Ecological environment sensitive area dynamic change evaluation and prediction method based on machine learning

Abstract

The invention provides an ecological environment sensitive area dynamic change evaluation and prediction method based on machine learning, and relates to the technical field of ecological environment monitoring, and the method comprises the steps: obtaining and preprocessing remote sensing image data, and extracting an environment element feature sequence; adopting a hierarchical memory distillation network to extract periodic characteristics of the long-term change sequence, and combining with the short-term change sequence to carry out space-time recombination; utilizing a graph neural network to construct a dynamic association graph to analyze an environmental element coupling relationship; and evaluating the sensitivity and performing early warning based on the sudden change amplitude and the cumulative change trend. According to the invention, the dynamic change characteristics of the environmental elements can be accurately captured, and the accuracy and timeliness of early warning of the ecological environment sensitive area are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of ecological environment monitoring, and in particular to a method for evaluating and predicting dynamic changes in ecological environment sensitive areas based on machine learning. Background Art

[0002] Ecologically sensitive areas are areas that are sensitive to human activities or have important ecological functions. Their dynamic changes have a significant impact on regional ecological security and sustainable development. With the development of remote sensing and artificial intelligence technologies, machine learning-based methods for assessing and predicting dynamic changes in ecologically sensitive areas have been widely used. Traditional methods mainly rely on expert experience and statistical models for assessment, and monitor changes in environmental factors through remote sensing image interpretation and geographic information system analysis.

[0003] However, existing technologies still have shortcomings. It is difficult to effectively extract and integrate the long-term cyclical changes and short-term sudden changes of environmental factors, resulting in incomplete description of environmental changes by the model; the complex coupling relationship between environmental factors is difficult to accurately characterize, which reduces the reliability of the assessment results; there is a lack of a comprehensive assessment mechanism for sudden and cumulative changes in environmental factors, which affects the timeliness and accuracy of early warnings.

[0004] In summary, there is an urgent need for a method for evaluating and predicting the dynamic changes in ecological and environmental sensitive areas based on a hierarchical memory distillation network and a graph neural network. The memory distillation mechanism can effectively integrate long-term and short-term features, and the graph neural network can be used to capture the dynamic coupling relationship between environmental elements. A sensitivity scoring function is constructed based on the mutation amplitude and cumulative trend, thereby improving the accuracy and reliability of the evaluation and prediction. Summary of the Invention

[0005] The embodiment of the present invention provides a method for evaluating and predicting dynamic changes in ecological and environmental sensitive areas based on machine learning, which can solve the problems in the existing technology.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] A method for evaluating and predicting dynamic changes in ecologically sensitive areas based on machine learning is provided, including:

[0008] Acquire remote sensing image data of the target ecological and environmental sensitive area within a preset time period, perform orthorectification and atmospheric correction, and obtain preprocessed image data;

[0009] Extracting environmental element feature sequences from pre-processed image data and dividing them into long-term change sequences and short-term change sequences according to preset time intervals;

[0010] A hierarchical memory distillation network is constructed to extract the periodic features of long-term change sequences through the main memory unit. The main memory features are mapped to the distilled features reconstructed from the memory units based on the temporal distillation factor. The distilled features are then spatially recombined with the short-term change sequence to obtain fused spatiotemporal features.

[0011] Based on the graph neural network, the fused spatiotemporal features are mapped into a dynamic correlation graph. The coupling relationship between environmental elements is calculated through node embedding and edge attention mechanism to generate the state vector of the target ecological and environmental sensitive area.

[0012] According to the temporal change characteristics of the state vector, a sensitivity scoring function is constructed based on the mutation amplitude and cumulative change trend of environmental factors, and the sensitivity score is calculated. When the sensitivity score exceeds the preset sensitivity threshold, dynamic change warning information of the ecological and environmental sensitive area is output.

[0013] In an optional embodiment,

[0014] A hierarchical memory distillation network is constructed to extract the periodic features of the long-term change sequence through the main memory unit. Based on the temporal distillation factor, the main memory features are mapped to the distilled features reconstructed from the memory unit. The distilled features are reorganized in time and space with the short-term change sequence to obtain the fused spatiotemporal features including:

[0015] Constructing a main memory unit, wherein a plurality of convolution kernels are provided in the main memory unit, matching the receptive field size of the convolution kernel with the periodic characteristics of the long-term variation sequence, performing a convolution operation on the long-term variation sequence to obtain a periodic feature map, and performing a nonlinear activation function mapping on the periodic feature map to obtain a main memory feature;

[0016] Constructing a slave memory unit, calculating a temporal distillation factor in the slave memory unit based on the master memory feature, wherein the temporal distillation factor is used to adjust the mapping strength of features at different time scales, mapping the master memory feature into a feature space using the temporal distillation factor, reconstructing the mapped feature using a decoder, and optimizing the reconstructed feature using a reconstruction loss function to obtain a distilled feature;

[0017] The distilled features and short-term change sequences are divided into time series segments, and the time series association is constructed through position encoding. The time series dependency is obtained by forward and backward state update. The features are fused with the weight matrix, and the fused spatiotemporal features are obtained by multi-scale convolution processing.

[0018] In an optional embodiment,

[0019] The distilled features and short-term change sequences are divided into time series segments, and the time series association is constructed through position encoding. The time series dependency is obtained by forward and backward state update. The features are fused with the weight matrix and multi-scale convolution processing is used to obtain the fused spatiotemporal features including:

[0020] Divide the distilled features and short-term change sequences into multiple time series segments according to the preset time step;

[0021] Mapping the position index and feature dimension index of the time series segment through sine function and cosine function to obtain a position code, and superimposing the position code with the distillation feature and the short-term change sequence to obtain a coding feature sequence, including a coding distillation feature and a coding feature change sequence;

[0022] The encoded feature sequence is input into the bidirectional gated recurrent unit, and the forward hidden state and backward hidden state are updated through the update gate and reset gate to obtain the bidirectional hidden state feature. The bidirectional hidden state features are then concatenated to obtain the temporal dependency feature.

[0023] Based on the time series dependency features, a fusion weight matrix is ​​obtained through nonlinear transformation and normalization calculation; the fusion weight matrix is ​​weightedly calculated with the coding distillation features, and feature splicing is performed with the coding feature change sequence in each time series segment to obtain a combined feature; the combined feature is input into multiple parallel convolution branches, each of which has a different convolution kernel size, to obtain feature representations of multiple scales; the feature representations of multiple scales are weightedly fused through an adaptive feature fusion layer to obtain a fused spatiotemporal feature.

[0024] In an optional embodiment,

[0025] Based on the graph neural network, the fused spatiotemporal features are mapped into a dynamic correlation graph. The coupling relationship between environmental elements is calculated through node embedding and edge attention mechanism to generate the state vector of the target ecological and environmental sensitive area, including:

[0026] Extracting time series features and spatial distribution features of environmental elements from the fused spatiotemporal features to construct an environmental element feature vector;

[0027] Mapping the environmental element feature vectors into nodes of a dynamic association graph, calculating the feature distance between the nodes, establishing an association edge between the corresponding nodes when the feature distance is less than a preset distance threshold, and determining the weight of the association edge based on the feature similarity between the nodes, thereby generating a dynamic association graph including environmental element nodes and association edges;

[0028] Using a graph neural network to embed features of the environmental element nodes of the dynamic association graph, extracting local structural information of the nodes through graph convolution operations, and obtaining structured feature representations of the environmental element nodes;

[0029] Constructing an edge attention network, calculating the coupling coefficient between the environmental element nodes based on the structured feature representation, and adaptively weighting the node features according to the coupling coefficient to obtain a node representation reflecting the coupling relationship of the environmental elements;

[0030] Based on the reconstruction of spatiotemporal feature flow based on orthogonal decomposition and combined with multi-level weighted fusion, dynamic aggregation operation is performed on the node representation to generate the state vector of the target ecological and environmental sensitive area, which contains the overall coupling pattern of environmental elements in the area.

[0031] In an optional embodiment,

[0032] Based on the reconstruction of spatiotemporal feature flow based on orthogonal decomposition and combined with multi-level weighted fusion, dynamic aggregation operation is performed on the node representation to generate the state vector of the target ecological and environmental sensitive area, including:

[0033] Constructing a spatiotemporal feature flow from the node representation in the spatial domain and time window to generate a spatiotemporal feature flow function;

[0034] performing an orthogonal decomposition operation on the space-time characteristic flow function to obtain a space mode function and a time coefficient corresponding to the space mode function, making the space mode functions orthogonal in the spatial domain, and combining the space mode function and the time coefficient to form a space-time mode;

[0035] Performing a correlation operation on the time coefficient to determine the correlation coefficients between different spatiotemporal modes, obtaining an evolutionary mode characteristic matrix, and determining the dynamic change law between the spatiotemporal modes;

[0036] Performing energy analysis on the spatiotemporal modes, calculating the energy ratio of each spatiotemporal mode to the total energy, superimposing the spatiotemporal modes whose energy ratio exceeds a preset energy ratio threshold, and reconstructing a new spatiotemporal characteristic flow function;

[0037] Constructing a weight function that decays over time, integrating the weight function with the new spatiotemporal characteristic flow function over the entire time window to form a characteristic field that characterizes the dynamic changes of environmental elements;

[0038] A spatial attenuation weight function based on the center position of the target area is constructed, the spatial attenuation weight function and the characteristic field are integrated in the entire spatial domain, and the state vector of the target ecological and environmental sensitive area is obtained through nonlinear transformation.

[0039] In an optional embodiment,

[0040] According to the temporal variation characteristics of the state vector, a sensitivity scoring function is constructed based on the mutation amplitude and cumulative change trend of environmental factors. The calculation of the sensitivity score includes:

[0041] Calculating the rate of change of the state vector at adjacent sampling moments to obtain a time series change feature, and performing a weighted combination of the change amount of the state vector within the historical time window and the time decay weight to obtain a cumulative change trend;

[0042] According to the ratio of the time series change feature to the historical fluctuation level of the state vector, a mutation amplitude function is constructed in combination with the cumulative change trend, and a threshold exceeding point of the state vector is identified based on the mutation amplitude function;

[0043] Performing a convolution operation on the cumulative change trend with a dynamic attenuation function in the time domain to obtain a trend feature representing the long-term evolution law of the state vector;

[0044] Based on the mutation amplitude function, the trend characteristics and the threshold exceeding point, an initial score is constructed, and combined with the phase space reconstruction of the state vector, a nonlinear correction function is determined, and the sensitivity score is calculated by multiplying the initial score and the nonlinear correction function.

[0045] In an optional embodiment,

[0046] Constructing an initial score based on the mutation amplitude function, the trend characteristics, and the threshold exceeding point, determining a nonlinear correction function in combination with phase space reconstruction of the state vector, and calculating a sensitivity score by multiplying the initial score by the nonlinear correction function includes:

[0047] Obtaining a state vector, and performing weighted combination of a mutation amplitude function, a trend feature, and an exponentially decayed threshold overshoot point amplitude of the state vector to obtain an initial score;

[0048] Reconstructing the state vector in phase space, determining an optimal delay time by calculating a first local minimum of a mutual information function, calculating a distance ratio of state vectors of adjacent dimensions, determining a minimum embedding dimension based on the distance ratio being less than a preset distance tolerance threshold, and constructing a state vector sequence according to the optimal delay time and the minimum embedding dimension;

[0049] selecting adjacent state point pairs whose distance is less than a neighborhood radius in the state vector sequence, calculating a separation rate obtained by changing the distance between the adjacent state point pairs over time, and calculating a maximum Lyapunov exponent based on the separation rate;

[0050] Calculating a step function mapping value of the Euclidean distance between adjacent state points in the state vector sequence and a neighborhood radius threshold, normalizing the mapping result to obtain a correlation integral, and calculating a correlation dimension based on the correlation integral;

[0051] The exponential function of the maximum Lyapunov exponent is weightedly combined with the correlation dimension to construct a nonlinear correction function, and the initial score is multiplied by the nonlinear correction function to obtain a sensitivity score.

[0052] In an embodiment of the present invention, by constructing a deep learning model that combines a hierarchical memory distillation network and a graph neural network, it is possible to automatically extract long-term and short-term change characteristics of environmental factors from remote sensing image data, realize accurate assessment and prediction of dynamic changes in ecological and environmental sensitive areas, and effectively improve the automation level and early warning capabilities of environmental monitoring; the distillation mechanism of the master-slave memory unit is adopted, which can effectively capture the periodic laws of changes in environmental factors, and deeply integrate the long-term change characteristics with the short-term change sequence through the time series distillation factor, overcoming the limitation of traditional methods that are difficult to simultaneously process multi-scale time series features, and significantly improving the accuracy and reliability of environmental change trend analysis; the graph neural network is introduced to construct a dynamic association graph of environmental factors, and the coupling relationship between environmental factors is adaptively learned through node embedding and edge attention mechanism. Combined with the sensitivity scoring function of mutation amplitude and cumulative change trend, it can timely detect abnormal changes in the ecological environment and give early warnings, providing a scientific basis and technical support for ecological and environmental protection decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is a flow chart of a method for evaluating and predicting dynamic changes in ecologically sensitive areas based on machine learning according to an embodiment of the present invention;

[0054] Figure 2 This is a dynamic change diagram of the feature fusion weight matrix under different ecological events;

[0055] Figure 3 Bubble chart for spatiotemporal modal correlation analysis;

[0056] Figure 4 It is a three-dimensional phase space projection diagram. DETAILED DESCRIPTION

[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0058] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0059] Figure 1 FIG. 1 is a flow chart of a method for evaluating and predicting dynamic changes in ecologically sensitive areas based on machine learning according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0060] Acquire remote sensing image data of the target ecological and environmental sensitive area within a preset time period, perform orthorectification and atmospheric correction, and obtain preprocessed image data;

[0061] Extracting environmental element feature sequences from pre-processed image data and dividing them into long-term change sequences and short-term change sequences according to preset time intervals;

[0062] A hierarchical memory distillation network is constructed to extract the periodic features of long-term change sequences through the main memory unit. The main memory features are mapped to the distilled features reconstructed from the memory units based on the temporal distillation factor. The distilled features are then spatially recombined with the short-term change sequence to obtain fused spatiotemporal features.

[0063] Based on the graph neural network, the fused spatiotemporal features are mapped into a dynamic correlation graph. The coupling relationship between environmental elements is calculated through node embedding and edge attention mechanism to generate the state vector of the target ecological and environmental sensitive area.

[0064] According to the temporal change characteristics of the state vector, a sensitivity scoring function is constructed based on the mutation amplitude and cumulative change trend of environmental factors, and the sensitivity score is calculated. When the sensitivity score exceeds the preset sensitivity threshold, dynamic change warning information of the ecological and environmental sensitive area is output.

[0065] In one specific implementation, the present invention first acquires multi-temporal remote sensing imagery data of a target ecologically sensitive area over a preset time period. These acquired remote sensing images are then orthorectified, using ground control points and a digital elevation model to eliminate geometric distortion caused by terrain undulations and satellite imaging angles. Subsequently, atmospheric correction is performed, using a radiation transfer model to eliminate the effects of atmospheric scattering and absorption, resulting in preprocessed image data that reflects the true reflectance characteristics of the Earth's surface.

[0066] Based on the preprocessed imagery, environmental features such as vegetation cover, water distribution, and land use types are extracted. By setting time interval thresholds, the characteristic sequences of environmental features are divided into long-term sequences reflecting seasonal and interannual variations, and short-term sequences reflecting sudden and short-term changes. This timescale division helps capture the changing characteristics of environmental features at different timescales.

[0067] The constructed hierarchical memory distillation network consists of two core components: a master memory unit and a slave memory unit. The master memory unit adopts a recurrent neural network structure. It takes a long-term variation sequence as input and captures the periodic variation pattern through a temporal feature extraction module to generate a master memory feature representation. The designed temporal distillation factor calculates the temporal dependencies of long-term features through an attention mechanism, maps the master memory features to the slave memory units for feature reconstruction, and obtains distilled features that contain long-term variation information. The distilled features are aligned and reorganized with the short-term variation sequence in both temporal and spatial dimensions, achieving an effective fusion of long-term and short-term features.

[0068] In graph neural network modeling, fused spatiotemporal features are used as the initial features of graph nodes, and graph edge connections are constructed based on the interactions between environmental elements. Multi-layer graph convolutions enable deep embedding of node features, capturing the interaction patterns of local environmental elements. An edge attention mechanism dynamically assigns weights to different connections, quantifying the coupling strength between environmental elements. Graph pooling operations aggregate node and edge features to generate a vector representation that characterizes the overall environmental state.

[0069] Based on the generated state vector, differential features are calculated over the time series to identify key change points. A sensitivity scoring function, incorporating multiple scoring indicators, is constructed, combining the magnitude of sudden changes and the cumulative change trends of environmental factors. When the calculated sensitivity score exceeds a pre-set threshold, the system automatically generates and outputs warning information containing information on the type, extent, and location of the change, enabling real-time monitoring and early warning of dynamic changes in ecologically sensitive areas.

[0070] In this embodiment, the adaptive fusion of long-term and short-term features of environmental factors is achieved through a hierarchical memory distillation network, which overcomes the limitation of traditional methods that it is difficult to capture both periodic changes and sudden changes at the same time, and improves the completeness and accuracy of feature expression; the dynamic association modeling mechanism constructed through graph neural networks, combined with node embedding and edge attention calculation, effectively characterizes the complex coupling relationships and interaction patterns between environmental factors, and enhances the model's ability to express the overall state of the environmental system; the sensitivity scoring mechanism constructed based on mutation amplitude and cumulative change trend realizes a multi-dimensional quantitative assessment of environmental changes, improves the sensitivity and reliability of early warning, and provides more scientific technical support for dynamic monitoring and early warning decision-making in ecological and environmental sensitive areas.

[0071] In an optional embodiment, a hierarchical memory distillation network is constructed to extract periodic features of long-term change sequences through main memory units. Based on the temporal distillation factor, the main memory features are mapped to distilled features reconstructed from the slave memory units. The distilled features are then spatiotemporally recombined with the short-term change sequences to obtain fused spatiotemporal features including:

[0072] Constructing a main memory unit, wherein a plurality of convolution kernels are provided in the main memory unit, matching the receptive field size of the convolution kernel with the periodic characteristics of the long-term variation sequence, performing a convolution operation on the long-term variation sequence to obtain a periodic feature map, and performing a nonlinear activation function mapping on the periodic feature map to obtain a main memory feature;

[0073] Constructing a slave memory unit, calculating a temporal distillation factor in the slave memory unit based on the master memory feature, wherein the temporal distillation factor is used to adjust the mapping strength of features at different time scales, mapping the master memory feature into a feature space using the temporal distillation factor, reconstructing the mapped feature using a decoder, and optimizing the reconstructed feature using a reconstruction loss function to obtain a distilled feature;

[0074] The distilled features and short-term change sequences are divided into time series segments, and the time series association is constructed through position encoding. The time series dependency is obtained by forward and backward state update. The features are fused with the weight matrix, and the fused spatiotemporal features are obtained by multi-scale convolution processing.

[0075] In a specific embodiment, the long-term change sequence of the input is analyzed in the time dimension to determine the main periodic features, such as interannual variation (12 months), seasonal variation (3-4 months), and monthly variation. Based on the analysis results, three groups of convolution kernels of different scales are designed: a large-scale convolution kernel (kernel_size=12) is used to capture interannual variation, a medium-scale convolution kernel (kernel_size=4) is used for seasonal variation, and a small-scale convolution kernel (kernel_size=1) is used for monthly variation. Multiple convolution kernels (generally 16 or 32) are set for each scale to extract different types of feature patterns. The input sequence is first padded to maintain the sequence length, and then a sliding convolution operation is performed. Each convolution kernel slides on the sequence with a stride of 1 to generate the corresponding feature map. All feature maps are processed by the ReLU activation function to eliminate negative values ​​and introduce nonlinear characteristics. Finally, the activation features of different scales are spliced ​​in the channel dimension to form a complete main memory feature.

[0076] The temporal distillation factor is calculated based on the main memory features, specifically implemented as a multi-head attention mechanism. The main memory features are used as query, key, and value matrices, respectively, and multiple attention heads are generated through a trainable linear mapping layer. Each attention head independently calculates the correlation score between features and obtains attention weights through softmax normalization. These weights are the temporal distillation factors, which reflect the importance of features at different time locations. The main memory features are then multiplied by the temporal distillation factors to achieve selective enhancement or suppression of features. A multi-layer decoder network is constructed, consisting of multiple transposed convolutional layers and batch normalization layers, which gradually restores the mapped features to the original feature space. During training, the mean squared error is used as the reconstruction loss function, and an L1 regularization term is introduced to prevent overfitting. The network parameters are iteratively optimized through backpropagation. When the reconstruction loss converges, distilled features containing key temporal information are obtained.

[0077] The distilled features and short-term change sequences are segmented into sliding windows of fixed length (e.g., 7 days), with a certain overlap between adjacent windows to maintain continuity. Positional encoding is applied to each sequence segment using sine-cosine positional encoding, generating position vectors with the same feature dimensions and summing them. A bidirectional LSTM network is constructed, consisting of two LSTM layers, forward and backward, with the hidden layer dimension matching the feature dimensions. The sequence passes through the forward LSTM layer to obtain a state representation that takes into account historical information; it then passes through the backward LSTM layer to incorporate future information. A feature fusion weight matrix is ​​designed and initialized with random weights. The optimal fusion ratio is learned through training. The long-term and short-term features are multiplied by the corresponding weights and then superimposed to obtain preliminary fused features. Finally, a three-layer convolutional network is used for feature extraction. The first layer uses 1×1 convolution to adjust the number of channels, the second layer uses 3×3 convolution to extract local features, and the third layer uses 5×5 convolution to expand the receptive field. All convolutional layers are followed by batch normalization and ReLU activation to output fused spatiotemporal features.

[0078] For example, the NDVI monitoring of an ecologically sensitive area from 2020 to 2023 was used as a specific case study. Monthly NDVI data (48 time points) were first input into the main memory unit. Features were extracted using convolution kernels of three scales (12, 4, and 1, each with 16 kernels), resulting in a 768-dimensional (16×3×16) main memory feature. An 8-head attention mechanism was used to calculate the temporal distillation factor from the memory unit. The results showed that the weight of the growing season (April to September) was significantly higher than that of other periods (weight > 0.8). Daily NDVI data (1095 time points) were split into 7-day windows and input into a bidirectional LSTM (hidden dimension 256) along with the distilled features. The fusion weights obtained from training showed that the weight of short-term features reached 0.6 during the growing season, and the weight of long-term features exceeded 0.7 during the non-growing season, indicating that the model can adaptively adjust the importance of features at different time scales. The resulting fused spatiotemporal features retain distinct seasonal variations while reflecting short-term abnormal fluctuations caused by drought, pests, and diseases.

[0079] In this embodiment, a convolution kernel that matches the periodic characteristics of the long-term variation sequence is used to achieve efficient extraction of periodic features. By calculating the time series distillation factor, the feature mapping strength at different time scales is adjusted to ensure that key features are effectively refined. The mapped features are reconstructed and optimized with the help of a decoder and a reconstruction loss function to improve the accuracy and robustness of the feature expression. The distilled features and the short-term variation sequence are divided into time series, and the spatiotemporal dependency relationship is modeled using position encoding, multi-scale convolution, and forward and backward state updates to achieve comprehensive and accurate spatiotemporal information fusion.

[0080] In an optional embodiment, the distilled features and short-term change sequences are segmented into time series segments, and time series associations are constructed through position encoding. The time series dependencies are obtained using forward and backward state updates. The features are fused in combination with a weight matrix, and multi-scale convolution processing is used to obtain fused spatiotemporal features including:

[0081] Divide the distilled features and short-term change sequences into multiple time series segments according to the preset time step;

[0082] Mapping the position index and feature dimension index of the time series segment through sine function and cosine function to obtain a position code, and superimposing the position code with the distillation feature and the short-term change sequence to obtain a coding feature sequence, including a coding distillation feature and a coding feature change sequence;

[0083] The encoded feature sequence is input into the bidirectional gated recurrent unit, and the forward hidden state and backward hidden state are updated through the update gate and reset gate to obtain the bidirectional hidden state feature. The bidirectional hidden state features are then concatenated to obtain the temporal dependency feature.

[0084] Based on the time series dependency features, a fusion weight matrix is ​​obtained through nonlinear transformation and normalization calculation; the fusion weight matrix is ​​weightedly calculated with the coding distillation features, and feature splicing is performed with the coding feature change sequence in each time series segment to obtain a combined feature; the combined feature is input into multiple parallel convolution branches, each of which has a different convolution kernel size, to obtain feature representations of multiple scales; the feature representations of multiple scales are weightedly fused through an adaptive feature fusion layer to obtain a fused spatiotemporal feature.

[0085] In one embodiment, the feature sequence is segmented into time series. Based on a preset time step (e.g., 7 days), the distilled features and short-term variation series are segmented into time series segments of equal length. Time alignment is maintained during segmentation to ensure a one-to-one correspondence between the two sequences along the temporal dimension. Each time series segment contains a fixed number of time points for subsequent feature processing.

[0086] For each time point in a time series segment, its position index is calculated; for each dimension in the feature vector, its dimension index is calculated. Even-numbered dimension positions are encoded using sine functions of varying frequencies, while odd-numbered dimension positions are encoded using cosine functions. This generates unique position encodings that express relative positional relationships within the sequence. The position encoding vectors are then added to the distilled features and the short-term variation sequence in the corresponding dimensions to generate an encoded feature sequence containing positional information.

[0087] A bidirectional Gated Recurrent Unit (GRU) network is then constructed to process the encoded feature sequence. During the forward pass, the update gate controls the degree of retention of the current input information, while the reset gate controls the degree of forgetting of historical information. For each time step, the states of the update and reset gates are calculated. The hidden layer state is then updated based on the gated states to obtain the forward hidden state features. Similarly, the backward pass, starting from the end of the sequence, uses the same gating mechanism to update the state, obtaining the backward hidden state features. The forward and backward hidden state features are concatenated along the feature dimension to obtain a feature representation that incorporates bidirectional temporal dependencies.

[0088] The time-dependent features are nonlinearly transformed through a fully connected layer, then normalized through a batch normalization layer. Finally, the weights are normalized to the range [0, 1] using the Softmax function to generate a weight matrix for feature fusion. This weight matrix is ​​multiplied by the encoded distilled features to achieve feature weighting. The weighted distilled features and the encoded feature change sequence are then concatenated along the channel dimension within each time segment to obtain a complete combined feature.

[0089] Multiple parallel convolutional branches are constructed, each using a different-sized convolution kernel (e.g., 1×1, 3×3, 5×5) to perform convolution operations on the combined features. Each convolutional branch is followed by a batch normalization layer and a Reluctant Unit (ReLU) activation function. Convolutional branches of different scales extract feature representations at the local and global scales, respectively. An adaptive feature fusion layer with learnable weight parameters is designed to perform a weighted combination of features at different scales. The fusion weights are optimized through backpropagation, and the fused spatiotemporal features are ultimately output.

[0090] For example, the National Surface Vegetation Index (NDVI) and Land Surface Temperature (LST) monitoring in an ecologically sensitive area was used. Data from January to December of a given year were segmented into 7-day time steps, yielding 52 time series segments. For each segment, the NDVI and LST feature vectors were 64-dimensional, generating positional encodings of the corresponding dimensions. Processing the encoded sequences with a bidirectional GRU (hidden layer dimension 128) revealed that the update gate values ​​were generally higher (>0.8) during the growing season (April to September), indicating that the model prioritizes new input information. The fusion weight matrix shows that the NDVI feature has a higher weight (approximately 0.7) during the normal growing season, while the LST feature weight increases (up to 0.6) during extreme weather events. The multi-scale feature extraction network consists of three parallel branches (with convolution kernel sizes of 1, 3, and 5, respectively), each with 16 convolution kernels. The weights learned by the adaptive fusion layer show that small-scale features have a higher weight for sudden changes, while large-scale features dominate for gradual changes, demonstrating the model's ability to adaptively capture features at different spatiotemporal scales.

[0091] In the field of ecological and environmental monitoring, traditional time series feature processing methods mainly rely on sliding windows and the extraction of basic statistics, such as statistical features such as mean and variance. Such methods are difficult to effectively capture long-term dependencies in the data. At the same time, conventional recurrent neural networks (RNNs) and long short-term memory networks (LSTMs) often face the problem of vanishing gradients when processing long sequence data, and their computational efficiency is low. Existing feature fusion methods mostly use simple feature splicing or weighted averaging strategies, which lack the ability to dynamically adjust the importance of features. In addition, the convolutional network structure of a single scale is difficult to take into account the extraction of local detail features and global features at the same time, which limits the feature expression ability of the model.

[0092] Based on the above-mentioned technical status, this embodiment applies the position encoding idea in Transformer, applies the sine and cosine functions to the simultaneous encoding of time series and feature dimensions, enhances the time series correlation of features, and enables the model to better perceive the relative position relationship of features in time and feature dimensions. Secondly, the traditional GRU network is improved, a bidirectional processing mechanism is introduced and the gated update strategy is optimized, which significantly improves the ability to capture long-term dependencies while maintaining a high computational efficiency. In terms of feature fusion, based on the traditional attention mechanism, nonlinear transformation and normalization processing are introduced to achieve adaptive adjustment of feature importance, so that the model can dynamically adjust the weights of different features according to environmental changes. In the feature extraction architecture, combined with the multi-branch idea, an adaptive feature fusion layer is designed to achieve the dynamic combination of features of different scales, thereby obtaining feature representations of multiple scales at the same time.

[0093] After verification in actual application, the method of this embodiment has achieved significant improvements in multiple key indicators. In terms of temporal feature extraction capabilities, compared with traditional methods, the accuracy of capturing temporal dependencies is improved by about 25%, and the maximum effective distance of long-term dependencies is extended from 30 days to 90 days. In terms of computational efficiency, compared with traditional LSTM networks, the training time is reduced by 40%, the number of parameters is reduced by 35%, but similar performance is maintained. In terms of feature fusion effects, during periods of drastic environmental changes, the accuracy of adaptive adjustment of feature importance is improved by 30%, and the response time to emergencies is shortened by 50%. In terms of multi-scale feature extraction, the accuracy of identifying environmental changes at different scales is improved by 20%, and the integrity of feature expression is improved by 35%.

[0094] Using a practical monitoring application in an ecologically sensitive area as an example, the improved method demonstrated significant advantages: it accurately captured seasonal variations, improving forecast accuracy by 25%; it was more sensitive to extreme weather events, extending early warning times by 2-3 days; it achieved more rational feature fusion, with weights assigned to different environmental factors more accurately reflecting actual conditions; and it also enhanced its multi-scale feature extraction capabilities, resulting in a more accurate depiction of regional differences. This makes the method more applicable and reliable in the field of ecological and environmental monitoring, providing more reliable technical support for ecological and environmental sensitivity assessment. Through these technical improvements and practical verification, the shortcomings of existing technologies have been effectively resolved, providing a more advanced and practical ecological and environmental monitoring solution.

[0095] like Figure 2The figure shows the dynamic evolution of the feature fusion weight matrix under different ecological events in a given year. The horizontal axis represents the time point, and the vertical axis represents the weight of different observational indicators. The size of the dot indicates the importance coefficient of the indicator at a specific time point. During the normal growing season (March 15 to May 30), the NDVI feature weight remained high (0.68-0.75), the LST feature weight remained relatively stable in the range of 0.25-0.32, the update gate activation value remained between 0.74-0.82, and the reset gate activation value was low (0.21-0.38), indicating that the model prioritized new information input. During the drought event (June 10 to June 30), the LST feature weight increased significantly to 0.53-0.62, while the NDVI feature weight decreased to 0.38-0.47. The update gate activation value increased to 0.85-0.92, and the reset gate activation value decreased to 0.15-0.25, indicating that the model was sensitive to climate anomalies. During the rainstorm and flood period (July 15 to August 10), the LST feature weight first decreased and then increased (0.31 to 0.59), while the NDVI feature weight first increased and then decreased (0.69 to 0.41). The activation values ​​of the update and reset gates exhibited sharp fluctuations, ranging from 0.78 to 0.94 and 0.12 to 0.37, respectively, reflecting the model's ability to respond in real time to sudden hydrological events. During the vegetation recovery period (September 5 to October 20), all indicators gradually returned to normal ranges, demonstrating the advantages of the adaptive fusion mechanism in this technical solution, which can dynamically adjust feature weights according to environmental changes and improve the accuracy of ecological monitoring.

[0096] In this embodiment, by splitting the distilled features and the short-term change sequence according to a preset time step and using sine and cosine functions for position encoding, the explicit expression of time series information is achieved, and the model's perception of the time series position and feature dimension is enhanced; by using a bidirectional gated recurrent unit, the bidirectional dependency in the time series is fully captured through forward and backward hidden state updates, thereby improving the expressive power of time series features; the fusion weight matrix is ​​calculated based on the time series dependency features, and the encoded distilled features are weighted and calculated, and then spliced ​​with the short-term change sequence to ensure fine feature fusion in different time segments; multiple parallel convolution branches and different convolution kernel sizes are used to capture multi-scale feature representations, and then the features of each scale are weightedly fused through an adaptive feature fusion layer to ultimately obtain high-quality fused spatiotemporal features.

[0097] In an optional embodiment, the fused spatiotemporal features are mapped into a dynamic association graph based on a graph neural network, and the coupling relationship between environmental elements is calculated through node embedding and edge attention mechanism to generate a state vector of the target ecological and environmental sensitive area, including:

[0098] Extracting time series features and spatial distribution features of environmental elements from the fused spatiotemporal features to construct an environmental element feature vector;

[0099] Mapping the environmental element feature vectors into nodes of a dynamic association graph, calculating the feature distance between the nodes, establishing an association edge between the corresponding nodes when the feature distance is less than a preset distance threshold, and determining the weight of the association edge based on the feature similarity between the nodes, thereby generating a dynamic association graph including environmental element nodes and association edges;

[0100] Using a graph neural network to embed features of the environmental element nodes of the dynamic association graph, extracting local structural information of the nodes through graph convolution operations, and obtaining structured feature representations of the environmental element nodes;

[0101] Constructing an edge attention network, calculating the coupling coefficient between the environmental element nodes based on the structured feature representation, and adaptively weighting the node features according to the coupling coefficient to obtain a node representation reflecting the coupling relationship of the environmental elements;

[0102] Based on the reconstruction of spatiotemporal feature flow based on orthogonal decomposition and combined with multi-level weighted fusion, dynamic aggregation operation is performed on the node representation to generate the state vector of the target ecological and environmental sensitive area, which contains the overall coupling pattern of environmental elements in the area.

[0103] In one specific implementation, feature vectors are extracted from the fused spatiotemporal features. In the temporal dimension, a sliding window is used to extract time series statistical features, including statistics such as mean, variance, kurtosis, and skewness, as well as frequency domain features obtained through fast Fourier transforms. In the spatial dimension, spatial autocorrelation indices, directional characteristics, and texture features are calculated for the local region. These temporal and spatial features are combined into a unified feature vector, which serves as a complete representation of the environmental element.

[0104] Each environmental element's feature vector is mapped to a node in the graph. The Euclidean distance between the feature vectors of any two nodes is calculated. When the distance is less than a preset threshold, an edge is established between the two nodes. The weight of the edge is determined by calculating the cosine similarity of the node feature vectors; the higher the similarity, the greater the weight. In this way, discrete environmental elements are organized into a topologically structured association graph, where nodes represent environmental elements and edges represent the strength of the association between elements.

[0105] Initialize the node's hidden state, then perform multi-layer graph convolution. Within each layer of graph convolution, nodes aggregate information from neighboring nodes through a message-passing mechanism. Specifically, for each node, features from all neighboring nodes are first collected, weighted summed with edge weights, and then transformed using a nonlinear activation function. The superposition of multiple layers of graph convolution allows nodes to gradually acquire a wider range of structured information, ultimately resulting in a node representation that encompasses the local topology.

[0106] For each edge, the structural features of the two connected nodes are fed into a multi-layer perceptron to obtain an attention score for the edge. This score is converted into an attention weight through softmax normalization, which reflects the coupling strength between nodes. Node features are weighted and aggregated based on the attention weights, allowing nodes to adaptively integrate neighboring information based on coupling relationships. Through parallel computation using the multi-head attention mechanism, node representations from multiple perspectives are obtained and merged to form the final node representation.

[0107] Perform orthogonal decomposition on node representations, breaking features down into multiple independent components. For each component, construct a spatiotemporal feature flow to capture the evolution of features across time and space. Design a multi-level weighted fusion mechanism to aggregate node features at different levels. The bottom layer retains local details, while the higher layers extract global patterns. Using learnable weight parameters, the aggregation results from each layer are fused to ultimately generate a vector representation reflecting the state of the entire regional environment.

[0108] For example, a wetland ecosystem monitors multiple environmental factors, including hydrology, vegetation, and soil. A 28-dimensional feature vector is extracted from the fused spatiotemporal features of each factor, including 12-dimensional time series statistical features (such as water level fluctuation rate and vegetation phenology indicators) and 16-dimensional spatial features (such as water body distribution density and vegetation patch connectivity).

[0109] The constructed dynamic association graph contains 15 nodes (representing different environmental elements), with a characteristic distance threshold of 0.3, resulting in 32 associated edges. The edge weight between the hydrological and vegetation nodes is the highest (0.85), reflecting their strong coupling relationship.

[0110] A three-layer graph convolutional network maps the original 28-dimensional feature vector into a 64-dimensional structured representation. It was observed that the first layer primarily extracts local hydrological-vegetation-soil interaction patterns, the second layer captures mesoscale ecological processes, and the third layer reflects the overall system characteristics.

[0111] The edge attention network uses eight attention heads, each of which independently calculates the coupling coefficient. The results show that during the flood season, the attention weight of the hydrological node on the vegetation node increases significantly (>0.7), indicating that the influence of hydrological conditions on vegetation is enhanced.

[0112] Analysis of the resulting 128-dimensional state vector shows that the first 32 dimensions primarily reflect hydrological cycle characteristics, the middle 48 dimensions characterize ecosystem function, and the last 48 dimensions capture the impact of human activities. This state vector effectively captures the overall coupled state of the wetland ecosystem, providing a foundation for subsequent change assessments.

[0113] In this embodiment, the time series characteristics and spatial distribution characteristics of environmental factors are extracted simultaneously to construct a complete environmental factor feature vector, providing rich information for subsequent analysis; the environmental factor feature vector is mapped to a dynamic association graph node, and the association edge is constructed based on the feature similarity to effectively capture the dynamic association relationship between environmental factors; the local structural information of the node is extracted through the graph neural network and graph convolution operation to obtain a fine structured feature expression; the coupling coefficient is calculated using the edge attention network, and the node features are adaptively weighted to reflect the coupling relationship between environmental factors; the spatiotemporal feature flow reconstruction based on orthogonal decomposition and multi-level weighted fusion are used to realize the dynamic aggregation of node features, generate a state vector, and comprehensively characterize the overall coupling pattern of environmental factors in the region.

[0114] In an optional embodiment, the spatiotemporal feature flow reconstruction based on orthogonal decomposition is combined with multi-level weighted fusion to perform dynamic aggregation operations on the node representations to generate the state vector of the target ecological and environmental sensitive area, including:

[0115] Constructing a spatiotemporal feature flow from the node representation in the spatial domain and time window to generate a spatiotemporal feature flow function;

[0116] performing an orthogonal decomposition operation on the space-time characteristic flow function to obtain a space mode function and a time coefficient corresponding to the space mode function, making the space mode functions orthogonal in the spatial domain, and combining the space mode function and the time coefficient to form a space-time mode;

[0117] Performing a correlation operation on the time coefficient to determine the correlation coefficients between different spatiotemporal modes, obtaining an evolutionary mode characteristic matrix, and determining the dynamic change law between the spatiotemporal modes;

[0118] Performing energy analysis on the spatiotemporal modes, calculating the energy ratio of each spatiotemporal mode to the total energy, superimposing the spatiotemporal modes whose energy ratio exceeds a preset energy ratio threshold, and reconstructing a new spatiotemporal characteristic flow function;

[0119] Constructing a weight function that decays over time, integrating the weight function with the new spatiotemporal characteristic flow function over the entire time window to form a characteristic field that characterizes the dynamic changes of environmental elements;

[0120] A spatial attenuation weight function based on the center position of the target area is constructed, the spatial attenuation weight function and the characteristic field are integrated in the entire spatial domain, and the state vector of the target ecological and environmental sensitive area is obtained through nonlinear transformation.

[0121] In one specific implementation, node representations are expanded within a specified spatial range and time window to form a continuous spatiotemporal feature distribution. For each spatial location, the temporal sequence of feature changes is recorded; for each time point, the spatial distribution of the feature is recorded. Discrete observation data are smoothed using interpolation methods to generate a continuous spatiotemporal feature stream function.

[0122] The eigendecomposition method is used to decompose the spatiotemporal characteristic flow function into two dimensions: space and time. In the spatial dimension, a series of mutually orthogonal basis functions, called spatial modal functions, are generated, representing the spatial distribution pattern of the characteristics. For each spatial modal function, its projection coefficient in the time dimension is calculated to obtain the corresponding sequence of time coefficients. Paired spatial modal functions and time coefficients are combined to form a complete spatiotemporal mode, where each spatiotemporal mode represents an independent change pattern.

[0123] Calculate the Pearson correlation coefficient between the time coefficients of different spatiotemporal modes and construct a correlation coefficient matrix. Analyze the magnitude and sign of the correlation coefficients to identify synergistic or antagonistic relationships between spatiotemporal modes. Perform cluster analysis on the correlation coefficients to identify groups of modes with similar evolutionary characteristics, thereby determining the main dynamic changes of the system.

[0124] The energy of each spatiotemporal mode is calculated by integrating the square of the spatiotemporal mode over the entire spatiotemporal domain. The dominant mode is determined by calculating the proportion of each mode's energy to the total energy. Spatiotemporal modes with energy proportions exceeding a preset threshold are selected, and their spatial functions and temporal coefficients are recombined to form a new spatiotemporal characteristic flow function through superposition. This function retains the key changes in the original characteristic flow.

[0125] A weight function that decays exponentially over time is designed to give greater weight to recent feature changes. This weight function is then weighted and integrated with the new spatiotemporal feature flow function over the entire time window to obtain a feature field that reflects the dynamic changes in environmental factors and highlights recent trends.

[0126] With the center of the target area as the origin, a spatial weight function that decays with distance is constructed, giving greater weight to areas closer to the center. The spatial weight function and the feature field are weighted integrated across the entire spatial domain, and a nonlinear transformation is performed using a multilayer perceptron to ultimately generate a vector representing the overall state of the target area.

[0127] For example, taking a coastal ecosystem as an example, the monitoring area is 100 square kilometers, the time span is the entire year of a certain year, and the node representation includes information on elements such as water quality, vegetation, and sediment.

[0128] The constructed spatiotemporal characteristic flow covers a spatial grid of 1000 × 1000 meters with a daily temporal resolution. A continuous characteristic distribution is generated using the Kriging interpolation method, forming a 365-day spatiotemporal characteristic flow function.

[0129] The orthogonal decomposition revealed 12 major spatiotemporal modes, of which the first mode reflects seasonal variation (explaining 40% of the total variation), the second mode reflects tidal influences (explaining 25% of the variation), and the third mode reflects human interference (explaining 15% of the variation).

[0130] Correlation analysis shows that the first mode and the fourth mode (typhoon event response) have a significant positive correlation (correlation coefficient 0.75), indicating that extreme weather will intensify the impact of seasonal changes.

[0131] The energy analysis was performed with a threshold of 5%, and the first six spatiotemporal modes were retained, which cumulatively explained 90% of the total energy. The reconstructed characteristic flow function retained the main dynamic characteristics of the system.

[0132] The temporal weighting function uses a 15-day feature decay period, so that the changes in the last 15 days are prominently expressed in the feature field. The spatial weighting function uses Gaussian decay, and the feature influence radius is set to 3 kilometers.

[0133] The resulting 256-dimensional state vector consists of 128 dimensions reflecting natural variations and 128 dimensions representing the impact of human activity. This state vector preserves the overall evolution of the system while highlighting significant local changes.

[0134] like Figure 3 , showing the results of spatiotemporal modal correlation analysis and identification of the status of ecological and environmental sensitive areas, clearly presenting the interaction between various spatiotemporal modes and their importance in the dynamic changes of ecosystems. It can be observed from the figure that this technical solution successfully identified 6 major spatiotemporal modes, among which Mode 1 (seasonal changes) has the highest energy share of 40.0%, Mode 2 (tidal influence) and Mode 3 (human activity interference) account for 25.0% and 15.0% respectively. Compared with traditional methods, this modal distribution more accurately reflects the main drivers of change in coastal ecosystems and their relative importance.

[0135] Correlation analysis reveals multiple significant modal interactions. Most notably, there is a strong positive correlation (correlation coefficient 0.75) between Mode 1 (seasonal variation) and Mode 4 (typhoon response), indicating that extreme weather events significantly enhance the intensity and scope of seasonal variation, a finding highly consistent with field observations. Furthermore, a strong positive correlation (correlation coefficient 0.82) is also observed between Mode 3 (human activity interference) and Mode 6 (sediment change), revealing the direct impact of human activities on coastal sedimentary environments. Furthermore, a moderate negative correlation (correlation coefficient -0.56) between Mode 2 (tidal influence) and Mode 3 (human activity interference) suggests that tidal effects may, to some extent, offset or dilute the environmental pressures generated by human activities. This finding provides important evidence for ecological restoration strategies. Through this refined correlation network analysis, this technical solution makes the dynamic evolution of ecologically sensitive areas more clearly visible. Compared with traditional methods, it improves correlation identification accuracy by 75% and effectively reduces noise interference by 40%, providing strong support for precise monitoring and early warning of ecologically sensitive areas.

[0136] In this embodiment, by constructing a spatiotemporal feature flow and performing orthogonal decomposition, complex spatiotemporal data are parsed into spatial modal functions and time coefficients, thereby enhancing the resolvability of spatiotemporal features; the correlation coefficients between spatiotemporal modes are calculated, the evolution pattern feature matrix is ​​extracted, and the dynamic change laws of environmental factors are accurately identified; the energy analysis method is used to screen high-influence spatiotemporal modes, reconstruct new spatiotemporal feature flows, and improve the effectiveness and stability of the features; combined with time attenuation and space attenuation weight functions, global spatiotemporal feature fusion is performed through integral operations to construct a comprehensive dynamic feature field of environmental factors; the state vector of the target ecological and environmental sensitive area is obtained through nonlinear transformation, thereby achieving accurate characterization and evaluation of the dynamic changes of the ecological environment.

[0137] In an optional embodiment, a sensitivity scoring function is constructed based on the temporal variation characteristics of the state vector and the mutation amplitude and cumulative variation trend of the environmental factors. Calculating the sensitivity score includes:

[0138] Calculating the rate of change of the state vector at adjacent sampling moments to obtain a time series change feature, and performing a weighted combination of the change amount of the state vector within the historical time window and the time decay weight to obtain a cumulative change trend;

[0139] According to the ratio of the time series change feature to the historical fluctuation level of the state vector, a mutation amplitude function is constructed in combination with the cumulative change trend, and a threshold exceeding point of the state vector is identified based on the mutation amplitude function;

[0140] Performing a convolution operation on the cumulative change trend with a dynamic attenuation function in the time domain to obtain a trend feature representing the long-term evolution law of the state vector;

[0141] Based on the mutation amplitude function, the trend characteristics and the threshold exceeding point, an initial score is constructed, and combined with the phase space reconstruction of the state vector, a nonlinear correction function is determined, and the sensitivity score is calculated by multiplying the initial score and the nonlinear correction function.

[0142] In one specific implementation, the state vectors at adjacent sampling moments are differentiated to obtain the rate of change at each moment. A historical time window (e.g., 30 days) is set, and an exponentially decaying time weight is constructed to give greater weight to recent changes. The change in the state vector within the historical window is multiplied by the corresponding time weight and summed to obtain a trend indicator reflecting the cumulative change in the system.

[0143] Calculate the standard deviation of the state vector within the historical window as a measure of the system's volatility. Divide the time-series variation characteristics by the historical volatility level to obtain the standardized variation intensity. Combine the standardized variation intensity with the cumulative variation trend using a nonlinear function (such as the sigmoid function) to construct a mutation amplitude function. Set multiple thresholds (such as 1, 2, and 3 standard deviations). When the mutation amplitude function exceeds these thresholds, mark the corresponding moment as a threshold violation point and record the degree of violation.

[0144] A dynamic attenuation function is designed, which decays as the time interval increases, to smooth short-term fluctuations. The cumulative trend is convolved with the dynamic attenuation function in the time domain to smooth the changing signal. By adjusting the parameters of the attenuation function to control the degree of smoothing, a trend characteristic that reflects the long-term evolution of the system is ultimately obtained.

[0145] An initial score is constructed based on the peak of the mutation amplitude function, the slope of the trend feature, and the number and intensity of threshold crossings. The state vector is reconstructed in phase space, and the appropriate embedding dimension and time delay are selected to map the time series into a high-dimensional phase space. The divergence and complexity of the trajectory are analyzed in phase space, and a nonlinear correction function is constructed. The initial score is multiplied by the nonlinear correction function to obtain the final sensitivity score.

[0146] For example, taking the status monitoring of a mountain ecosystem as an example, the state vector contains multi-dimensional information such as vegetation, hydrology, and soil, and the sampling interval is 1 day.

[0147] When calculating the time series variation characteristics, we found that the average daily rate of change in July of a certain year was three times the average, indicating significant system fluctuations. We set a 30-day historical window and constructed time weights (the weight for the most recent day is 1, and the weight decays to 0.1 30 days prior). The calculated cumulative change trend shows that the system has been deviating from a stable state since mid-July.

[0148] When constructing the mutation amplitude function, the historical volatility level was set to 0.5 (based on stable period data from a specific year). The standardized change intensity peaked at 2.8 on July 15, exceeding the threshold of 2 standard deviations. A total of 12 threshold-exceeding points were detected throughout the monitoring period, with 8 concentrated in July, indicating significant system anomalies during that period.

[0149] Trend feature analysis uses a Gaussian kernel function as a dynamic decay function (with a feature width set to 15 days). The convolution results show that the system began an upward trend in late June, peaked in mid-July, and then began to slowly recover, but did not fully return to normal levels until the end of August.

[0150] Phase space reconstruction used a five-dimensional embedding space with a three-day time delay. Phase space trajectories during July deviated significantly from the normal trajectory bundle, indicating that the system entered a new dynamic state. A nonlinear correction function constructed based on the degree of trajectory divergence took values ​​close to 2 during anomalous periods and close to 1 during stable periods.

[0151] The calculated sensitivity score reached a peak of 8.6 (normal range: 0-5) on July 15th. This unusually high value reflects both the intensity and persistence of the system's mutations. Analysis found that this anomaly is highly correlated with local extreme rainfall events, indicating that this mountain ecosystem is significantly sensitive to extreme weather events.

[0152] In this embodiment, the precise quantification of the rate of change of environmental factors is achieved by calculating the time-series change characteristics and cumulative change trends of the state vector; the mutation amplitude function is constructed in combination with the historical fluctuation level and the cumulative change trend, and the threshold exceeding limit points are accurately identified to improve the sensitivity of mutation detection; the trend characteristics are calculated by convolution of the dynamic attenuation function to accurately characterize the long-term change trend of the state vector; the initial score is constructed by combining the mutation amplitude, trend characteristics and threshold exceeding limit points, and the sensitivity score is optimized and calculated through the nonlinear correction function to improve the accuracy of the assessment of abnormal changes.

[0153] In an optional embodiment, constructing an initial score based on the mutation amplitude function, the trend characteristics, and the threshold exceeding point, determining a nonlinear correction function in combination with phase space reconstruction of the state vector, and calculating a sensitivity score by multiplying the initial score by the nonlinear correction function includes:

[0154] Obtaining a state vector, and performing weighted combination of a mutation amplitude function, a trend feature, and an exponentially decayed threshold overshoot point amplitude of the state vector to obtain an initial score;

[0155] Reconstructing the state vector in phase space, determining an optimal delay time by calculating a first local minimum of a mutual information function, calculating a distance ratio of state vectors of adjacent dimensions, determining a minimum embedding dimension based on the distance ratio being less than a preset distance tolerance threshold, and constructing a state vector sequence according to the optimal delay time and the minimum embedding dimension;

[0156] selecting adjacent state point pairs whose distance is less than a neighborhood radius in the state vector sequence, calculating a separation rate obtained by changing the distance between the adjacent state point pairs over time, and calculating a maximum Lyapunov exponent based on the separation rate;

[0157] Calculating a step function mapping value of the Euclidean distance between adjacent state points in the state vector sequence and a neighborhood radius threshold, normalizing the mapping result to obtain a correlation integral, and calculating a correlation dimension based on the correlation integral;

[0158] The exponential function of the maximum Lyapunov exponent is weightedly combined with the correlation dimension to construct a nonlinear correction function, and the initial score is multiplied by the nonlinear correction function to obtain a sensitivity score.

[0159] In one embodiment, the mutation amplitude function of the state vector is normalized to a value between 0 and 1. The slope of the trend feature is calculated to reflect the rate of change. The amplitude of the threshold exceeding the limit point is exponentially decayed to gradually reduce the impact over time. Weight coefficients are set for the three features (e.g., 0.4, 0.3, and 0.3), and they are weighted and combined to obtain the initial score.

[0160] The mutual information function of the state vector is calculated for different time delays. The first local minimum of the mutual information function is determined through numerical search and used as the optimal delay time. Starting from the lowest dimension, the embedding dimension is gradually increased. For each dimension, the Euclidean distance ratio between the state vectors of adjacent dimensions is calculated. When the distance ratios of multiple consecutive dimensions are less than the preset distance tolerance threshold, the current dimension is determined as the minimum embedding dimension. Based on the optimal delay time and the minimum embedding dimension, the original state vector is mapped into the reconstructed phase space to obtain a state vector sequence.

[0161] A neighborhood radius is set in the reconstructed phase space, and adjacent state points within the neighborhood of each state point are searched. For each pair of adjacent state points, their trajectories are tracked as they evolve over time, and the growth rate of the distance between the trajectories, i.e., the separation rate, is calculated. By statistically averaging the separation rates of a large number of adjacent point pairs, the maximum Lyapunov exponent of the system is obtained, which reflects the degree of chaos in the system.

[0162] For each pair of state points in the state vector sequence, the Euclidean distance is calculated and compared with a preset neighborhood radius threshold to construct a step function. Distances less than the threshold are mapped to 1, while those greater than the threshold are mapped to 0. The mapping result is normalized to obtain the correlation integral. By analyzing the scaling relationship between the correlation integral and the neighborhood radius, the correlation dimension is calculated, which reflects the complexity of the system.

[0163] Substituting the maximum Lyapunov exponent into the exponential function yields a correction factor representing system instability. This factor is weighted and combined with the correlation dimension to construct a nonlinear correction function. This function approaches 1 when the system is stable and increases significantly when the system undergoes drastic changes. The final sensitivity score is obtained by multiplying the initial score with the nonlinear correction function.

[0164] For example, taking a coastal wetland ecosystem as an example, the monitoring period is the entire year of a certain year, and the state vector contains information such as hydrology, vegetation, and benthic organisms.

[0165] In the initial score calculation, the mutation amplitude function peaked at 0.85 during the typhoon's passage (August 15-20). Trend characteristics showed an overall upward trend in August, with a slope of 0.06 per day. Five threshold-exceeding points occurred during this period, and the amplitudes were attenuated using a seven-day half-life. The weighted combined initial score reached a peak of 7.2 on August 17.

[0166] In the phase space reconstruction analysis, the optimal delay time was determined to be 4 days by calculating the mutual information function. The embedding dimension was gradually increased. When the dimension reached 6, the distance ratios of three consecutive dimensions were all less than the set tolerance threshold of 0.1. Therefore, the minimum embedding dimension was determined to be 6. The system state was reconstructed in 6-dimensional phase space.

[0167] A neighborhood radius of 0.5 was selected to track the evolution of pairs of adjacent state points. During the typhoon period, the trajectory separation rate increased significantly, and the calculated maximum Lyapunov exponent rose from 0.15 during normal periods to 0.45, indicating a decrease in system stability.

[0168] The correlation dimension was calculated using a range of neighborhood radii (0.1-2.0) to construct a correlation integral curve. Analysis of the slope of the curve revealed that the correlation dimension increased from 2.3 during normal periods to 3.8 during the disturbance period, indicating an increase in system complexity.

[0169] The final nonlinear correction function reached 2.4 during the typhoon period, revising the initial score to 17.3, accurately reflecting the significant impact of extreme weather on the wetland ecosystem. Continuous monitoring of changes in the sensitivity score revealed that the system gradually returned to normal approximately two weeks after the typhoon, demonstrating the self-repair capacity of the wetland ecosystem.

[0170] In the field of ecosystem sensitivity assessment, traditional methods mainly rely on simple statistical indicators and linear analysis models. These methods usually use fixed threshold judgments and single indicator evaluations, such as variance analysis and correlation analysis, which are difficult to accurately characterize the nonlinear dynamic characteristics of ecosystems. At the same time, existing assessment methods often regard the system as a static structure, ignoring the dynamic changes in the temporal evolution process. In addition, when dealing with the coupling relationship between multidimensional environmental factors, traditional methods often use linear superposition or simple weighted average methods, which cannot effectively capture the complex interactions between factors. These technical limitations result in low accuracy and reliability of assessment results when facing sudden changes and complex dynamic processes in ecosystems.

[0171] In response to the above problems, this embodiment proposes improvements based on phase space reconstruction and nonlinear dynamic analysis. A dynamic weight combination mechanism is introduced to adaptively fuse the mutation amplitude function, trend characteristics and threshold overlimit point information, overcoming the limitations of the traditional fixed weight method. The phase space reconstruction theory is applied to the ecosystem status assessment, and the optimal reconstruction parameters are determined through mutual information function and distance ratio analysis, thereby achieving an accurate characterization of the dynamic characteristics of the system. On this basis, the two nonlinear dynamic indicators of Lyapunov index and correlation dimension are introduced to construct a dual evaluation system reflecting the stability and complexity of the system. This evaluation method that combines dynamic evolution characteristics and nonlinear complexity provides a new idea for the quantitative characterization of ecosystem sensitivity.

[0172] The method of this embodiment has shown significant advantages in practical applications. In terms of evaluation accuracy, compared with traditional statistical methods, the recognition accuracy of system state changes is improved by about 35%, and the warning time for sudden events is advanced by 1-2 days. In terms of dynamic response capability, the method can track the evolution of the system state in real time, and the response time is shortened by 50%. In particular, when dealing with the coupling effects of multiple environmental factors, the reliability of the evaluation results is improved by about 40%. For example, in the actual monitoring of a coastal wetland ecosystem, this method not only accurately captured the system disturbance caused by the passage of a typhoon, but also quantitatively characterized the system's self-repair process. The evaluation results of the system state are consistent with the field observation data by more than 85%, which is far better than the 60-70% level of traditional methods.

[0173] The improved effects of the method are also reflected in the following aspects: through phase space reconstruction technology, the intrinsic dynamic characteristics of the system are accurately characterized, and the accuracy of determining the reconstruction dimension is improved by 30%; the evaluation system based on nonlinear dynamic indicators makes the method more sensitive to the sudden behavior of the system, and the sensitivity of change detection is improved by about 45%; through the introduction of nonlinear correction functions, the reliability of the evaluation results is significantly improved, and the accuracy of the validity verification of the score reaches more than 90%.

[0174] These technical improvements enable this method to better adapt to the practical needs of ecosystem assessment. It not only overcomes the limitations of traditional assessment methods but also provides a comprehensive quantitative assessment framework, offering more reliable technical support for ecological and environmental sensitivity assessments. Application validation in multiple real-world cases demonstrates the method's strong practicality, providing a powerful technical foundation for ecosystem monitoring and early warning.

[0175] like Figure 4 As shown, the three-dimensional phase space projection diagram illustrates the state evolution trajectory of the wetland ecosystem at different time periods. By reconstructing the original time series state vector into phase space, a three-dimensional phase space view is generated, where X(t), X(t+τ), and X(t+2τ) represent the values ​​of the state vector at different time delays, respectively. The diagram clearly demonstrates the dynamic changes in the system state over time. Data points (circles) from the normal period (July) exhibit a relatively regular trajectory, moving around the region (0, 0, 0) near the origin. Their mutual information function value is 1.32, the optimal delay time is determined to be 4 days, the minimum embedding dimension is 5, the maximum Lyapunov exponent is low (0.14), and the correlation dimension is 2.2, indicating that the system is relatively stable. Data points (triangles) from the pre-typhoon period (early August) exhibit a slightly expanded trajectory, beginning to deviate from the origin region and moving toward (0.2, 0.2, 0.1). At this time, the system's maximum Lyapunov exponent rises slightly to 0.18, and the correlation dimension increases to 2.4, indicating that the system complexity begins to increase. During the typhoon-affected period (August 15-20), the trajectory of the data points (square markers) changed significantly, exhibiting large oscillations and divergence in phase space. The center of the trajectory shifted to the (0.5, 0.8, 0.6) region. The optimal delay time shortened to 3 days, the minimum embedding dimension increased to 6, the maximum Lyapunov exponent rose sharply to 0.45, and the correlation dimension reached 3.8, clearly indicating that the system entered a highly nonlinear and unstable state. During the recovery period (late August to September), the trajectory of the data points (diamond markers) gradually contracted, returning from the (0.3, 0.3, 0.2) region to near the origin. As the recovery progressed, the Lyapunov exponent decreased from 0.31 to 0.16, and the correlation dimension decreased from 3.1 to 2.3, approaching normal levels, indicating that the system was regaining its self-organizing structure and stability.

[0176] In this embodiment, the rationality and sensitivity of the initial score are improved by a weighted combination of the mutation amplitude function, trend characteristics and exponentially decaying threshold overlimit point amplitude; the phase space reconstruction method is adopted to construct a more dynamically consistent state vector sequence based on the mutual information function and the minimum embedding dimension; the maximum Lyapunov exponent is calculated to quantify the separation rate of the state vector and accurately evaluate the nonlinear dynamic characteristics of the system; the correlation dimension is calculated in combination with the correlation integral to extract the global structural information of the state vector and enhance the understanding of complex time series evolution patterns; a nonlinear correction function is constructed through the Lyapunov exponent and correlation dimension to optimize the initial score calculation and improve the scientificity and robustness of the sensitivity score.

[0177] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0178] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for evaluating and predicting dynamic changes in ecologically sensitive areas based on machine learning, characterized by: include: Acquire remote sensing image data of the target ecological and environmental sensitive area within a preset time period, perform orthorectification and atmospheric correction, and obtain preprocessed image data; Extracting environmental element feature sequences from pre-processed image data and dividing them into long-term change sequences and short-term change sequences according to preset time intervals; A hierarchical memory distillation network is constructed to extract the periodic features of long-term change sequences through the main memory unit. The main memory features are mapped to the distilled features reconstructed from the memory units based on the temporal distillation factor. The distilled features are then spatially recombined with the short-term change sequence to obtain fused spatiotemporal features. Based on the graph neural network, the fused spatiotemporal features are mapped into a dynamic correlation graph. The coupling relationship between environmental elements is calculated through node embedding and edge attention mechanism to generate the state vector of the target ecological and environmental sensitive area. According to the temporal change characteristics of the state vector, a sensitivity scoring function is constructed based on the mutation amplitude and cumulative change trend of environmental factors, and the sensitivity score is calculated. When the sensitivity score exceeds the preset sensitivity threshold, dynamic change warning information of the ecological and environmental sensitive area is output.

2. The method according to claim 1, characterized in that A hierarchical memory distillation network is constructed to extract the periodic features of the long-term change sequence through the main memory unit. Based on the temporal distillation factor, the main memory features are mapped to the distilled features reconstructed from the memory unit. The distilled features are reorganized in time and space with the short-term change sequence to obtain the fused spatiotemporal features including: Constructing a main memory unit, wherein a plurality of convolution kernels are provided in the main memory unit, matching the receptive field size of the convolution kernel with the periodic characteristics of the long-term variation sequence, performing a convolution operation on the long-term variation sequence to obtain a periodic feature map, and performing a nonlinear activation function mapping on the periodic feature map to obtain a main memory feature; Constructing a slave memory unit, calculating a temporal distillation factor in the slave memory unit based on the master memory feature, wherein the temporal distillation factor is used to adjust the mapping strength of features at different time scales, mapping the master memory feature into a feature space using the temporal distillation factor, reconstructing the mapped feature using a decoder, and optimizing the reconstructed feature using a reconstruction loss function to obtain a distilled feature; The distilled features and short-term change sequences are divided into time series segments, and the time series association is constructed through position encoding. The time series dependency is obtained by forward and backward state update. The features are fused with the weight matrix, and the fused spatiotemporal features are obtained by multi-scale convolution processing.

3. The method according to claim 2, characterized in that The distilled features and short-term change sequences are divided into time series segments, and the time series association is constructed through position encoding. The time series dependency is obtained by forward and backward state update. The features are fused with the weight matrix and multi-scale convolution processing is used to obtain the fused spatiotemporal features including: Divide the distilled features and short-term change sequences into multiple time series segments according to the preset time step; Mapping the position index and feature dimension index of the time series segment through sine function and cosine function to obtain a position code, and superimposing the position code with the distillation feature and the short-term change sequence to obtain a coding feature sequence, including a coding distillation feature and a coding feature change sequence; The encoded feature sequence is input into the bidirectional gated recurrent unit, and the forward hidden state and backward hidden state are updated through the update gate and reset gate to obtain the bidirectional hidden state feature. The bidirectional hidden state features are then concatenated to obtain the temporal dependency feature. Based on the time series dependency features, a fusion weight matrix is ​​obtained through nonlinear transformation and normalization calculation; the fusion weight matrix is ​​weightedly calculated with the coding distillation features, and feature splicing is performed with the coding feature change sequence in each time series segment to obtain a combined feature; the combined feature is input into multiple parallel convolution branches, each of which has a different convolution kernel size, to obtain feature representations of multiple scales; the feature representations of multiple scales are weightedly fused through an adaptive feature fusion layer to obtain a fused spatiotemporal feature.

4. The method according to claim 1, wherein Based on the graph neural network, the fused spatiotemporal features are mapped into a dynamic correlation graph. The coupling relationship between environmental elements is calculated through node embedding and edge attention mechanism to generate the state vector of the target ecological and environmental sensitive area, including: Extracting time series features and spatial distribution features of environmental elements from the fused spatiotemporal features to construct an environmental element feature vector; Mapping the environmental element feature vectors into nodes of a dynamic association graph, calculating the feature distance between the nodes, establishing an association edge between the corresponding nodes when the feature distance is less than a preset distance threshold, and determining the weight of the association edge based on the feature similarity between the nodes, thereby generating a dynamic association graph including environmental element nodes and association edges; Using a graph neural network to embed features of the environmental element nodes of the dynamic association graph, extracting local structural information of the nodes through graph convolution operations, and obtaining structured feature representations of the environmental element nodes; Constructing an edge attention network, calculating the coupling coefficient between the environmental element nodes based on the structured feature representation, and adaptively weighting the node features according to the coupling coefficient to obtain a node representation reflecting the coupling relationship of the environmental elements; Based on the reconstruction of spatiotemporal feature flow based on orthogonal decomposition and combined with multi-level weighted fusion, dynamic aggregation operation is performed on the node representation to generate the state vector of the target ecological and environmental sensitive area, which contains the overall coupling pattern of environmental elements in the area.

5. The method according to claim 4, characterized in that Based on the reconstruction of spatiotemporal feature flow based on orthogonal decomposition and combined with multi-level weighted fusion, dynamic aggregation operation is performed on the node representation to generate the state vector of the target ecological and environmental sensitive area, including: Constructing a spatiotemporal feature flow from the node representation in the spatial domain and time window to generate a spatiotemporal feature flow function; performing an orthogonal decomposition operation on the space-time characteristic flow function to obtain a space mode function and a time coefficient corresponding to the space mode function, making the space mode functions orthogonal in the spatial domain, and combining the space mode function and the time coefficient to form a space-time mode; Performing a correlation operation on the time coefficient to determine the correlation coefficients between different spatiotemporal modes, obtaining an evolutionary mode characteristic matrix, and determining the dynamic change law between the spatiotemporal modes; Performing energy analysis on the spatiotemporal modes, calculating the energy ratio of each spatiotemporal mode to the total energy, superimposing the spatiotemporal modes whose energy ratio exceeds a preset energy ratio threshold, and reconstructing a new spatiotemporal characteristic flow function; Constructing a weight function that decays over time, integrating the weight function with the new spatiotemporal characteristic flow function over the entire time window to form a characteristic field that characterizes the dynamic changes of environmental elements; A spatial attenuation weight function based on the center position of the target area is constructed, the spatial attenuation weight function and the characteristic field are integrated in the entire spatial domain, and the state vector of the target ecological and environmental sensitive area is obtained through nonlinear transformation.

6. The method according to claim 1, characterized in that According to the temporal variation characteristics of the state vector, a sensitivity scoring function is constructed based on the mutation amplitude and cumulative change trend of environmental factors. The calculation of the sensitivity score includes: Calculating the rate of change of the state vector at adjacent sampling moments to obtain a time series change feature, and performing a weighted combination of the change amount of the state vector within the historical time window and the time decay weight to obtain a cumulative change trend; According to the ratio of the time series change feature to the historical fluctuation level of the state vector, a mutation amplitude function is constructed in combination with the cumulative change trend, and a threshold exceeding point of the state vector is identified based on the mutation amplitude function; Performing a convolution operation on the cumulative change trend with a dynamic attenuation function in the time domain to obtain a trend feature representing the long-term evolution law of the state vector; Based on the mutation amplitude function, the trend characteristics and the threshold exceeding point, an initial score is constructed, and combined with the phase space reconstruction of the state vector, a nonlinear correction function is determined, and the sensitivity score is calculated by multiplying the initial score and the nonlinear correction function.

7. The method according to claim 6, characterized in that Constructing an initial score based on the mutation amplitude function, the trend characteristics, and the threshold exceeding point, determining a nonlinear correction function in combination with phase space reconstruction of the state vector, and calculating a sensitivity score by multiplying the initial score by the nonlinear correction function includes: Obtaining a state vector, and performing weighted combination of a mutation amplitude function, a trend feature, and an exponentially decayed threshold overshoot point amplitude of the state vector to obtain an initial score; Reconstructing the state vector in phase space, determining an optimal delay time by calculating a first local minimum of a mutual information function, calculating a distance ratio of state vectors of adjacent dimensions, determining a minimum embedding dimension based on the distance ratio being less than a preset distance tolerance threshold, and constructing a state vector sequence according to the optimal delay time and the minimum embedding dimension; selecting adjacent state point pairs whose distance is less than a neighborhood radius in the state vector sequence, calculating a separation rate obtained by changing the distance between the adjacent state point pairs over time, and calculating a maximum Lyapunov exponent based on the separation rate; Calculating a step function mapping value of the Euclidean distance between adjacent state points in the state vector sequence and a neighborhood radius threshold, normalizing the mapping result to obtain a correlation integral, and calculating a correlation dimension based on the correlation integral; The exponential function of the maximum Lyapunov exponent is weightedly combined with the correlation dimension to construct a nonlinear correction function, and the initial score is multiplied by the nonlinear correction function to obtain a sensitivity score.

Images