specifically, according to the formula, a multi-head attention mechanism is used to dynamically allocate modal weights and calculate the fusion weights corresponding to visible image features, infrared image features, 3D scanning features, and meteorological parameter features. Schematically, the fusion weight h of the i-th head is iIt can be the fusion weight of any feature among visible image features, infrared image features, 3D scanning features and meteorological parameter features, and feature i can be any feature among visible image features, infrared image features, 3D scanning features and meteorological parameter features. Optionally, the classification function softmax is used to convert the input vector into a probability distribution to ensure that the sum of the weights is 1. Schematically, the query matrix Q of the i-th head is i Used to extract key information of feature i. Optionally, the key matrix of the i-th head Used to match feature i. Schematically, the feature dimension d is the sum of the dimensions of visible image features, infrared image features, 3D scanning features, and meteorological parameter features. Optionally, the total number of attention heads H is used to learn multiple sets of weights in parallel; the position encoding matrix P is used to inject the relative position information of the space where the feature is located; the value matrix V of the i-th head i Used to pass feature data of feature i.

[0149] S704, introduce a gating mechanism to dynamically adjust the fusion weights, dynamically fuse the visible image features, infrared image features, 3D scanning features, and meteorological parameter features, and obtain a primary fusion feature vector.

[0150] Specifically, by introducing a gating mechanism to optimize the attention weights, adaptive fusion is achieved, the aforementioned fusion weights are dynamically adjusted, and the fusion weights and corresponding features are weightedly combined to obtain a primary fusion feature vector.

[0151] S705 , using principal component analysis technology to compress the primary fusion feature vector to obtain a fusion feature vector.

[0152] Specifically, principal component analysis is used to select the first k principal components of the primary fusion feature vector, retaining n% of the information and reducing feature redundancy to obtain the fused feature vector. In principle, k and n can be set based on the actual requirements for feature fusion, prediction accuracy, and budget. The default setting is n = 95.

[0153] S706, based on the federated learning technology, globally integrate the fused feature vector to obtain the ecological feature vector.

[0154] Specifically, the control terminal integrates the data of edge points involved in data detection based on federated learning technology to obtain an ecological feature vector that contains a global primary fusion feature vector.

[0155] This embodiment adopts the isolation forest technology to perform data cleaning on visible image data, infrared image data, three-dimensional scanning data and meteorological parameter data to obtain corresponding ideal visible image data, ideal infrared image data, ideal three-dimensional scanning data and ideal meteorological parameter data, and performs linear projection and feature extraction on the features, and maps them to a unified feature space to obtain corresponding visible image features, infrared image features, three-dimensional scanning features and meteorological parameter features. Based on the attention mechanism and the introduction of the gating mechanism, the fusion weights are calculated and dynamically fused to obtain a primary fusion feature vector. The principal component analysis technology is adopted to extract the principal components of the primary fusion feature vector and remove feature redundancy to obtain a fusion feature vector. The global fusion feature vector is integrated based on the federated learning technology to obtain an ecological feature vector, which effectively improves the quality of data cleaning, realizes adaptive weight distribution of meteorological conditions, and integrates the global sensor data for fusion, providing a reliable data basis for the prediction of the multi-task neural network model.

[0156] In one embodiment, based on federated learning technology, the fused feature vectors are globally integrated to obtain an ecological feature vector, including:

[0157] According to the fusion feature vector, based on the federated learning technology, the ecological feature vector is calculated using the following formula:

[0158]

[0159] Among them, F g is the ecological feature vector, N is the total number of edge points participating in federated learning, is the fusion feature vector of the i-th edge point, is the alignment matrix of the i-th edge point.

[0160] Specifically, the control terminal integrates the fusion feature vectors of all edge points participating in federated learning to obtain the ecological feature vector. Schematically, the ecological feature vector F g The feature vector is obtained by integrating the global fusion feature vector, which contains the sensor data features of all edge points participating in the federated learning. Optionally, the total number of edge points N participating in the federated learning can be set according to the number and type of sensors participating in the detection of the monitoring system in actual work. Schematically, the fusion feature vector of the i-th edge point is is the fused feature vector obtained by S705 for the i-th edge point. Optionally, the alignment matrix of the i-th edge point is Used to ensure the feature space consistency of the fused feature vectors of all edge points.

[0161] This embodiment integrates the global fusion feature vector based on the federated learning technology to obtain the ecological feature vector, integrates the global sensor data for fusion, and provides a reliable data basis for the prediction of the multi-task neural network model.

[0162] In one embodiment, an ecological evolution digital twin is generated based on ecological feature vectors, biodiversity identification results, vegetation index inversion results, and environmental anomaly detection results, including:

[0163] S901, based on the ecological feature vector, biodiversity identification results, vegetation index inversion results and environmental anomaly detection results, the distribution feature contribution is calculated based on cooperative game theory to obtain the corresponding biodiversity contribution, vegetation index contribution and environmental anomaly contribution.

[0164] Specifically, the control terminal uses cooperative game theory to calculate the contribution of each feature in the ecological feature vector to the biodiversity identification results, vegetation index inversion results, and environmental anomaly detection results, respectively, to obtain the corresponding biodiversity contribution, vegetation index contribution, and environmental anomaly contribution. Schematically, the Shapley value calculation formula is used to calculate the contribution of each feature in the ecological feature vector to the biodiversity identification results, vegetation index inversion results, and environmental anomaly detection results.

[0165] S902 , projecting the biodiversity contribution, vegetation index contribution, and environmental anomaly contribution into a preset detection area to obtain a feature importance heat map.

[0166] Specifically, the biodiversity contribution, vegetation index contribution, and environmental anomaly contribution are projected onto the geographic space of the pre-set detection area to generate a feature importance heat map. Illustratively, a larger value for a coordinate in the feature importance heat map indicates a more significant impact of that region's feature on the prediction. Optionally, the pre-set detection area's geographic space can be in the form of geographic grid cells.

[0167] S903: constructing a detection area network model based on the preset three-dimensional scanning model of the detection area.

[0168] Specifically, the control terminal constructs a network model of the detection area based on a preset 3D scanning model of the detection area. Illustratively, the preset 3D scanning model of the detection area can be constructed based on the 3D scanning data obtained in S301, or it can be constructed based on the preset 3D scanning data, where the 3D scanning model only displays the geographical characteristics of the detection area.

[0169] S904 , performing vertex association between the biodiversity identification results, the vegetation index inversion results, and the environmental anomaly detection results and the detection region network model to obtain a fused region network model.

[0170] Specifically, the control terminal associates the biodiversity identification results, vegetation index inversion results, and environmental anomaly detection results with the detection area network model. It then accurately projects their location information onto the detection area network model to create a fused regional network model. This fused regional network model illustratively displays the biodiversity identification results, vegetation index inversion results, and environmental anomaly detection results within the coordinate system of a 3D scan model of the detection area.

[0171] S905, mapping the feature importance heat map into the texture and transparency of the fusion regional network model to obtain the ecological evolution digital twin.

[0172] Specifically, the feature importance heatmap is mapped to the texture and transparency of the fused regional network model to obtain the ecological evolution digital twin. In principle, the lower the confidence in the feature importance heatmap, the more transparent the fused regional network model.

[0173] This embodiment calculates the contribution of the features in the ecological feature vector to the biodiversity identification results, vegetation index inversion results and environmental anomaly detection results, calculates the feature importance heat map, constructs a detection area network model based on three-dimensional scanning data, and associates and accurately projects the biodiversity identification results, vegetation index inversion results and environmental anomaly detection results with the detection area network model to obtain a fused area network model, maps the feature importance heat map to the texture and transparency of the fused area network model, and obtains an ecological evolution digital twin, so as to realize the visualization of biodiversity identification results, vegetation index inversion results and environmental anomaly detection results in a three-dimensional model, thereby improving the efficiency of forest supervision.

[0174] In one embodiment, the method further comprises:

[0175] S1001: Divide the ecological feature vectors, biodiversity identification results, vegetation index inversion results, and environmental anomaly detection results into blocks according to time and space windows to obtain standardized data blocks.

[0176] Specifically, the control terminal divides the ecological feature vectors, biodiversity identification results, vegetation index inversion results, and environmental anomaly detection results into blocks according to spatiotemporal windows, generating standardized data blocks. For example, the spatiotemporal window block format can be configured based on the actual data storage standards used in practice, such as a 1km×1km grid with a 1-hour time slice.

[0177] S1002, using the SHA-256 algorithm to generate a unique identifier corresponding to the standardized data block to obtain a data hash.

[0178] Specifically, the control terminal uses the SHA-256 algorithm to convert the standardized data block into a uniquely identified data hash.

[0179] S1003, based on the data hash and the standardized data block, the blockchain technology is used for storage to obtain the analysis result storage information.

[0180] Specifically, the control terminal uses blockchain technology to obtain the corresponding block hash based on the data hash and standardized data block. It then adds the standardized data block to the IPFS network, obtaining a corresponding unique CID (content identifier). Based on the block hash and CID, a spatiotemporal index tree is constructed to store the analysis results. This storage allows for sub-second spatiotemporal data retrieval, enabling long-term analysis of biodiversity identification, vegetation index inversion, and environmental anomaly detection.

[0181] This embodiment divides ecological feature vectors, biodiversity identification results, vegetation index inversion results, and environmental anomaly detection results into blocks according to time and space windows to obtain standardized data blocks. It uses hash algorithms and blockchain technology to achieve distributed storage and version traceability of monitoring data, retaining data support for forest supervision decisions and providing data support for long-term analysis.

[0182] In the above-mentioned long-term monitoring method of forest ecosystems based on machine learning, multi-source detection data of the forest detection area is obtained; the multi-source detection data includes visible image data, infrared image data, three-dimensional scanning data and meteorological parameter data; a distributed model training mechanism is established through a federated learning framework, and the multi-source detection data is cleaned and integrated to obtain an ecological feature vector; a multi-task deep neural network model is used to analyze the ecological feature vector in multiple branches to obtain biodiversity identification results, vegetation index inversion results and environmental anomaly detection results; based on the ecological feature vector, biodiversity identification results, vegetation index inversion results and environmental anomaly detection results, an ecological evolution digital twin is generated; the ecological evolution digital twin is used to dynamically display the trend map of the detection area and generate ecological early warning signals, realize the integration of visible image data, infrared image data, three-dimensional scanning data and meteorological parameter data, conduct comprehensive analysis and visualization of multiple forest supervision aspects, and improve the efficiency of forest supervision.

[0183] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0184] The above-described embodiments merely represent several implementation methods of the embodiments of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that a person skilled in the art may make various modifications and improvements without departing from the concept of the embodiments of the present application, and these modifications and improvements fall within the scope of protection of the embodiments of the present application.

Claims

1. A long-term monitoring system for forest ecosystems based on machine learning, characterized in that: The system comprises: Sensors for acquiring sensing data of forest detection areas; A control terminal, connected to the sensor, configured to analyze and process the sensor data, obtain monitoring results, and generate an ecological evolution digital twin based on the monitoring results; a server connected to the control terminal and configured to provide meteorological information of the forest detection area; Visualization hardware, connected to the control terminal, for displaying the monitoring results based on the ecological evolution digital twin; The analyzing and processing of the sensor data to obtain monitoring results includes: Acquiring multi-source detection data of the forest detection area; the multi-source detection data includes visible image data, infrared image data, three-dimensional scanning data and meteorological parameter data; Establishing a distributed model training mechanism through a federated learning framework to clean and fuse the multi-source detection data to obtain an ecological feature vector; A multi-task deep neural network model is used to perform multi-branch analysis of the ecological feature vectors to obtain biodiversity identification results, vegetation index inversion results, and environmental anomaly detection results; An ecological evolution digital twin is generated based on the ecological feature vector, the biodiversity identification result, the vegetation index inversion result and the environmental anomaly detection result; the ecological evolution digital twin is used to dynamically display the trend map of the detection area and generate an ecological early warning signal.

2. The system according to claim 1, wherein: The system further comprises: A mobile unit, connected to the sensor, and configured to drive the sensor to move and collect the sensing data; An energy supply unit is connected to the sensor, the control terminal, the server, the visualization hardware and the mobile unit, and is used to provide energy.

3. A long-term monitoring method for forest ecosystems based on machine learning, applied to the control terminal according to claim 1 or 2, characterized in that: include: Acquiring multi-source detection data of the forest detection area; the multi-source detection data includes visible image data, infrared image data, three-dimensional scanning data and meteorological parameter data; Establishing a distributed model training mechanism through a federated learning framework to clean and fuse the multi-source detection data to obtain an ecological feature vector; A multi-task deep neural network model is used to perform multi-branch analysis of the ecological feature vectors to obtain biodiversity identification results, vegetation index inversion results, and environmental anomaly detection results; An ecological evolution digital twin is generated based on the ecological feature vector, the biodiversity identification result, the vegetation index inversion result and the environmental anomaly detection result; the ecological evolution digital twin is used to dynamically display the trend map of the detection area and generate an ecological early warning signal.

4. The method according to claim 3, characterized in that The multi-task deep neural network model is used to analyze the ecological feature vector in multiple branches to obtain biodiversity identification results, vegetation index inversion results and environmental anomaly detection results, including: Based on a multi-task neural network model, the ecological feature vectors are analyzed to obtain preliminary biodiversity identification results, preliminary vegetation index inversion results, preliminary environmental anomaly detection results and fusion shared features; A Monte Carlo Dropout confidence algorithm is adopted to adjust the preliminary biodiversity identification result, the preliminary vegetation index inversion result and the preliminary environmental anomaly detection result according to the fused shared features to obtain the biodiversity identification result, the vegetation index inversion result and the environmental anomaly detection result.

5. The method according to claim 4, characterized in that The multi-task neural network model is used to analyze the ecological feature vector to obtain preliminary biodiversity identification results, preliminary vegetation index inversion results, preliminary environmental anomaly detection results, and fusion shared features, including: The following formula is used to extract cross-task features of the ecological feature vector through a fully connected layer and a dynamic feature routing mechanism to obtain fused shared features: Among them, h s is the fusion shared feature, F g is the ecological characteristic vector, c ij is the dynamic routing weight, W ij is the capsule weight from input capsule i to output capsule j, b ij is the number of routing pairs from input capsule i to output capsule j; Among them, the iterative update formula of dynamic routing weight is: Among them, c ij is the dynamic routing weight, b ij is the number of routing pairs from input capsule i to output capsule j, b ik is the number of routing pairs from input capsule i to output capsule k; Performing multi-category classification on the fused shared features to obtain the preliminary biodiversity identification result, the preliminary vegetation index inversion result, and the preliminary environmental anomaly detection result; Among them, the following formula is used to balance the multi-class classification weights using uncertainty weighting: in, is the total loss for multi-class classification, is the loss function of the i-th task, is the noise variance of the ith task.

6. The method according to claim 4, characterized in that The Monte Carlo Dropout confidence algorithm is used to adjust the preliminary biodiversity identification result, the preliminary vegetation index inversion result, and the preliminary environmental anomaly detection result according to the fusion shared features to obtain the biodiversity identification result, the vegetation index inversion result, and the environmental anomaly detection result, including: Based on the Dropout sampling inference algorithm, the fused shared features are forward inferred to obtain a confident prediction result; Use the following formula to calculate the prediction confidence variance corresponding to the confidence prediction result: in, is the prediction confidence variance, T is the number of samples, y t is the confidence prediction result of the t-th sampling; The prediction confidence variance is combined with the preliminary biodiversity identification result, the preliminary vegetation index inversion result, and the preliminary environmental anomaly detection result to obtain the corresponding biodiversity identification result, the vegetation index inversion result, and the environmental anomaly detection result.

7. The method according to claim 3, characterized in that The distributed model training mechanism is established through the federated learning framework to clean and fuse the multi-source detection data to obtain the ecological feature vector, including: Using isolation forest technology, the visible image data, infrared image data, three-dimensional scanning data and meteorological parameter data are cleansed to obtain corresponding ideal visible image data, ideal infrared image data, ideal three-dimensional scanning data and ideal meteorological parameter data; performing linear projection and feature extraction on the ideal visible image data, the ideal infrared image data, the ideal three-dimensional scanning data, and the ideal meteorological parameter data to obtain corresponding visible image features, infrared image features, three-dimensional scanning features, and meteorological parameter features; The fusion weight is calculated based on the attention mechanism using the following formula: Among them, h i is the fusion weight of the i-th head, softmax is the classification function, Q i is the query matrix of the i-th head, is the key matrix of the i-th head, d is the feature dimension, H is the total number of attention heads, P is the position encoding matrix, V i is the value matrix of the i-th head; A gating mechanism is introduced to dynamically adjust the fusion weights, and the visible image features, the infrared image features, the three-dimensional scanning features, and the meteorological parameter features are dynamically fused to obtain a primary fusion feature vector; Using principal component analysis technology, compressing the primary fusion feature vector to obtain a fusion feature vector; Based on the federated learning technology, the fusion feature vector is globally integrated to obtain the ecological feature vector.

8. The method according to claim 7, characterized in that The federated learning technology is used to globally integrate the fusion feature vector to obtain the ecological feature vector, including: According to the fusion feature vector, based on the federated learning technology, the ecological feature vector is calculated using the following formula: Among them, F g is the ecological feature vector, N is the total number of edge points participating in federated learning, is the fusion feature vector of the i-th edge point, is the alignment matrix of the i-th edge point.

9. The method according to claim 3, characterized in that The generating of an ecological evolution digital twin according to the ecological feature vector, the biodiversity identification result, the vegetation index inversion result, and the environmental anomaly detection result includes: Calculate and allocate feature contributions based on the cooperative game theory according to the ecological feature vector, the biodiversity identification result, the vegetation index inversion result, and the environmental anomaly detection result, to obtain corresponding biodiversity contribution, vegetation index contribution, and environmental anomaly contribution; Projecting the biodiversity contribution, vegetation index contribution, and environmental anomaly contribution into a preset detection area to obtain a feature importance heat map; Based on the preset three-dimensional scanning model of the detection area, a detection area network model is constructed; Associating the biodiversity identification results, vegetation index inversion results, and environmental anomaly detection results with the detection area network model to obtain a fusion area network model; The feature importance heat map is mapped to the texture and transparency of the fusion regional network model to obtain the ecological evolution digital twin.

10. The method according to claim 3, characterized in that The method further comprises: Dividing the ecological feature vector, the biodiversity identification result, the vegetation index inversion result, and the environmental anomaly detection result into blocks according to a spatiotemporal window to obtain standardized data blocks; Using the SHA-256 algorithm, a unique identifier corresponding to the standardized data block is generated to obtain a data hash; According to the data hash and the standardized data block, blockchain technology is used for storage to obtain analysis result storage information.

Images