A soil mapping method, system and program combining convolutional neural network and structural equation model

Abstract

The present invention relates to the technical field of soil property mapping, and in particular to a soil mapping method, system and program that combine a convolutional neural network and a structural equation model. This method establishes a causal relationship network between soil properties and environmental factors through a structural equation model, and transforms this causal structure into the network topology of a CNN to guide the design of the neural network structure and parameter training. By introducing latent variable constraints and a joint loss function, the prediction accuracy, physical consistency and interpretability of the model are improved. Experimental results show that this method has higher accuracy and generalization ability in soil property spatial prediction, significantly superior to existing single data-driven models.

Description

Technical Field

[0001] The present invention relates to the technical field of soil property mapping, and in particular, to a soil mapping method, system and program combining a convolutional neural network and a structural equation model. Background Art

[0002] Soil is an important basic resource for agricultural production and ecological environment, and its quality and distribution have crucial impacts on crop yields, land use planning, and ecological environment protection. To accurately grasp the soil property characteristics within a region, soil mapping is an essential basic task. Currently, Digital Soil Mapping (DSM) methods have been widely applied, which associate soil properties with environmental factors (such as climate, terrain, vegetation, parent material, and human activities, etc.) through quantitative models to achieve spatial interpolation or prediction of soil properties in the target area.

[0003] In the prior art, the commonly used soil mapping methods mainly include the following two categories:

[0004] 1. Traditional statistical and regression methods: such as Kriging interpolation method in geostatistics, multiple regression models, etc. These methods are relatively simple to implement and can, to a certain extent, reveal the linear or approximately linear relationship between soil properties and environmental variables. However, with the increasing complexity and non-linearity of the soil environment system, the prediction accuracy and generalization ability of traditional statistical methods often fail to meet the requirements of high-precision mapping.

[0005] 2. Machine learning or deep learning methods: With the improvement of computer hardware capabilities and data acquisition means, there has been an increasing amount of research on soil mapping using models such as Random Forest, Support Vector Machine (SVM), and Deep Neural Network (DNN). These data-driven models have demonstrated excellent capabilities in capturing the complex non-linear relationships between soil properties and environmental factors. However, since the structures of most deep learning or machine learning models rely on experience for construction, and the prediction process of the models is usually like a "black box" and lacks interpretability, it is difficult to effectively reflect the causal mechanism in the soil formation process. In addition, limited by the quality and quantity of sample data, when facing multi-scale and multi-source environmental factors, the models are prone to problems such as overfitting or insufficient robustness.

[0006] As disclosed in a Chinese invention patent (Publication No.: CN116206011A, Publication Date: June 2, 2023), a digital soil mapping method and system based on multi-source data are provided. The method includes obtaining the soil use types and satellite remote sensing image raster data of the area to be measured, dividing the area to be measured into multiple soil regions, where each soil use type includes at least one soil region, identifying all soil regions to obtain the topographic features of each soil region and the distribution characteristics of soil regions of the same soil use type according to the identification results; determining whether the soil regions of the same soil use type are adjacent; if not, obtaining the sample points of each soil unit, and obtaining the sample data of each sample point according to the sample points for soil mapping. This patent divides the area to be measured into multiple soil regions, and then determines the sample points of each soil region and the corresponding sample data of the sample points respectively, so that the obtained sample data is more representative and improves the prediction accuracy of the prediction model.

[0007] As disclosed in a Chinese invention patent (Publication No.: CN117333579A, Publication Date: January 2, 2024), a method for digital mapping of soil properties based on geographic information technology is provided. The specific steps of the method for digital mapping of soil properties based on geographic information technology are as follows: S1: Soil data collection; S2: Data preprocessing based on GEE; S3: Automatic segmentation of soil covariate data based on ArcGIS; S4: Data augmentation; S5: Multi-modal fusion deep neural network model; S6: Soil data prediction; S7: Automatic digital mapping. This solution can automatically import multi-modal soil data and perform methods and strategies for accurate prediction and mapping. It effectively improves the efficiency and accuracy of the work of geographers and effectively reduces the time wasted in the process.

[0008] Structural Equation Model (SEM) has been widely used in fields such as ecology, environmental science, and soil science because it can explicitly depict the causal path relationship between latent variables and observed variables. Through SEM, prior knowledge and measured data can be integrated to evaluate the direct or indirect effects of environmental factors on soil properties, thereby improving the interpretability and mechanistic nature of the model to a certain extent. However, SEM often assumes that the relationships between variables are approximately linear or simply non-linear, and it is difficult to accurately depict the highly complex or highly non-linear processes existing in the soil-environment system.

[0009] To make full use of the simulation ability of deep learning for non - linear relationships and the advantages of SEM in causal analysis and interpretation, explorations of combining simple artificial neural networks with SEM for soil property prediction or moisture inversion have emerged in the existing literature, and the prediction effect has been improved to a certain extent. However, there is currently no relevant report on combining convolutional neural networks (CNNs) with the causal relationship network extracted by SEM. CNNs have significant advantages in capturing spatial structure features and processing high - dimensional image data, and are particularly suitable for scenarios in soil mapping that require processing multi - source spatial information such as remote sensing images and terrain data. Therefore, how to combine the structural relationship of SEM with the spatial feature extraction ability of CNNs to construct a soil mapping model with interpretability, mechanism, and high - precision prediction ability has become an urgent problem in this field. Summary of the Invention

[0010] To solve the above - mentioned technical problems, the purpose of the present invention is to provide a soil mapping method that combines a convolutional neural network and a structural equation model. By integrating the description of soil - environment causal relationships by SEM into the topological structure and training process of CNN, more accurate and mechanistically interpretable spatial prediction of soil properties is achieved. The application of this technology can significantly improve the accuracy and reliability of digital soil mapping, providing a scientific basis for agricultural production, environmental management, and land - use planning.

[0011] To achieve the above - mentioned purpose, the present invention adopts the following technical solutions:

[0012] A soil mapping method that combines a convolutional neural network (CNN) and a structural equation model (SEM), characterized in that the method comprises the following steps:

[0013] 1) Data pre - processing:

[0014] Collect ground - measured sample point data of soil properties and environmental factor data; remove outliers and clean noise from the sample point data; perform format conversion, projection correction, and spatial resampling on the environmental factor data to ensure the consistency of input data format and resolution;

[0015] 2) Construct a conceptual model:

[0016] According to the prior knowledge of soil formation and development, determine climate, topography, vegetation, parent material, and human activities as latent variables, as well as the corresponding observed variables for each latent variable; define the unidirectional action relationships between latent variables and the latent variable of soil properties to form a conceptual model for soil - environment causal analysis;

[0017] 3) Establish a structural equation model:

[0018] Based on the conceptual model, a structural equation model is established, where the measurement model is used to describe the relationship between observed variables and latent variables, and the structural model is used to describe the causal paths between latent variables; the model parameters are estimated using maximum likelihood estimation or other appropriate methods, and the model is tested and modified through goodness-of-fit indices to obtain the optimal structural equation model; the predicted values of each latent variable are calculated and output through the optimal structural equation model;

[0019] 4) Convert the structural equation model relationship network into a CNN topological structure:

[0020] Map the latent variables and their causal paths in the structural equation model to the network levels and connection relationships of the CNN, so that the hierarchical design of the CNN better reflects the logical relationship of soil-environment factors; according to the influence of different latent variables on soil properties, set the node structures and connection methods of the convolutional layer, fully connected layer, and output layer of the CNN.

[0021] 5) Convolutional neural network training and soil mapping:

[0022] Input the predicted values of the latent variables, environmental factors, and soil property sample data into the CNN model together; on the basis of defining the joint loss function, use the backpropagation algorithm to train and optimize the network parameters; use the trained CNN model to infer and predict the environmental factor data of the target area to obtain a soil property distribution map with higher spatial resolution and higher accuracy.

[0023] Preferably, when preprocessing the environmental factor data in step 1), it further includes: performing cloud removal, noise removal, radiometric correction, and index calculation on spatial data such as remote sensing images to obtain high-quality derived data such as vegetation indices, terrain indices, or night light indices.

[0024] Preferably, when determining the latent variables and observed variables in step 2), it further includes:

[0025] 2.1) The climate latent variable is characterized by annual average temperature and annual precipitation;

[0026] 2.2) The terrain latent variable is characterized by elevation and slope;

[0027] 2.3) The vegetation latent variable is characterized by NPP and NDVI;

[0028] 2.4) The parent material latent variable is characterized by soil type or lithological data;

[0029] 2.5) The human activity latent variable is characterized by night light and population density;

[0030] 2.5) The soil property is used as a single target variable, or multiple soil properties are integrated to construct a soil latent variable.

[0031] Preferably, in step 3), the measurement model and the structural model of the structural equation model are calculated according to the following formula;

[0032] The measurement model describes how the observed variables are related to the latent variables, and the formula is as follows:

[0033] ,

[0034] where y is the observed soil property variable, x is the observed environmental factor variable, η is the potential or composite soil property variable, ξ is the potential environmental factor variable, Λ and Λ x are the loading matrices, ∈ and δ are the measurement errors;

[0035] The structural model describes the paths between the latent variables, and the formula is as follows:

[0036] ,

[0037] where B represents the relationship matrix between the latent variables, Γ represents the influence matrix of the exogenous variables on the endogenous variables, and ζ is the structural error term;

[0038] And / or, when evaluating and modifying the structural equation model in step 3), the goodness-of-fit indexes adopted include the chi-square test value, the root mean square error of approximation (RMSEA), the comparative fit index (CFI), or any combination thereof.

[0039] Preferably, when converting the structural equation model relationship network into a CNN topological structure in step 4), the mapping rules adopted include:

[0040] 4.1) Corresponding the latent variables in the structural equation model to the key hidden nodes or convolutional kernel groups of the CNN model;

[0041] 4.2) Converting the unidirectional causal relationship between the latent variables into the connection direction between the CNN layers or between the feature maps;

[0042] 4.3) Corresponding to the soil property latent variable at the output layer of the CNN, or adding a separate node in the CNN hidden layer to represent the soil latent variable, for comparison with the latent variable value output by the structural equation model.

[0043] Preferably, the joint loss function defined in step 5) is:

[0044] ,

[0045] where Loss SEMs represents the error between the soil latent variable output by the CNN hidden layer and the soil latent variable predicted by the structural equation model, Loss predictIt represents the prediction error of soil properties. α and β are weight coefficients, which can be determined by cross-validation;

[0046] And / or, in step 5), the Adam optimizer is used to train the CNN model, and the mean square error RMSE and the coefficient of determination R² are used for evaluation on the validation dataset. If the results do not meet the expectations, the model is retrained and optimized by adjusting the network depth, the size of the convolutional kernel, or the latent variable mapping method.

[0047] Furthermore, the present invention also provides a soil mapping system for implementing the above method, including:

[0048] A data preprocessing module, which is used to clean, project and convert, and unify the format of the collected soil property sample point data and environmental factor data;

[0049] A structural equation modeling module, which is used to construct and solve a structural equation model, and output the causal path of soil-environment factors and the predicted values of latent variables;

[0050] A CNN topology generation module, which is used to map the causal network obtained from the above structural equation model into the hierarchical structure of a convolutional neural network;

[0051] A model training module, which is used to perform joint loss function training by combining latent variable constraints and soil property prediction errors, and optimize the CNN network parameters;

[0052] A mapping output module, which is used to receive the inference results of the trained CNN model and generate a spatial distribution map of soil properties in the target area.

[0053] Preferably, the data preprocessing module further includes:

[0054] A sub-module interfaced with remote sensing data, which is used to obtain multi-source remote sensing images and perform standardization processing;

[0055] And / or, a sub-module interfaced with a geographic information system (GIS), which is used to read geographic vector data, digital elevation models, and human activity data and unify the coordinate system and resolution;

[0056] And / or, a visualization and evaluation sub-module, which is used to output the predicted soil property distribution results in raster or vector form, and overlay them on the geographic background base map for visualization; calculate RMSE, R², and other accuracy indicators, and generate error statistics and distribution maps to guide the further optimization of the structural equation model and the convolutional neural network.

[0057] Furthermore, the present invention also provides a computer-readable storage medium, on which a computer program or instruction is stored. When the computer program or instruction is executed by a processor, the above method is implemented.

[0058] Furthermore, the present invention also provides a computer program product, including a computer program or instructions, which, when executed by a processor, implement the above method.

[0059] Due to the adoption of the above technical solution, the present invention combines the non-linear modeling ability of the convolutional neural network (CNN) with the advantages of the structural equation model (SEM) in causal relationship analysis and interpretation, achieving more accurate and comprehensive spatial prediction and mapping of soil properties, which is specifically reflected in the following aspects:

[0060] 1. Significantly improve the prediction accuracy of the model: CNN can make full use of the local features and spatial correlations of multi-source spatial data such as remote sensing images and digital elevation models; extract the causal path information between soil and environmental factors through SEM and embed it into the network structure of CNN, enabling the model to better identify and refine the key variables and influencing factors in the process of soil formation and evolution based on prior knowledge in high-dimensional spatial data, thus effectively improving the prediction accuracy.

[0061] 2. Enhance the physical consistency and interpretability of the model: Existing deep learning models often rely on experience for network structure design, resulting in insufficient interpretability and stability; SEM provides a clear causal mechanism between variables. By using the structural relationships determined by SEM as the topological framework and training constraints of CNN, the logical causal relationships of the soil-environment system can be reflected in the input, hidden layer, and even output layer of the network, making the prediction results more interpretable and credible in terms of physical and ecological meanings.

[0062] 3. Make full use of prior knowledge and measured data: SEM can effectively integrate prior knowledge in the study of the soil-environment system, such as the influence mechanisms of climate, topography, parent material, vegetation cover, and human activities on soil formation; use the predicted values of latent variables or structural relationships output by SEM as the guidance and constraints for the CNN training process, improving the training efficiency of CNN and reducing the risk of overfitting; enabling the model to still have good generalization ability and stable performance in the case of relatively limited data volume or unbalanced sample distribution.

[0063] 4. Improve the efficiency and practicality of soil mapping: Compared with simply relying on the linear / weak non-linear assumptions of SEM, the present invention uses CNN to capture highly non-linear features, avoiding over-simplification of the complex relationships in the soil-environment system; in practical applications, an integrated model can be directly established based on multi-source environmental data and a small number of soil measured sample points to achieve high-precision mapping of soil properties in large-scale regions, meeting diverse needs such as agricultural production, environmental protection, and land planning.

[0064] In summary, the present invention overcomes the defect of the difficulty in dealing with non - linear relationships when only using SEM, and makes up for the deficiency of ordinary deep learning methods in terms of mechanism interpretability. By introducing the prior relationship network of the structural equation model and leveraging the powerful spatial feature extraction ability of CNN, the efficiency, accuracy, and interpretability of the soil mapping process are greatly improved, providing a practical technical solution for subsequent applications in the fields of agriculture, ecology, and environment. Brief Description of the Drawings

[0065] Figure 1 is the flowchart of the present invention. Detailed Embodiments

[0066] Next, in combination with the embodiments of the present invention, the technical solutions in the embodiments will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without making creative efforts fall within the protection scope of the present invention.

[0067] As shown in Figure 1, a soil mapping method combining a CNN convolutional neural network and a structural equation model has the following specific implementation steps:

[0068] Step 1: Data pre - processing. Collect ground - measured sample point data of soil properties and environmental parameter data, clean the sample point data to remove outliers and noise, and perform format conversion, data resampling, etc. on the environmental parameter data to ensure that the input data formats and resolutions are consistent.

[0069] Step 2: Define a conceptual model of the relationship between latent variables and observed variables based on the prior knowledge of the impact of the environment on soil. Its specific steps can be further divided into the following two points:

[0070] 1) Define latent variables and observed variables. Determine five types of environmental factors - climate, terrain, vegetation, parent material, and human activities - as latent variables, and determine the direct observed variables for each latent variable. For example, the climate latent variable is characterized by the annual average temperature and annual precipitation, the terrain latent variable is characterized by elevation and slope, the vegetation latent variable is characterized by NPP and NDVI, the parent material latent variable is characterized by soil type, and the human activity latent variable is characterized by night - time light and population density. In addition, the target variable can use a single soil property or construct a soil latent variable through multiple soil properties. The selection of the above - mentioned latent variables and observed variables can be adjusted and extended according to specific requirements.

[0071] 2) Define the interaction relationships between latent variables. In the present invention, the interaction relationships of latent variables include the interactions between environmental latent variables and the influence of environmental latent variables on soil latent variables. All the interaction relationships between latent variables are set to be unidirectional, that is, the influence of one latent variable on another latent variable can only be propagated unidirectionally. These relationships are represented by unidirectional arrows in the structural equation model (SEM), such as climate → vegetation, terrain → climate, terrain → vegetation, etc. Based on this logic, a conceptual model can be constructed in combination with prior knowledge to accurately describe the interaction mechanism between each latent variable.

[0072] Step 3: Construct a structural equation model. On the basis of determining the observed variables, latent variables and their relationships in Step 2, use the structural equation model (SEM) to quantify the path relationships between variables. Evaluate the influence intensity and significance between variables through model fitting, output the optimal structural equation model relationship network, and output the predicted values of latent variables. Its specific steps can be further divided into the following three points:

[0073] 1) Calculate the measurement model and the structural model of the structural equation model according to the following formula.

[0074] The measurement model describes how the observed variables are related to the latent variables, and the formula is as follows:

[0075] ,

[0076] where y is the observed soil property variable, x is the observed environmental factor variable, η is the potential or composite soil property variable, ξ is the potential environmental factor variable, Λ and Λx are the loading matrices, ∈ and δ are the measurement errors.

[0077] The structural model describes the paths between latent variables, and the formula is as follows:

[0078] ,

[0079] where B represents the relationship matrix between latent variables, Γ represents the influence matrix of exogenous variables on endogenous variables, and ζ is the structural error term;.

[0080] 2) Model evaluation and optimization. Use methods such as maximum likelihood estimation to estimate the model parameters, and calculate the estimated values and standard errors of the model parameters. Use fitting indices (such as chi-square test, root mean square error of approximation (RMSEA), comparative fit index (CFI), etc.) to evaluate the model fit, and check the significance of the path coefficients and factor loadings of the model. Adjust the model structure according to the fitting results to improve the fit. The revised model needs to be re-identified, estimated and evaluated, and the optimal model is output.

[0081] 3) Use the optimal structural equation model to predict the values of latent variables, which can be used as input variables for the convolutional neural network or as constraint terms for the loss function.

[0082] Through the above steps and formulas, an intuitive path diagram can be generated to quantify the mutual relationships between soil environmental variables.

[0083] Step 4: Convert the structural equation model relationship network obtained in Step 3 into the topological structure of a CNN neural network, and construct a model that combines the CNN neural network and the structural equation relationship network. Its specific steps can be further divided into the following four points:

[0084] 1) Input data preparation: Use the latent variable values of climate, vegetation, terrain, parent material, and human activities calculated by the structural equation model in Step 3 as new input variables, and input them into the convolutional neural network model together with other original environmental factors. Taking the location of the sample point as the center, extract an environmental factor window with a size of w × w to form input data with a dimension of w × w × m , where m is the dimension of the input variable.

[0085] 2) Construct a convolutional neural network (CNN): Convert the structural equation model relationship network obtained in Step 3 into the topological structure of a CNN neural network, and set several convolutional layers, fully connected layers, and output layers. The number of hidden layer nodes connected to the output layer is kept consistent with the number of latent variables in the SEMs model.

[0086] 3) Define the loss function: Use the soil latent variable values calculated by the structural equation model in Step 3 as label data, calculate the error with the predicted values output by the CNN hidden layer nodes, and define the error as Loss SEMs . Incorporate Loss SEMs into the loss function and perform weighted calculation in combination with the standard prediction error:

[0087] .

[0088] where α and β are weights used to balance the contributions of different terms, and the optimal parameters can be determined through cross-validation.

[0089] 4) Model training optimization and backpropagation: Use the Adam optimizer for iterative training until the model converges. Test the model performance on an independent validation dataset, and optimize the model structure according to the experimental results for soil mapping tasks. Use the independent validation dataset to evaluate the accuracy of the generated mapping results, and calculate the root mean square error (RMSE) and coefficient of determination (R 2Based on indicators such as ), combined with the results of error analysis, adjust the model parameters or optimize the SEMs and CNN structures to further improve the prediction accuracy and stability of the model.

[0090] Those skilled in the art should understand that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0091] The present application is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate for implementing in the process Figure 1 one process or multiple processes and / or blocks Figure 1 a device for the functions specified in one block or multiple blocks.

[0092] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device, and the instruction device implements in the process Figure 1 one process or multiple processes and / or blocks Figure 1 the functions specified in one block or multiple blocks.

[0093] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0094] In a typical configuration, a computing device includes one or more processors (CPUs), an input / output interface, a network interface, and memory.

[0095] The memory may include non-permanent memory in the form of computer-readable media, random access memory (RAM) and / or non-volatile memory such as read-only memory (ROM) or flash RAM. The memory is an example of computer-readable media.

[0096] Computer-readable media includes permanent and non-permanent, removable and non-removable media and can store information by any method or technology. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette tapes, magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory computer-readable media such as modulated data signals and carrier waves.

[0097] Experimental Examples

[0098] The following content is an experimental example designed based on the above invention principle. By testing real or simulated soil and environmental factor data, the technical effects of the method described in the present invention in terms of accuracy, interpretability and stability are verified.

[0099] I. Experimental Area and Data Sources

[0100] Experimental Area: An area of approximately 10,000 square kilometers within a certain province is selected as the study area. This area has significant topographic undulations and diverse climate and vegetation types, which can well represent typical complex soil-environment relationships.

[0101] Measured Soil Property Data: 2000 sample points are evenly distributed in this area, and the target soil properties (such as pH, soil organic matter, soil water content or total soil nitrogen, etc.) are collected and detected, denoted as D s .

[0102] Environmental Factor Data (may include but is not limited to the following types):

[0103] 1) Climate Data: Annual average temperature, annual precipitation, from meteorological station interpolation or meteorological reanalysis data;

[0104] 2) Topographic data: digital elevation model (DEM), slope, aspect, with a resolution of 30 m;

[0105] 3) Vegetation data: Normalized Difference Vegetation Index (NDVI) and Net Primary Productivity (NPP) calculated from remote sensing images;

[0106] 4) Parent material data: soil type or lithology layer;

[0107] 5) Human activity data: night light intensity, population density, etc.

[0108] Data preprocessing: The environmental factor data were projected and resampled to a uniform spatial resolution (30 m), and extreme outliers were removed. The data set is denoted as D e .

[0109] 2. Experimental Design

[0110] 1. Method comparison

[0111] In order to verify the effectiveness of the method of the present invention (denoted as SEM-CNN), the following control group and comparison model were established:

[0112] SEM alone model (only structural equation modeling was used to predict soil properties without deep learning);

[0113] CNN-alone model (using only convolutional neural networks without incorporating the causal structure of the structural equation model);

[0114] Traditional machine learning models (such as random forest RF or support vector machine SVM) as a reference for conventional data-driven methods;

[0115] The method of the present invention (SEM-CNN) combines the causal network structure of SEM with the convolutional and fully connected layer deep modeling of CNN.

[0116] 2. Experimental Procedure

[0117] 1) Construction of structural equation model

[0118] According to the prior knowledge of soil formation, five latent variables, namely "climate", "topography", "vegetation", "parent material" and "human activities", were set. The measurement model and structural model were used to model the 2,000 sample data based on maximum likelihood estimation and appropriate fit index to obtain the causal paths between the latent variables and their effects on the soil latent variables (target soil properties).

[0119] 2) CNN topology design

[0120] Map the causal network extracted by SEM to the network structure of CNN: make the latent variable influence path correspond to the connection between different layers in CNN; select several convolutional layers (such as 2 - 4 layers), and add nodes or output layers corresponding to "soil latent variables" at the end of the fully connected layer.

[0121] 3) Training and validation

[0122] Randomly split the dataset (D s +D e ) into a training set (80%) and a validation / test set (20%); train the four types of models respectively and make predictions on a unified test set; mainly select RMSE (root mean square error) and R 2 (coefficient of determination) as evaluation metrics, and record the training duration or inference efficiency of each model.

[0123] 4) Result comparison and analysis

[0124] Systematically compare the control groups (SEM, CNN, random forest, etc.) with the method of the present invention in terms of soil prediction accuracy, interpretability, stability, etc.; compare the prediction differences before and after introducing latent variable constraints in the SEM - CNN model to illustrate the strengthening effect of the structural equation model.

[0125] IV. Experimental results and data examples

[0126] The following table gives an example of the evaluation results of each model on the test set in the soil pH prediction task (the values are for illustration only and are not real data):

[0127]

[0128] Note: The smaller the RMSE value and the higher the R 2 value, the better the prediction accuracy of the model; the standard deviation reflects the difference in the stability of the model under repeated experiments; the training time is for reference only, and different hardware environments will lead to differences in time consumption.

[0129] It can be seen from the above results:

[0130] 1. The SEM - CNN model has the highest accuracy in predicting soil pH, with the RMSE dropping from 0.47 of the pure CNN to 0.42, and the R 2 also increasing from 0.75 to 0.81;

[0131] 2. Although SEM has an advantage in explaining causal relationships, when used alone, SEM is insufficient in capturing highly non - linear and complex spatial features, so its RMSE is greater than 0.5, and the R 2 is only 0.70;

[0132] 3. Traditional machine learning (random forest) can also obtain relatively reliable results when there are a certain number of training samples, but its prediction accuracy and robustness are still lower than those of SEM-CNN that integrates causal relationships and deep learning.

[0133] 4. Although the training time of SEM-CNN is slightly higher than that of pure CNN, it maintains the stability and high accuracy of the prediction results in multiple experiments and has good interpretability.

[0134] V. Conclusion

[0135] 1. Significant improvement in accuracy: The RMSE and R of the SEM-CNN model on the test set are both better than those of the control group, and it comprehensively integrates the causal network of SEM and the non-linear feature extraction ability of CNN. 2

[0136] 2. Enhanced interpretability: The mechanism of action between each environmental latent variable and soil properties can be clearly demonstrated from the perspective of the causal path, avoiding the defect of the pure "black box" of CNN.

[0137] 3. Application and promotion value: In actual regional soil mapping, this model can use limited measured sample point data and multi-source environmental data to achieve high-precision inference of soil properties in large-scale regions, which is of great significance for agricultural production, ecological monitoring, land use planning, etc.

[0138] In summary, the experimental data and results show that the soil mapping method combining convolutional neural network and structural equation model proposed in the present invention can effectively improve the accuracy and reliability of soil mapping, and has good interpretability and applicability, providing a new method and technical support for digital soil mapping.

[0139] The above is the description of the embodiments of the present invention. Through the above description of the disclosed embodiments, those skilled in the art can implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but will be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A soil mapping method combining a convolutional neural network and a structural equation model, characterized in that, The method includes the following steps: 1) Data preprocessing: Collect ground measured sample point data of soil properties and environmental factor data; eliminate outliers and clean noise from the sample point data; perform format conversion, projection correction, and spatial resampling on the environmental factor data to ensure the consistency of the input data format and resolution; 2) Construct a conceptual model: According to the prior knowledge of soil formation and development, determine climate, terrain, vegetation, parent material, and human activities as latent variables, as well as the observed variables corresponding to each latent variable; define the unidirectional relationship between latent variables and the latent variable of soil properties to form a conceptual model for soil-environment causal analysis; 3) Establish a structural equation model SEM: Based on the conceptual model, establish a structural equation model, where the measurement model is used to describe the relationship between observed variables and latent variables, and the structural model is used to describe the causal paths between latent variables; use maximum likelihood estimation or other appropriate methods to estimate model parameters, and test and correct the model through goodness-of-fit indicators to obtain the optimal structural equation model; calculate and output the predicted values of each latent variable through the optimal structural equation model; 4) Convert the structural equation model relationship network into a convolutional neural network CNN topological structure: Map the latent variables and their causal paths in the structural equation model to the network levels and connection relationships of the CNN, so that the hierarchical design of the CNN better reflects the logical relationship between soil-environment factors; set the node structure and connection method of the convolutional layer, fully connected layer, and output layer of the CNN according to the influence of different latent variables on soil properties; 5) Convolutional neural network training and soil mapping: Input the predicted values of the latent variables, environmental factors, and soil property sample data into the CNN model together; based on the defined joint loss function, use the backpropagation algorithm to train and optimize the network parameters; use the trained CNN model to reason and predict the environmental factor data of the target area to obtain a soil property distribution map with higher spatial resolution and accuracy; In step 3), the measurement model and structural model of the structural equation model are calculated according to the following formula; The measurement model describes how observed variables are related to latent variables, and the formula is as follows: ， where y is the observed soil property variable, x is the observed environmental factor variable, η is the latent or composite soil property variable, ξ is the latent environmental factor variable, Λ and Λ x are the loading matrices, and ϵ and δ are the measurement errors; The structural model describes the paths between latent variables, and the formula is as follows: ， Among them, 𝐵 represents the relationship matrix between latent variables, Γ represents the influence matrix of exogenous variables on endogenous variables, and ζ is the structural error term.

2. The method according to claim 1, characterized in that When preprocessing the environmental factor data in step 1), it further includes: performing cloud removal, noise removal, radiation correction, and index calculation on the remote sensing image spatial data to obtain derived data including vegetation index, terrain index, or night light index.

3. The method according to claim 1, characterized in that When determining latent variables and observed variables in step 2), it further includes: 2.1) The climate latent variable is characterized by annual average temperature and annual precipitation; 2.2) The terrain latent variable is characterized by elevation and slope; 2.3) The vegetation latent variable is characterized by NPP and NDVI; 2.4) The parent material latent variable is characterized by soil type or lithology data; 2.5) The human activity latent variable is characterized by night light and population density; 2.5) The soil properties are used as a single target variable, or multiple soil properties are integrated to construct a soil latent variable.

4. The method according to claim 1, characterized in that When evaluating and modifying the structural equation model in step 3), the goodness-of-fit indices used include the chi-square test value, the root mean square error approximation RMSEA, the comparative fit index CFI, or any combination thereof.

5. The method according to claim 1, wherein When converting the structural equation model relationship network into a CNN topological structure in step 4), the mapping rules used include: 4.1) Corresponding the latent variables in the structural equation model to the key hidden nodes or convolutional kernel groups in the CNN model; 4.2) Converting the unidirectional causal relationships between latent variables into the connection directions between CNN layers or feature maps; 4.3) Corresponding to the soil property latent variable at the output layer of the CNN, or adding a separate node in the CNN hidden layer to represent the soil latent variable, for comparison with the latent variable values output by the structural equation model.

6. The method according to claim 1, wherein The joint loss function defined in step 5) is: ， Among them, Loss SEMs represents the error between the soil latent variable output by the CNN hidden layer and the soil latent variable predicted by the structural equation model. Loss predict represents the prediction error of soil properties. α and β are weight coefficients, which can be determined by cross-validation.

7. The method according to claim 1, characterized in that, In step 5), the Adam optimizer is used to train the CNN model, and the mean square error RMSE and the coefficient of determination R² are used for evaluation on the validation data set. If the results do not meet the expectations, the model is retrained and optimized by adjusting the network depth, the convolutional kernel size, or the latent variable mapping method.

8. A soil mapping system for implementing the method according to any one of claims 1-7, characterized in that, Including: A data preprocessing module for cleaning, projection transformation, and format unification of the collected soil property sample point data and environmental factor data; A structural equation modeling module for constructing and solving a structural equation model, and outputting the causal path and latent variable prediction values of the soil-environmental factors; A CNN topology generation module for mapping the causal network obtained from the above structural equation model into the hierarchical structure of a convolutional neural network; A model training module for performing joint loss function training by combining latent variable constraints and soil property prediction errors, and optimizing the CNN network parameters; A mapping and output module for receiving the inference results of the trained CNN model and generating a spatial distribution map of the soil properties in the target area.

9. The system according to claim 8, wherein, The data preprocessing module further includes: A sub-module interfacing with remote sensing data for obtaining multi-source remote sensing images and performing standardization processing.

10. The system according to claim 8, wherein The data preprocessing module further includes: A sub-module interfacing with the geographic information system GIS for reading geographic vector data, digital elevation models, and human activity data and unifying the coordinate system and resolution.

11. The system according to claim 8, characterized in that, The data preprocessing module further includes: A visualization and evaluation sub-module for outputting the predicted soil property distribution results in raster or vector form, and overlaying them on the geographic background base map for visualization; calculating RMSE, R², and other accuracy indices, and generating error statistics and distribution maps to guide the further optimization of the structural equation model and the convolutional neural network.

12. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instruction is executed by a processor, it implements the method according to any one of claims 1-7.

13. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instruction is executed by a processor, it implements the method according to any one of claims 1-7.

Images