A vegetation coverage inversion method based on an ecological hydrological constraint graph convolution network

Abstract

The present application provides a kind of vegetation coverage inversion method based on ecological hydrology constraint graph convolution network, comprising: obtaining multispectral remote sensing image and preprocessing to obtain effective pixel, the spectral characteristics including water index are calculated to construct multidimensional feature set;Water is extracted based on water index, and water influence factor is generated using distance attenuation function;With effective pixel as node, edge is built based on spatial adjacency relationship, and ecological weighted graph structure is formed by fusing spatial weight and difference determined ecological similarity weight;Multidimensional features and water influence factor are fused and input into trained model to obtain vegetation coverage simulation value;The vegetation coverage of the region to be studied is inverted using the trained model, and the spatial distribution result of vegetation coverage is obtained.The present application realizes the high-precision inversion of vegetation coverage under complex surface environment, and enhances the ecological rationality and interpretability of model results.

Description

Technical Field

[0001] This invention relates to the field of eco-hydrology technology, specifically providing a method for vegetation cover inversion based on eco-hydrological constraint graph convolutional networks. Background Technology

[0002] Vegetation cover (FVC) is an important ecological parameter characterizing the growth status and spatial distribution pattern of surface vegetation, and a crucial indicator for measuring the structure and function of regional ecosystems. In arid and semi-arid regions, changes in vegetation cover are particularly sensitive to ecological and environmental evolution, serving as a vital basis for evaluating the effectiveness of ecological restoration and desertification control. Therefore, accurately obtaining spatial distribution information on vegetation cover is of great significance for watershed ecological environmental protection and sustainable development.

[0003] In existing technologies, various technical approaches have been developed for the inversion of vegetation cover. Among the methods based on optical remote sensing and vegetation indices, one proposed method for vegetation cover inversion based on high-resolution satellite data, which estimates vegetation cover by calculating the Normalized Difference Vegetation Index (NDVI) and combining it with a neural network model; another proposed method is to correct the NDVI before inversion to improve the accuracy of vegetation cover inversion. These methods mainly rely on spectral information to construct models, which improves inversion accuracy to some extent, but they are still limited to spectral feature representation and cannot reflect the impact of complex ecological and environmental conditions on vegetation growth.

[0004] Among the methods based on multi-source remote sensing data and machine learning, a vegetation cover inversion method based on multi-remote sensing factor random forest is proposed, which achieves inversion by fusing multiple indices and constructing a random forest model; a vegetation cover inversion method combining geographic zoning is proposed, which improves the inversion accuracy of different regions through zonal modeling; in addition, a vegetation cover inversion optimization method based on interpretable artificial intelligence and an inversion method using XGBoost model and feature optimization method are also included. These methods improve the accuracy of vegetation cover inversion to some extent by introducing machine learning and multi-source features, but overall they are still mainly data-driven and lack explicit modeling of spatial neighborhood relationships and ecological environment impacts.

[0005] In spatial information modeling, with the development of deep learning technology, methods such as Graph Convolutional Networks (GCNs) are gradually being applied to the field of remote sensing. For example, a method for water body extraction from remote sensing images based on graph convolutional neural networks has been proposed. By constructing a graph structure to realize the propagation of spatial neighborhood information, the accuracy of water body extraction is improved. This type of method shows that introducing spatial neighborhood relationships can effectively enhance the expressive power of remote sensing information. However, existing research mainly focuses on tasks such as target extraction, and its application in vegetation cover inversion is still relatively limited. Moreover, it usually only constructs graph structures based on spatial adjacency relationships, lacking ecological and environmental constraint mechanisms.

[0006] Furthermore, water bodies, as important ecological and environmental elements, significantly influence the growth of surrounding vegetation through processes such as regulating soil moisture and affecting local climate, and their effects typically exhibit spatial attenuation characteristics. However, existing vegetation cover inversion methods often rely on masking or removal of water body information, lacking a quantitative characterization of the eco-hydrological impact of water bodies and failing to reflect the intrinsic relationship between vegetation spatial distribution and the hydrological environment. Summary of the Invention

[0007] In order to overcome the above-mentioned defects, the present invention is proposed to provide a solution or at least a partial solution to the above-mentioned problems.

[0008] This invention provides a vegetation cover inversion method based on eco-hydrological constraint graph convolutional networks, comprising the following steps: Multispectral remote sensing images of the area to be studied are acquired and preprocessed to obtain effective pixels. Various spectral features, including water index, are calculated to construct a pixel-level multidimensional feature set. Based on the water body index, water body information of the area under study is extracted. Using water body pixels as the source, the distance from the effective pixel to the nearest water body pixel and the preset distance decay function are used to generate the water body influence factor WI, which characterizes the spatial influence intensity of the water body on the surrounding eco-hydrological conditions. Using effective pixels as graph nodes, edges are established between nodes based on spatial adjacency; spatial weights are determined based on the spatial distance between nodes, and ecological similarity weights are determined based on the differences in the water body influence factor (WI) between nodes. The spatial weights and ecological similarity weights are fused to obtain the comprehensive weight of the edges, forming an ecologically weighted graph structure. A graph convolutional network model is trained based on a joint loss function of error loss and ecological ranking constraint loss; wherein, the ecological ranking constraint loss is based on the difference of water body influence factor WI between adjacent nodes, which constrains the ranking relationship between the simulated vegetation cover values ​​output by the model. The pixel-level multidimensional feature set is fused with the water body influence factor WI as node features. The node features and the ecological weighted graph structure are input into the trained graph convolutional network model to output the simulated vegetation coverage value of the effective pixels. Using the trained graph convolutional network model, vegetation cover inversion is performed on the study area to obtain the spatial distribution results of vegetation cover.

[0009] Preferably, the water body index is the Improved Normalized Difference Water Body Index (MNDWI); the water body influence factor WI is constructed by the following formula: In the formula, Let be the Euclidean distance from the i-th valid pixel to the nearest body of water; This is the distance attenuation control parameter.

[0010] Preferably, the distance attenuation control parameter is adaptively determined based on the spatial resolution of the remote sensing image or the scale of the area under study, or is automatically optimized through model optimization methods.

[0011] Preferably, determining the spatial weights based on the spatial distance between nodes includes: Spatial weights are constructed based on the Euclidean distance between pixels, with greater weights for closer pixels. The expression is as follows: In the formula, , These are the row and column coordinates of the corresponding pixels for nodes i and j, respectively.

[0012] Preferably, determining the ecological similarity weight based on the differences in water body impact factors (WI) between nodes includes: based on the differences between nodes. and Based on the differences, an ecological similarity weight is constructed, expressed as: In the formula, , These are the water body influence factors corresponding to nodes i and j, respectively. The ecological similarity weight between node i and node j reflects the degree of similarity between the two nodes in terms of eco-hydrological conditions; Preferably, the comprehensive weight of the edges is expressed as: .

[0013] Preferably, the graph convolutional network model adopts a two-layer graph convolutional structure. The number of layers in the two-layer graph convolutional structure is configured such that the information propagation range between nodes is limited to the local neighborhood, so as to maintain the distinguishability of node features under different water body influence factor WI gradients and avoid the excessive smoothing of node features caused by deep graph convolution.

[0014] Preferably, the two-layer graph convolutional structure is specifically as follows: The first graph convolutional layer has the same input dimension as the fused node feature dimension and an output dimension of 64. The second graph convolutional layer has an input dimension of 64 dimensions and an output dimension of 32 dimensions. A fully connected layer maps the 32-dimensional features output by the second graph convolutional layer to a 1-dimensional output to obtain the simulated value of vegetation coverage. Nonlinear activation functions are introduced between layers to achieve feature mapping.

[0015] Preferably, the method further includes: using a mean squared error loss function to constrain the difference between the simulation results and the measured values ​​to obtain an error loss function. : In the formula, The number of samples; This represents the simulated vegetation cover of the i-th pixel; This represents the measured vegetation cover of the i-th pixel; Based on the adjacency relationships between nodes in a graph structure, adjacent node pairs... Constructing an ecological ranking constraint loss function : In the formula, , These represent the water body influence factors of node i and node j, respectively. Represents the set of edges in the graph; , These represent the simulated vegetation cover of nodes i and j, respectively. The interval parameter controls the degree of relaxation of the sorting constraints. = At that time, no sorting constraint is imposed on the node pair; The joint loss function is expressed as follows: In the formula, These are weighting coefficients used to balance the impact between error constraints and ecological constraints.

[0016] Preferably, the spatial adjacency relationship adopts the 8-neighborhood rule: traverse the 8 neighboring cells around each valid cell, and if the neighboring cell is a valid cell, then establish an undirected edge between the corresponding nodes.

[0017] The beneficial effects of this invention are as follows: 1. By organically integrating remote sensing spectral information, spatial neighborhood relationships and eco-hydrological mechanisms, it overcomes the shortcomings of existing vegetation cover inversion methods that mainly rely on spectral information and lack spatial and ecological constraints, and realizes high-precision inversion of vegetation cover under complex surface environments.

[0018] 2. It enhances the ecological rationality and interpretability of the model results, improves the methodological system for remote sensing vegetation parameter inversion, and provides technical support for ecological environment monitoring and assessment.

[0019] 3. This invention is not limited to vegetation cover inversion applications. It can be flexibly extended to other vegetation parameter inversion tasks according to the ecological environment characteristics and data conditions of different regions. Furthermore, based on the hydrological conditions, topographic conditions, or water body influence characteristics of specific study areas, corresponding eco-hydrological constraints and graph structure construction strategies can be introduced to achieve adaptive modeling and refined expression of different surface ecological processes. Attached Figure Description

[0020] The disclosure of this invention will become more readily understood with reference to the accompanying drawings. It will be readily understood by those skilled in the art that these drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Furthermore, similar numbers in the drawings are used to denote similar components, wherein: Figure 1 This is a flowchart illustrating a vegetation cover inversion method based on an eco-hydrological constraint graph convolutional network according to an embodiment of the present invention.

[0021] Figure 2 This is a measured distribution map of vegetation cover in a local area of ​​a tail lake in an arid region of Northwest China, which is an example of the present invention.

[0022] Figure 3 This is a schematic diagram illustrating the construction of the water body influence factor WI in an example of the present invention.

[0023] Figure 4 This is a schematic diagram of a graph convolutional network model structure according to an example of the present invention.

[0024] Figure 5 This is a scatter plot of the model's training, validation, and testing phases in an example of the present invention. Detailed Implementation

[0025] Some embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0026] like Figure 1-5 As shown, this invention provides a vegetation cover inversion method based on eco-hydrological constraint graph convolutional networks, comprising the following steps: Step S1: Acquire multispectral remote sensing images of the area to be studied and preprocess them to obtain effective pixels. Calculate various spectral features, including water index, and construct a pixel-level multidimensional feature set.

[0027] In this embodiment, Landsat series multispectral remote sensing images of the study area are acquired, including surface reflectance data in the blue (B2), green (B3), red (B4), near-infrared (B5), and shortwave infrared (B6) bands. Vegetation coverage reference data of the corresponding area are obtained through UAV remote sensing measurements to establish a remote sensing basic dataset. Figure 2 This is a measured distribution map of vegetation coverage in a local area of ​​a tail lake in an arid region of Northwest China, as an example of the present invention.

[0028] The acquired remote sensing images are preprocessed, including removing invalid pixels such as clouds and shadows based on quality control bands, performing reflectivity scaling on the images, extracting the set of valid pixels, and obtaining the spatial location information of each pixel.

[0029] Spectral characteristics include spectral indices and multi-band surface reflectance. Preferably, spectral indices include the Normalized Difference Vegetation Index (NDVI), the Modified Normalized Difference Water Index (MNDWI), and the Enhanced Vegetation Index (EVI), which can be flexibly selected and combined according to the surface cover characteristics of the area under study. Multi-band surface reflectance may include reflectance data from the blue, green, red, near-infrared, and shortwave infrared bands.

[0030] Based on spectral indices and combined with multi-band surface reflectance information, a pixel-level 8-dimensional basic feature set is constructed. The input features are standardized to eliminate the impact of differences in feature dimensions on model training.

[0031] Step S2: Based on the water body index, extract water body information of the area to be studied. Using water body pixels as the source, and utilizing the distance from the effective pixel to the nearest water body pixel and a preset distance attenuation function, generate a water body influence factor WI that characterizes the spatial influence intensity of the water body on the surrounding eco-hydrological conditions.

[0032] In this embodiment, the specific method for water body extraction is as follows: water bodies are identified based on a threshold set according to the MNDWI index. Preferably, pixels with an MNDWI greater than 0 are identified as water body pixels, and a water body distribution mask is generated.

[0033] Of course, water body extraction methods are not limited to the MNDWI thresholding method. Other water body extraction methods such as Normalized Differential Water Index (NDWI), support vector machine classification, and deep learning semantic segmentation can also be used, as long as the water body distribution information of the study area can be accurately obtained.

[0034] Based on water body extraction, the Euclidean distance from each effective pixel to the nearest water body pixel is calculated using a distance transformation method. Distance transformations can include Euclidean distance transformation, urban street distance transformation, or chessboard distance transformation, with Euclidean distance transformation being the preferred method.

[0035] In one embodiment, the water body impact factor WI is constructed by the following formula: In the formula, Let be the Euclidean distance from the i-th valid pixel to the nearest body of water; This is the distance attenuation control parameter.

[0036] In one embodiment, the distance attenuation control parameter can be determined based on the spatial resolution of the remote sensing image. For example, when the spatial resolution is 30 meters, it can be set to a value within the range of 100 to 500 meters; it can also be adaptively determined based on the scale of the area under study; or it can be automatically optimized using model optimization methods. As model hyperparameters, their optimal values ​​are automatically selected during training based on validation set performance. Regardless of the method used to determine them, the key is to adjust the spatial attenuation range of the water body's influence.

[0037] It should be noted that the water body impact factor (WI) characterizes the spatial intensity of the influence of water bodies as eco-hydrological elements on the growth conditions of vegetation in the surrounding area. The closer to the water body, the larger the WI, indicating that the eco-hydrological influence of the water body on vegetation growth is more significant through processes such as regulating soil moisture and influencing local microclimate. This design elevates the eco-hydrological influence of water bodies from traditional masking or removal treatments to a continuous influence gradient expression with spatial decay characteristics.

[0038] Step S3: Using effective pixels as graph nodes, establish edges between nodes based on spatial adjacency; determine spatial weights based on the spatial distance between nodes, determine ecological similarity weights based on the differences in the water body influence factor WI between nodes, and fuse the spatial weights and ecological similarity weights to obtain the comprehensive weights of the edges, forming an ecologically weighted graph structure.

[0039] In this embodiment, the graph is constructed as follows: each valid cell is treated as a graph node, and whether an edge connection is established between nodes is determined based on spatial adjacency.

[0040] In one embodiment, the spatial adjacency relationship adopts the 8-neighborhood rule: traverse the neighboring cells in the eight directions (up, down, left, right and four diagonal directions) around each valid cell. If the neighboring cell is a valid cell, then an undirected edge is established between the corresponding nodes to form the edge set of the graph.

[0041] Alternatively, spatial adjacency can be implemented using a 4-neighborhood rule (up, down, left, and right directions), or a wider neighborhood can be used depending on the pixel spatial resolution and the scale of the study area. For remote sensing images with coarser spatial resolution, the neighborhood range can be appropriately expanded.

[0042] Of course, the neighborhood rules for constructing graph structures are not limited to the methods listed above, and can be flexibly selected and adjusted according to the spatial resolution of remote sensing images and the degree of surface heterogeneity of the area under study.

[0043] In one embodiment, determining spatial weights based on the spatial distance between nodes includes: Spatial weights are constructed based on the Euclidean distance between pixels, with greater weights for closer pixels. The expression is as follows: In the formula, , These are the row and column coordinates of the corresponding pixels for nodes i and j, respectively.

[0044] In one embodiment, determining the ecological similarity weight based on the differences in the water body impact factor (WI) between nodes includes: based on the differences between nodes. and Based on the differences, an ecological similarity weight is constructed, expressed as: In the formula, , These are the water body influence factors corresponding to nodes i and j, respectively. The ecological similarity weight between node i and node j reflects the degree of similarity between the two nodes in terms of eco-hydrological conditions.

[0045] In this embodiment, by encoding the similarity of water body influencing factors as part of the edge weights, the graph structure not only encodes "whether pixels are spatially adjacent" but also "whether pixels are similar in an eco-hydrological sense," thereby giving nodes with similar eco-hydrological characteristics a greater weight for mutual influence during information propagation.

[0046] In one embodiment, the combined weight of the edges is expressed as: .

[0047] Of course, the fusion method of comprehensive weighting is not limited to multiplicative fusion. Other fusion methods such as weighted summation, minimum value taking, and harmonic average can also be used. Those skilled in the art can flexibly choose the fusion strategy based on prior knowledge of the relative importance between spatial factors and ecological factors.

[0048] Step S4: Train the graph convolutional network model based on the joint loss function of error loss and ecological ranking constraint loss; wherein, the ecological ranking constraint loss is based on the difference of water body influence factor WI between adjacent nodes, and constrains the ranking relationship between the simulated vegetation coverage values ​​output by the model.

[0049] In this embodiment, the training objective of the graph convolutional network model is to minimize the error between the model's simulated values ​​and the measured values, while enabling the model to learn the spatial distribution characteristics of vegetation cover that conform to ecological laws.

[0050] In one embodiment, the method further includes: using a mean squared error loss function to constrain the difference between the simulation results and the measured values ​​to obtain an error loss function. : In the formula, The number of samples; This represents the simulated vegetation cover of the i-th pixel; This represents the measured vegetation cover of the i-th pixel; Based on the adjacency relationships between nodes in a graph structure, adjacent node pairs... Constructing an ecological ranking constraint loss function : In the formula, , These represent the water body influence factors of node i and node j, respectively. Represents the set of edges in the graph; , These represent the simulated vegetation cover of nodes i and j, respectively. The interval parameter controls the degree of relaxation of the sorting constraints. = At that time, no sorting constraint is imposed on the node pair; The joint loss function is expressed as follows: In the formula, These are weighting coefficients used to balance the impact between error constraints and ecological constraints.

[0051] The value of the weighting coefficient λ is determined comprehensively based on the intensity of eco-hydrological influences and the required inversion accuracy in the study area. For example, in arid tailwater regions where the eco-hydrological influences on water bodies are significant, a larger λ value can be set to strengthen the ecological ordination constraint; in humid regions, a smaller λ value can be set to avoid over-constraint. Of course, λ can also be optimized and determined through cross-validation or hyperparameter search methods.

[0052] Preferably, during model training, the effective pixel samples are randomly divided into training, validation, and test sets in a ratio of 7:1.5:1.5. The Adam optimizer is used to train the model until the validation set loss converges. During training, the joint loss function is used as the optimization objective, and the model parameters are iteratively updated through backpropagation. To quantitatively evaluate the model's inversion accuracy, the root mean square error (RMSE) and coefficient of determination (R²) are introduced on the training, validation, and test sets, respectively. 2 The consistency and fitting effect between the model's simulated values ​​and the measured values ​​are evaluated.

[0053] Alternatively, in an alternative approach, the sample split ratio can be flexibly adjusted to 6:2:2 or 8:1:1, depending on the total sample size; the optimizer can also employ other optimization algorithms such as SGD or RMSprop. Regardless of the sample split ratio or optimization algorithm used, the goal is to minimize the loss function while ensuring the model's generalization ability.

[0054] Step S5: Fuse the pixel-level multidimensional feature set with the water body influence factor WI as node features, input the node features and the ecological weighted graph structure into the trained graph convolutional network model, and output the simulated vegetation coverage value of the effective pixels.

[0055] In this embodiment, the node features are constructed by concatenating and fusing the various spectral indices and band reflectances in the pixel-level multidimensional feature set constructed in step S1 with the water body influence factor WI calculated in step S2 to form a complete feature vector for each graph node. For example, when the multidimensional feature set contains 8-dimensional features, the fusion results in a 9-dimensional node feature vector.

[0056] In one embodiment, the graph convolutional network model adopts a two-layer graph convolutional structure. The number of layers in the two-layer graph convolutional structure is configured to limit the information propagation range between nodes to a local neighborhood, so as to maintain the distinguishability of node features under different water body influence factor WI gradients and avoid excessive smoothing of node features caused by deep graph convolution.

[0057] It should be noted that in multi-layered graph convolutional networks, as the number of convolutional layers increases, node features gradually converge (i.e., oversmoothing). When the number of network layers is too large, the feature differences between high-WI nodes near water bodies and low-WI nodes far from water bodies will be smoothed out, causing the model to lose its ability to distinguish eco-hydrological gradients. The two-layer graph convolutional structure used in this invention effectively limits the information propagation range to a local neighborhood, maintaining the feature differences between nodes at different eco-hydrological gradient locations, ensuring that the model can accurately capture the spatial gradient changes of vegetation cover radiating outward from the water body. This layer design is suitable for the application requirements of this invention in eco-hydrological modeling in arid and semi-arid regions.

[0058] In one embodiment, the two-layer graph convolutional structure is specifically as follows: The first graph convolutional layer has the same input dimension as the fused node feature dimension and an output dimension of 64. The second graph convolutional layer has an input dimension of 64 dimensions and an output dimension of 32 dimensions. A fully connected layer maps the 32-dimensional features output by the second graph convolutional layer to a 1-dimensional output to obtain the simulated value of vegetation coverage. Nonlinear activation functions are introduced between layers to achieve feature mapping. These nonlinear activation functions can be ReLU, LeakyReLU, ELU, etc., with ReLU being the preferred choice to balance computational efficiency and model performance.

[0059] It should be noted that the number of layers and the dimensions of each layer in a graph convolutional network model are not limited to the specific values ​​listed above. The output dimension of the first graph convolutional layer can be 32, 64, or 128, or other dimensions; the output dimension of the second graph convolutional layer can be 16, 32, or 64, or other dimensions. The network structure can be flexibly adjusted according to the dimension of the input features, the sample size, and the surface complexity of the region under study.

[0060] Alternatively, in an alternative approach, graph convolutional networks can employ a single-layer graph convolutional structure with fully connected layers, or a graph convolutional structure with three or more layers, combined with improvements such as residual connections and attention mechanisms to alleviate the problem of deep oversmoothing. Regardless of the specific network structure used, as long as it can effectively learn the representation of node features that integrate eco-hydrological information, it is acceptable.

[0061] Step S6: Using the trained graph convolutional network model, perform vegetation cover inversion in the study area to obtain the spatial distribution results of vegetation cover.

[0062] In this embodiment, the node features and ecological weighted graph structure of all valid pixels in the study area are input into the graph convolutional network model trained in steps S4 and S5. The model outputs a simulated vegetation cover value corresponding to each valid pixel. Based on the spatial location information of each valid pixel, the simulated value is mapped back to the geographic space of the study area to generate a spatial distribution map of vegetation cover.

[0063] Based on this, to quantitatively evaluate the inversion accuracy of the model, the root mean square error (RMSE) and coefficient of determination (R²) as shown in Table 1 were introduced on the training set, validation set, and test set, respectively. 2 The consistency and fitting effect between the model's simulated values ​​and measured values ​​are evaluated to verify the effectiveness and stability of the constructed model in vegetation cover inversion. Figure 5This is a scatter plot of the model's fitting during the training, validation, and testing phases in an example of the present invention.

[0064] Table 1

[0065] The application of the vegetation cover inversion method based on eco-hydrologically constrained graph convolutional networks in this invention is as follows: It can be applied to high-precision inversion modeling of vegetation cover in complex ecological environments. By constructing a graph convolutional network model that integrates remote sensing spectral information, spatial neighborhood relationships, and eco-hydrological mechanisms, and introducing an ecological ranking constraint mechanism based on water body influence factors (WI) during model training, the model's ability to express the spatial distribution characteristics of vegetation cover is improved. Utilizing the modeling ability of graph convolutional networks to depict spatial relationships between pixels and the guiding role of eco-hydrological constraints on physical laws, a refined inversion of vegetation cover is achieved. This overcomes the shortcomings of traditional methods based on empirical indices or single-pixel regression, which struggle to characterize spatial dependencies and ecological constraints. It improves the vegetation parameter inversion method system driven by multi-source remote sensing information, providing technical support for regional vegetation dynamic monitoring, ecological environment assessment, and hydro-ecological regulation.

[0066] It should be noted that this invention is not limited to vegetation cover inversion applications. It can be flexibly extended to other vegetation parameter inversion tasks according to the ecological environment characteristics and data conditions of different regions. Furthermore, based on the hydrological conditions, topographic conditions, or water body influence characteristics of specific study areas, corresponding eco-hydrological constraints and graph structure construction strategies can be introduced to achieve adaptive modeling and refined expression of different surface ecological processes.

[0067] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the original technical features, and the technical solutions resulting from these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for vegetation cover inversion based on eco-hydrological constraint graph convolutional networks, characterized in that, Includes the following steps: Multispectral remote sensing images of the area to be studied are acquired and preprocessed to obtain effective pixels. Various spectral features, including water index, are calculated to construct a pixel-level multidimensional feature set. Based on the water body index, water body information of the area under study is extracted. Using water body pixels as the source, the distance from the effective pixel to the nearest water body pixel and the preset distance decay function are used to generate the water body influence factor WI, which characterizes the spatial influence intensity of the water body on the surrounding eco-hydrological conditions. Using effective pixels as graph nodes, edges are established between nodes based on spatial adjacency; spatial weights are determined based on the spatial distance between nodes, and ecological similarity weights are determined based on the differences in the water body influence factor (WI) between nodes. The spatial weights and ecological similarity weights are fused to obtain the comprehensive weight of the edges, forming an ecologically weighted graph structure. A graph convolutional network model is trained based on a joint loss function of error loss and ecological ranking constraint loss; wherein, the ecological ranking constraint loss is based on the difference of water body influence factor WI between adjacent nodes, which constrains the ranking relationship between the simulated vegetation cover values ​​output by the model. The pixel-level multidimensional feature set is fused with the water body influence factor WI as node features. The node features and the ecological weighted graph structure are input into the trained graph convolutional network model to output the simulated vegetation coverage value of the effective pixels. Using the trained graph convolutional network model, vegetation cover inversion is performed on the study area to obtain the spatial distribution results of vegetation cover.

2. The method according to claim 1, characterized in that, The water body index is the Improved Normalized Difference Water Body Index (MNDWI); the water body influence factor WI is constructed by the following formula: In the formula, Let be the Euclidean distance from the i-th valid pixel to the nearest body of water; This is the distance attenuation control parameter.

3. The method according to claim 2, characterized in that, The distance attenuation control parameters are adaptively determined based on the spatial resolution of the remote sensing image or the scale of the area under study, or are automatically optimized through model optimization methods.

4. The method according to claim 2, characterized in that, Determining spatial weights based on the spatial distance between nodes includes: Spatial weights are constructed based on the Euclidean distance between pixels, with greater weights for closer pixels. The expression is as follows: In the formula, , These are the row and column coordinates of the corresponding pixels for nodes i and j, respectively.

5. The method according to claim 4, characterized in that, The ecological similarity weight is determined based on the differences in water body impact factors (WI) between nodes, including: based on the differences between nodes. and Based on the differences, an ecological similarity weight is constructed, expressed as: In the formula, , These are the water body influence factors corresponding to nodes i and j, respectively. The ecological similarity weight between node i and node j reflects the degree of similarity between the two nodes in terms of eco-hydrological conditions.

6. The method according to claim 5, characterized in that, The overall weight of the edge is expressed as: 。 7. The method according to claim 1, characterized in that, The graph convolutional network model adopts a two-layer graph convolutional structure. The number of layers in the two-layer graph convolutional structure is configured to limit the information propagation range between nodes to the local neighborhood, so as to maintain the distinguishability of node features under different water body influence factors WI gradients and avoid the excessive smoothing of node features caused by deep graph convolution.

8. The method according to claim 7, characterized in that, The two-layer graph convolutional structure is specifically as follows: The first graph convolutional layer has the same input dimension as the fused node feature dimension and an output dimension of 64. The second graph convolutional layer has an input dimension of 64 dimensions and an output dimension of 32 dimensions. A fully connected layer maps the 32-dimensional features output by the second graph convolutional layer to a 1-dimensional output to obtain the simulated value of vegetation coverage. Nonlinear activation functions are introduced between layers to achieve feature mapping.

9. The method according to claim 1, characterized in that, The method further includes: using a mean squared error loss function to constrain the difference between the simulation results and the measured values ​​to obtain the error loss function. : In the formula, The number of samples; This represents the simulated vegetation cover of the i-th pixel; This represents the measured vegetation cover of the i-th pixel; Based on the adjacency relationships between nodes in a graph structure, adjacent node pairs... Constructing an ecological ranking constraint loss function : In the formula, , These represent the water body influence factors of node i and node j, respectively. Represents the set of edges in the graph; , These represent the simulated vegetation cover of nodes i and j, respectively. The interval parameter controls the degree of relaxation of the sorting constraints. = At that time, no sorting constraint is imposed on the node pair; The joint loss function is expressed as follows: In the formula, These are weighting coefficients used to balance the impact between error constraints and ecological constraints.

10. The method according to claim 1, characterized in that, The spatial adjacency relationship adopts the 8-neighborhood rule: traverse the 8 neighboring cells around each valid cell, and if the neighboring cell is a valid cell, then establish an undirected edge between the corresponding nodes.

Images