CN-121982129-A - Digital soil mapping method and device based on horizontal-vertical bidirectional graph convolution network

CN121982129ACN 121982129 ACN121982129 ACN 121982129ACN-121982129-A

Abstract

The invention discloses a digital soil mapping method and a device based on a horizontal-vertical bidirectional graph convolution network, which construct a horizontal connecting edge between soil samples with the same soil type and adjacent spatial positions, construct a vertical connecting edge between soil samples with adjacent depth layers in the same soil section, correlate the soil samples with close soil attribute relationship in the spatial direction and the vertical section direction through the horizontal connecting edge and the vertical connecting edge, and more fully utilize structural correlation information between the soil samples based on a constructed soil sample graph structure training graph convolution model, effectively improve the precision and stability of a target soil attribute prediction model, reduce the inconsistency of a prediction result in the spatial and section directions, and provide more reliable technical support for the fields of soil resource management, agricultural production, ecological environment assessment and the like.

Inventors

CHEN SONGCHAO
CHEN ZHONGXING
SHI ZHOU

Assignees

浙江大学

Dates

Publication Date: 20260505
Application Date: 20260318

Claims (9)

1. A digital soil mapping method based on a horizontal-vertical bipartite graph convolutional network, comprising: (1) Taking the obtained soil sample data and corresponding environment covariate data as soil samples, constructing a soil sample set by a plurality of soil samples, and dividing the soil sample set into a training set and a verification set, wherein the soil sample data comprises spatial position information of sampling points, sampling depth information and target soil attribute values; (2) Constructing a horizontal connecting edge between soil samples with the same soil type and adjacent spatial positions by taking each soil sample as a node, constructing a vertical connecting edge between soil samples with adjacent depth layers in the same soil profile, so as to form a soil sample graph structure containing a horizontal and vertical double-phase relation, inputting the soil sample graph structure into a graph rolling network model to obtain a predicted soil attribute value, and training the graph rolling network model through a loss function based on a plurality of soil sample graph structures corresponding to a training set to obtain a horizontal-vertical bidirectional graph rolling network; (3) When the method is applied, the current soil sample data and the corresponding environment covariate data are input into a horizontal-vertical bidirectional graph convolution network to obtain the current predicted soil attribute value, so that digital soil mapping is completed.
2. A digital soil mapping method based on a horizontal-vertical bipartite graph convolutional network according to claim 1, wherein the method of obtaining soil sample data comprises: Collecting a plurality of soil profile samples in a region to be researched, wherein the soil profile samples cover a depth range of 0-200cm, and the soil profile samples are subjected to layered sampling based on a soil coverage type and a topography condition to obtain a plurality of sampling points; And collecting a plurality of sub-samples at each sampling point position, mixing the sub-samples to form a soil sample, simultaneously recording the spatial position information and sampling depth information of the sampling point positions corresponding to the soil sample, and measuring the target soil attribute value of the soil sample, thereby obtaining soil sample data, wherein the spatial position information is longitude and latitude information.
3. The method for digital soil mapping based on horizontal-vertical bipartite graph convolutional network according to claim 2, wherein the soil samples with the same longitude and latitude information are divided into one group, and the soil samples of the training set and the soil samples of the verification set are respectively located in different groups.
4. A digital soil mapping method based on a horizontal-vertical bipartite graph convolutional network according to claim 2 or 3, wherein said target soil property value is soil organic carbon content, said soil organic carbon content being determined by laboratory analysis methods including at least one of dry combustion or elemental analysis.
5. A digital soil mapping method based on a horizontal-vertical bipartite graph convolutional network according to claim 1, wherein the environmental covariate data comprises at least one of climate factors, topography factors, land utilization information, and remote sensing derivative variables.
6. The method for digital soil mapping based on horizontal-vertical bipartite graph convolutional network according to claim 1, wherein the environment covariate data is subjected to unified spatial resolution processing and coordinate system conversion to ensure spatial consistency with the soil sample data.
7. A digital soil mapping method based on a horizontal-vertical bipartite graph convolutional network according to claim 1, wherein the features of the nodes of the soil sample graph structure are constituted by feature vectors of the environment covariates of the corresponding soil samples.
8. The method for digital soil mapping based on horizontal-vertical bipartite graph convolutional network according to claim 1, wherein the soil samples of the validation set are input into the horizontal-vertical bipartite graph convolutional network to obtain predicted soil attribute values, and the accuracy of the horizontal-vertical bipartite graph convolutional network is evaluated by determining coefficients and root mean square errors based on the predicted soil attribute values.
9. A digital soil mapping device based on a horizontal-vertical bidirectional graph convolution network, comprising a computer memory, a computer processor and a computer program stored in the computer memory and executable on the computer processor, wherein the computer memory is provided with the horizontal-vertical bidirectional graph convolution network constructed by the digital soil mapping method based on the horizontal-vertical bidirectional graph convolution network according to any one of claims 1-8; The computer processor, when executing the computer program, performs the steps of: And inputting the current soil sample data and the corresponding environment covariate data into a horizontal-vertical bidirectional graph convolution network to obtain a current predicted soil attribute value, thereby completing digital soil mapping.

Description

Digital soil mapping method and device based on horizontal-vertical bidirectional graph convolution network Technical Field The invention relates to the technical field of digital soil mapping, in particular to a digital soil mapping method and device based on a horizontal-vertical bidirectional graph convolution network. Background The soil is used as an important component of the land ecological system, and has irreplaceable functions in the aspects of maintaining the function of the ecological system, guaranteeing the safety of grains, regulating the global carbon circulation and the like. However, soil systems exhibit significant complexity and heterogeneity in spatial distribution and cross-sectional structure, subject to factors such as climate change, land use change, and human activity disturbance. Therefore, the method for acquiring the soil attribute information with high precision and multiple scales has important significance for scientifically managing soil resources and making related decisions. Digital soil mapping is widely applied to soil information acquisition in regional and global scale as a technical means for establishing a spatial prediction model by using environment covariates and soil observation data. The existing digital soil mapping method is based on a regression model or a machine learning algorithm, and predicts the target soil attribute through an environment covariate, so that the spatial prediction accuracy of the soil attribute is improved to a certain extent. The invention discloses a method for predicting soil mineral combined state organic carbon based on random forest and environment covariates, which comprises the steps of collecting real soil mineral combined state organic carbon content and related environment covariates of a plurality of soil samples, screening the environment covariates based on a recursive characteristic elimination algorithm to obtain important environment covariates, taking the real soil mineral combined state organic carbon content and the corresponding screened environment covariates as sample sets, dividing the sample sets into modeling sets and independent verification sets, training an initial random forest prediction model based on the modeling sets to obtain a random forest prediction model, evaluating prediction accuracy of the random forest prediction model based on the independent verification sets by adopting a decision coefficient and root mean square error, and obtaining a final random forest prediction model when reaching a prediction accuracy threshold, and further discloses a soil mineral combined state organic carbon prediction device based on the random forest and the environment covariates. However, the above method generally regards soil samples at different spatial positions or samples of different depth layers within the same soil section as mutually independent observation objects, and it is difficult to simultaneously characterize the proximity correlation of soil properties in the spatial direction and the interlayer continuity in the vertical section direction. In practical soil systems, there is often a significant structural association between spatially adjacent soil samples and between layers of different depths within the same soil profile. If the spatial and vertical coupling characteristics are not effectively modeled, the insufficient characterization of the soil property change rule by the prediction model is easily caused, so that the further improvement of the prediction precision and stability of the model is limited. In recent years, a graph convolution network is used as a deep learning method capable of carrying out feature propagation and learning on graph structure data, and provides a new technical idea for expressing complex structure relations. By representing the samples as nodes in the graph structure and utilizing a graph convolution mechanism to realize information transmission among the nodes, the association relationship among the samples can be described at the model level, and new possibility is provided for improving the digital soil mapping method. Based on the above, it is necessary to provide a digital soil mapping method capable of simultaneously describing the spatial correlation of soil properties and the vertical continuity of the profile, so as to further improve the accuracy and stability of the soil property prediction model. Disclosure of Invention The invention provides a digital soil mapping method based on a horizontal-vertical bidirectional graph convolution network, which can enhance soil space prediction modeling and further promote the establishment of a more accurate digital soil map. Compared with the prior art, the invention has the beneficial effects that: According to the method, soil samples are used as nodes, horizontal connecting edges are constructed between soil samples with the same soil type and adjacent spatial positions, vertical connecting edges are constr