Digital soil mapping method and device based on horizontal-vertical bidirectional graph convolutional network

Abstract

The invention discloses a digital soil mapping method and device based on a horizontal-vertical bidirectional graph convolutional network, and the method comprises the steps: taking soil samples as nodes, and constructing horizontal connection edges between soil samples which are of the same soil type and are adjacent in spatial positions; the method comprises the following steps: constructing a vertical connecting edge between soil samples in adjacent depth layers in the same soil profile, and associating the soil samples with close soil attribute relationship in a space direction and a vertical profile direction through a horizontal connecting edge and the vertical connecting edge; based on the constructed soil sample graph structure training graph convolution model, structural correlation information between soil samples can be fully utilized, the precision and stability of a target soil attribute prediction model are effectively improved, and the inconsistency of prediction results in the space and section directions is reduced. And a more reliable technical support is provided for the fields of soil resource management, agricultural production, ecological environment evaluation and the like.

Description

Technical Field

[0001] This invention relates to the field of digital soil mapping technology, and more specifically to a digital soil mapping method and apparatus based on a horizontal-vertical bidirectional graph convolutional network. Background Technology

[0002] Soil, as a vital component of terrestrial ecosystems, plays an irreplaceable role in maintaining ecosystem functions, ensuring food security, and regulating the global carbon cycle. However, influenced by factors such as climate change, land-use change, and human activities, soil systems exhibit significant complexity and heterogeneity in spatial distribution and profile structure. Therefore, obtaining high-precision, multi-scale soil property information is crucial for the scientific management of soil resources and the formulation of related decisions.

[0003] Digital soil mapping, a technique that utilizes environmental covariates and soil observation data to establish spatial prediction models, has been widely applied to soil information acquisition at regional and global scales. Existing digital soil mapping methods are mostly based on regression models or machine learning algorithms, using environmental covariates to predict target soil properties, which has improved the spatial prediction accuracy of soil properties to some extent.

[0004] Patent application CN115758270A discloses a method for predicting soil mineral-bound organic carbon based on random forest and environmental covariates. The method includes: analyzing the true soil mineral-bound organic carbon content and related environmental covariate sets of multiple soil samples; filtering the environmental covariate set using a recursive feature elimination algorithm to obtain important environmental covariates; using the true soil mineral-bound organic carbon content and the corresponding filtered environmental covariate set as a sample set, dividing the sample set into a modeling set and an independent validation set; training an initial random forest prediction model based on the modeling set to obtain the random forest prediction model; evaluating the prediction accuracy of the random forest prediction model using the coefficient of determination and root mean square error based on the independent validation set; and obtaining the final random forest prediction model when the prediction accuracy threshold is reached. This invention also discloses a device for predicting soil mineral-bound organic carbon based on random forest and environmental covariates.

[0005] However, the aforementioned methods typically treat soil samples from different spatial locations or from different depths within the same soil profile as independent observation objects, making it difficult to simultaneously characterize the spatial proximity correlation of soil properties and the interlayer continuity in the vertical profile direction. In real soil systems, significant structural correlations often exist between spatially adjacent soil samples and between different depths within the same soil profile. If these spatial and vertical coupling characteristics are not effectively modeled, the prediction model may fail to adequately characterize the patterns of soil property variation, thus limiting further improvements in model prediction accuracy and stability.

[0006] In recent years, graph convolutional networks (GCNNs), as a deep learning method capable of feature propagation and learning on graph-structured data, have provided a new technical approach for representing complex structural relationships. By representing samples as nodes in a graph structure and utilizing graph convolution mechanisms to achieve information propagation between nodes, the correlation between samples can be characterized at the model level, offering new possibilities for improving digital soil mapping methods. Based on this, it is necessary to propose a digital soil mapping method that can simultaneously characterize the spatial correlation of soil properties and the vertical continuity of soil profiles, in order to further improve the accuracy and stability of soil property prediction models. Summary of the Invention

[0007] This invention provides a digital soil mapping method based on a horizontal-vertical bidirectional graph convolutional network, which can enhance soil spatial prediction modeling and thus promote the creation of more accurate digital soil maps.

[0008] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention constructs a system using soil samples as nodes, establishing horizontal connecting edges between soil samples of the same soil type and spatially adjacent locations, and vertical connecting edges between soil samples at adjacent depths within the same soil profile. These horizontal and vertical connecting edges associate soil samples with closely related soil properties in both spatial and vertical profile directions. The convolutional model trained based on this constructed soil sample map structure can more fully utilize the structural correlation information between soil samples, effectively improving the accuracy and stability of the target soil property prediction model, reducing inconsistencies in prediction results in spatial and profile directions, and providing more reliable technical support for soil resource management, agricultural production, and ecological environment assessment. Attached Figure Description

[0009] Figure 1 A flowchart of a digital soil mapping method based on a horizontal-vertical bidirectional graph convolutional network is provided in this embodiment of the invention. Figure 2 The scatter plot of measured and predicted values ​​of target soil property content in the validation set provided in the embodiments of the present invention represents the accuracy of the target soil property prediction model estimated by the present invention. Detailed Implementation

[0010] To address the challenge of existing digital soil mapping methods in simultaneously characterizing the spatial correlation and vertical continuity of soil properties, this invention proposes a digital soil mapping method based on a horizontal-vertical bidirectional graph convolutional network. This method constructs soil samples into a graph structure that simultaneously includes spatial proximity relationships and inter-layer relationships at depth within the profile. It then utilizes a graph convolutional network to propagate and aggregate information from the node features of the soil samples, achieving collaborative modeling of spatial and vertical dimensional information. This improves the accuracy and stability of the target soil property prediction model.

[0011] This invention provides a digital soil mapping method based on a horizontal-vertical bidirectional graph convolutional network, such as... Figure 1 As shown, it includes: (1) Obtain a soil sample set and divide the soil sample set into a training set and a validation set: take the obtained soil sample data and the corresponding environmental covariate data as soil samples, construct a soil sample set from multiple soil samples, divide the soil sample set into a training set and a validation set, and the soil sample data includes the spatial location information of the sampling point, the sampling depth information and the target soil attribute value.

[0012] In one specific embodiment, this embodiment provides a method for obtaining soil sample data, including: Multiple soil profile samples were collected in the area to be studied. These soil profile samples covered a depth range of 0-200cm, forming depth layers of different depths. Based on the land cover type and topographic conditions, multiple soil profile samples were stratified and sampled to obtain multiple sampling points. That is, multiple sampling points were formed by sampling at different depth layers.

[0013] In this embodiment, multiple sub-samples are collected at each sampling point and mixed to form a soil sample. At the same time, the spatial location information and sampling depth information of the sampling point corresponding to the soil sample are recorded, and the target soil attribute value of the soil sample is measured to obtain soil sample data. The spatial location information is latitude and longitude information.

[0014] In one specific embodiment, this embodiment assigns soil profile numbers to different soil samples based on latitude and longitude information. Since the soil profile numbers differ for different latitude and longitude, the samples are grouped according to the soil profile numbers. The soil samples in the training set and the soil samples in the validation set are located in different groups, ensuring that samples within the same soil profile do not appear in the training set and the validation set at the same time. The sample set is divided into a training set and a validation set for model training and prediction accuracy evaluation. Since soil samples with the same latitude and longitude, i.e., soil samples located in the same profile, have similar soil properties, if similar soil properties are located in the training set and the validation set respectively, the validation effect of the validation set will be reduced, affecting the accuracy evaluation of the model.

[0015] The target soil attribute value provided in this specific embodiment of the invention is soil organic carbon content. This soil organic carbon content is determined by laboratory analysis methods, including at least one of the dry burning method or elemental analysis. Environmental covariate data undergoes unified spatial resolution processing and coordinate system transformation to ensure spatial consistency with soil sample data.

[0016] (2) Construct a soil sample map structure. Based on the multiple soil sample map structures corresponding to the training set, train the graph convolutional network model through the loss function to obtain a horizontal-vertical bidirectional graph convolutional network: take each soil sample as a node, construct horizontal connecting edges between soil samples of the same soil type and adjacent spatial locations, and construct vertical connecting edges between soil samples located in adjacent depth layers within the same soil profile, thereby forming a soil sample map structure containing both horizontal and vertical bidirectional relationships. Input the soil sample map structure into the graph convolutional network model to obtain predicted soil attribute values. Based on the multiple soil sample map structures corresponding to the training set, train the graph convolutional network model through the loss function to obtain a horizontal-vertical bidirectional graph convolutional network.

[0017] Specifically, each soil sample is constructed as a node in a graph structure, with the node's features consisting of the feature vectors of its corresponding environmental covariates. Based on the spatial proximity of soil samples (i.e., adjacent spatial locations) and assuming the same soil type, horizontal connecting edges are constructed between corresponding soil sample nodes to represent the correlation between spatially adjacent samples with the same soil type; if the soil types are different, the correlation is weaker. Simultaneously, based on the hierarchical relationship between different depth layers within the same soil profile, vertical connecting edges are constructed between soil sample nodes corresponding to adjacent depth layers to represent the continuous variation of soil properties along the vertical profile direction. Through this method, a bidirectional soil sample graph structure containing both horizontal and vertical connecting edges is constructed.

[0018] In a specific embodiment of the present invention, a graph convolutional network model is constructed based on the aforementioned bidirectional soil sample graph structure to propagate and aggregate node features. During model training, node features propagate along horizontal and vertical connection edges within the graph structure, enabling spatial neighborhood information and inter-layer profile depth information to be collaboratively modeled within the graph structure, thereby obtaining node embedding features that integrate spatial and vertical relationships.

[0019] In a specific embodiment of the present invention, a target soil property prediction model is constructed based on the node embedding features, outputting the predicted values ​​of target soil properties for each soil sample at its corresponding spatial location and different depth layers. The model's prediction performance is evaluated using a validation set, and the final prediction model is determined when the model's prediction accuracy meets the preset requirements.

[0020] (3) When applying, the current soil sample data and the corresponding environmental covariate data are input into the horizontal-vertical bidirectional graph convolutional network to obtain the current predicted soil attribute values, thereby completing digital soil mapping.

[0021] In one specific embodiment, the soil samples of the validation set are input into a horizontal-vertical bidirectional graph convolutional network to obtain predicted soil attribute values. The accuracy of the horizontal-vertical bidirectional graph convolutional network is evaluated based on the predicted soil attribute values ​​using the coefficient of determination and root mean square error.

[0022] The accuracy of the model provided by this invention is evaluated using the coefficient of determination and the root mean square error, wherein the coefficient of determination R... 2 for: The root mean square error (RMSE) is: in, n For the number of samples, For the first i The measured content of the target soil properties in each sample data point For the first i The predicted target soil property content of each sample was generated by the prediction model. ， This represents the average measured content of all samples.

[0023] On the other hand, the present invention also provides a digital soil mapping device based on a horizontal-vertical bidirectional graph convolutional network, including a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor, characterized in that the computer memory uses a horizontal-vertical bidirectional graph convolutional network constructed using the aforementioned digital soil mapping method based on a horizontal-vertical bidirectional graph convolutional network. When the computer processor executes the computer program, it performs the following steps: The current soil sample data and corresponding environmental covariate data are input into a horizontal-vertical bidirectional graph convolutional network to obtain the current predicted soil attribute values, thereby completing digital soil mapping.

[0024] A specific embodiment of the present invention provides a digital soil mapping device based on a horizontal-vertical bidirectional graph convolutional network, comprising: The dataset module uses the soil dataset and its corresponding set of environmental covariates as the sample set. The sample dataset is divided into a training set and a validation set. Through random sampling, 70% of the data in the soil dataset is assigned to the training set, and the remaining 30% is assigned to the validation set. These sample data all contain the target soil attributes and environmental covariates.

[0025] In the training module, based on the spatial location and profile depth information of soil samples, each soil sample is constructed as a node in a graph structure. Horizontal connecting edges are constructed according to the spatial proximity relationship between soil samples, and vertical connecting edges are constructed according to the upper and lower layer relationships between different depth layers within the same soil profile, thus forming a soil sample graph structure containing both horizontal and vertical bidirectional relationships. Based on this graph structure, a graph convolutional network is used to propagate and aggregate node features, enabling information in the spatial and vertical dimensions to propagate collaboratively within the graph structure, obtaining node feature representations that integrate horizontal and vertical relationships, thereby constructing a target soil attribute prediction model. Finally, a digital soil mapping model based on a horizontal-vertical bidirectional graph convolutional network is constructed.

[0026] The validation module uses validation set data to evaluate the prediction effect of the target soil property prediction model. By comparing the difference between the target soil property content predicted by the model and the actual measured value, the prediction accuracy and reliability of the model are evaluated. When the prediction accuracy meets the preset requirements, the final prediction model is determined.

[0027] The output module is used to input the soil sample data to be tested into the corresponding final model to obtain the predicted value of the target soil property.

[0028] In summary, this invention, a digital soil mapping device based on a horizontal-vertical bidirectional graph convolutional network, effectively improves the accuracy and stability of target soil attribute prediction in digital soil mapping by constructing a graph structure that includes spatial proximity and profile depth relationships, and utilizing graph convolutional networks to achieve information propagation and fusion within the graph structure. This device enhances the predictive power and application reliability of digital soil mapping models, providing a more reliable technical means for fields such as soil resource surveys, agricultural management, and ecological environment assessment.

[0029] Example 1 In this embodiment, China is selected as the study area. Based on soil sample data obtained within China, a target soil property prediction model is constructed. The soil sample data includes soil organic carbon observation values ​​at different depths of multiple soil profiles, and records the corresponding latitude and longitude information. Combining other soil physicochemical property data and environmental covariates obtained through remote sensing technology, the digital soil mapping method based on horizontal-vertical bidirectional graph convolutional networks proposed in this invention is used for modeling, thereby improving the prediction accuracy of the target soil properties. The basic steps of this method are as described in steps (1) to (7) of the aforementioned embodiment, and will not be completely repeated. The following mainly shows the specific data and implementation details: Step (1): Soil sample data were obtained within the study area in China. The soil samples were derived from multiple soil profiles and covered multiple standard depth layers. The soil samples were analyzed in the laboratory to determine their soil organic carbon content and related physicochemical properties. At the same time, the latitude, longitude and sampling depth information of each soil sample were recorded.

[0030] Step (2): Based on the latitude and longitude information of the soil samples, extract environmental covariates related to soil organic carbon distribution from multi-source remote sensing images and environmental data products. These environmental covariates include climate factors, topographic factors, land use information, and remote sensing-derived variables. All environmental covariates undergo uniform spatial resolution processing and coordinate system transformation to ensure spatial consistency with the soil sample data.

[0031] Step (3): In this embodiment, the environmental covariate extraction process is completed in a Python programming environment. The latitude and longitude information of the soil sample points is spatially matched with the environmental covariate raster data to extract the environmental covariate values ​​corresponding to each sampling point, thus constructing an environmental feature set for the soil samples. The environmental covariate extraction and data processing process is implemented using the numerical computation and geospatial data processing function library in Python, and the extracted environmental covariates are stored in tensor form for subsequent input to the deep learning model.

[0032] Step (4): Construct a sample set from soil sample data and corresponding environmental covariates, and group them according to soil profile number. While ensuring that samples from the same profile do not appear simultaneously in the training and validation sets, allocate approximately 70% of the profile samples to the training set and the remaining 30% to the validation set. Construct each soil sample as a node in a graph structure, where node features are composed of the corresponding environmental covariate feature vectors. Based on the spatial proximity between soil samples and under the condition of the same soil type, construct horizontal connecting edges between corresponding soil sample nodes to represent the correlation between spatially adjacent samples with the same soil type. Simultaneously, based on the upper and lower layer relationships between different depth layers within the same soil profile, construct vertical connecting edges between soil sample nodes corresponding to adjacent depth layers to represent the continuous variation characteristics of soil properties in the vertical profile direction. Through the above methods, construct a bidirectional soil sample graph structure containing horizontal and vertical connecting edges, and input it into the graph convolutional network model in the form of a sparse adjacency matrix.

[0033] Step (5): In this embodiment, based on the aforementioned soil sample graph structure containing horizontal and vertical connecting edges, a graph convolutional network model is constructed using a Python deep learning environment. The model is implemented based on the PyTorch deep learning framework and its graph neural network extension library, PyTorchGeometric. By performing multi-layer graph convolution operations on node features in the graph structure, joint modeling of spatial neighborhood information and inter-layer information of profile depth is achieved. During model construction, a multi-layer graph convolutional structure is used to perform non-linear mapping on node features, and activation functions and Dropout mechanisms are introduced in the hidden layers to enhance the model's generalization ability. In this embodiment, the graph convolutional network model is set to a three-layer structure, with the hidden layer feature dimension set to 64 and the Dropout ratio set to 0.2. During model training, the Adam optimization algorithm is used to update the model parameters, with a learning rate set to 0.0002 and the mean squared error loss function (MSE) used as the loss function. The above parameter combination is the optimal parameter configuration obtained through experimental verification, used to construct the final soil organic carbon prediction model.

[0034] Step (6): Train the graph convolutional network model using the training set. During training, continuously optimize the model parameters through multiple iterations to improve the model's ability to express the spatial distribution characteristics of soil organic carbon. Specifically, in this embodiment, the maximum number of training rounds is set to 200, and an early stopping strategy is introduced during training. When the model's prediction performance on the validation set does not improve further in several consecutive training rounds, the training process is terminated early to prevent overfitting and improve training efficiency. After the model training is completed, the model's prediction performance is evaluated based on the validation set. The model's prediction accuracy is quantitatively analyzed by comparing the difference between the model's predicted soil organic carbon content and the measured value.

[0035] This embodiment divides the data into a training set (39,564 samples) and a validation set (16,956 samples), used for fitting the prediction model and objectively evaluating the model's accuracy, respectively. The data is evaluated on the same validation data at a Chinese regional scale, such as... Figure 2 As shown, the soil organic carbon density prediction model based on a horizontal-vertical bidirectional graph convolutional network has a determination coefficient R0 on the validation set. 2 The value is 0.49, and the root mean square error (RMSE) is 4.58 kg m. -2 In contrast, the prediction model built using the traditional random forest model has a determination coefficient of approximately R0 on the same validation set. 2 = 0.35, corresponding to a root mean square error (RMSE) of 5.22 kg m -2 The results show that in modeling soil organic carbon density in China, the horizontal-vertical bidirectional graph convolutional network model outperforms the random forest model, with R...2 It increased by 0.14, and RMSE decreased by 0.64 kg m -2 This demonstrates the effectiveness and superiority of the method in this embodiment.

[0036] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. A digital soil mapping method based on a horizontal-vertical bidirectional graph convolutional network, characterized in that, include: (1) The obtained soil sample data and the corresponding environmental covariate data are used as soil samples. Multiple soil samples are used to construct a soil sample set. The soil sample set is divided into a training set and a validation set. The soil sample data includes the spatial location information of the sampling point, the sampling depth information, and the target soil attribute value. (2) Take each soil sample as a node, construct horizontal connecting edges between soil samples of the same soil type and adjacent spatial locations, and construct vertical connecting edges between soil samples of adjacent depth layers in the same soil profile, thereby forming a soil sample map structure containing both horizontal and vertical relationships. Input the soil sample map structure into the graph convolutional network model to obtain predicted soil attribute values. Based on the multiple soil sample map structures corresponding to the training set, train the graph convolutional network model through the loss function to obtain a horizontal-vertical bidirectional graph convolutional network. (3) When applying, the current soil sample data and the corresponding environmental covariate data are input into the horizontal-vertical bidirectional graph convolutional network to obtain the current predicted soil attribute values, thereby completing digital soil mapping.

2. The digital soil mapping method based on a horizontal-vertical bidirectional graph convolutional network according to claim 1, characterized in that, Methods for obtaining soil sample data include: Multiple soil profile samples were collected in the area to be studied, covering a depth range of 0-200cm. Based on land cover type and topographic conditions, the multiple soil profile samples were stratified to obtain multiple sampling points. Multiple sub-samples are collected at each sampling point and mixed to form a soil sample. At the same time, the spatial location information and sampling depth information of the sampling point corresponding to the soil sample are recorded, and the target soil attribute value of the soil sample is measured to obtain soil sample data. The spatial location information is latitude and longitude information.

3. The digital soil mapping method based on a horizontal-vertical bidirectional graph convolutional network according to claim 2, characterized in that, Soil samples with the same latitude and longitude information are grouped together, with the soil samples in the training set and the soil samples in the validation set located in different groups.

4. The digital soil mapping method based on a horizontal-vertical bidirectional graph convolutional network according to claim 2 or 3, characterized in that, The target soil attribute value is soil organic carbon content, which is determined by laboratory analysis methods, including at least one of dry burning method or elemental analysis method.

5. The digital soil mapping method based on a horizontal-vertical bidirectional graph convolutional network according to claim 1, characterized in that, The environmental covariate data include at least one of climate factors, topographic factors, land use information, and remote sensing derived variables.

6. The digital soil mapping method based on a horizontal-vertical bidirectional graph convolutional network according to claim 1, characterized in that, Environmental covariate data are processed with uniform spatial resolution and coordinate system transformation to ensure spatial consistency with soil sample data.

7. The digital soil mapping method based on a horizontal-vertical bidirectional graph convolutional network according to claim 1, characterized in that, The features of the nodes in the soil sample map structure are composed of the feature vectors of the corresponding environmental covariates of the soil sample.

8. The digital soil mapping method based on a horizontal-vertical bidirectional graph convolutional network according to claim 1, characterized in that, Soil samples from the validation set are input into a horizontal-vertical bidirectional graph convolutional network to obtain predicted soil attribute values. The accuracy of the horizontal-vertical bidirectional graph convolutional network is evaluated based on the predicted soil attribute values ​​using the coefficient of determination and root mean square error.

9. A digital soil mapping device based on a horizontal-vertical bidirectional graph convolutional network, comprising a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor, characterized in that, The computer memory uses a horizontal-vertical bidirectional graph convolutional network constructed according to any one of claims 1 to 8. When the computer processor executes the computer program, it performs the following steps: The current soil sample data and corresponding environmental covariate data are input into a horizontal-vertical bidirectional graph convolutional network to obtain the current predicted soil attribute values, thereby completing digital soil mapping.

Images