A deep learning-based multi-source soil environment data fusion method

Abstract

The application relates to the field of soil environment monitoring and discloses a multi-source soil environment data fusion method based on deep learning, which comprises the following steps: collecting multi-source soil environment original data and carrying out spatio-temporal alignment preprocessing, adopting a double-branch feature extraction network to respectively extract static global features and dynamic spatio-temporal features, generating a multi-modal fusion feature vector through a cross-modal interaction attention fusion module, obtaining a fusion model through a training framework with soil field knowledge constraint and dynamic weight adjustment, and finally outputting high spatio-temporal resolution fusion data and pixel-level uncertainty evaluation results after multi-source decomposition; the application solves the problems that in the prior art, multi-source soil data fusion is insufficient in complementary information utilization, pure data-driven prediction violates soil physical and chemical rules, and pixel-level multi-dimensional uncertainty quantization cannot be realized.

Description

Technical Field

[0001] This invention relates to the field of soil environmental monitoring, and in particular to a method for fusing multi-source soil environmental data based on deep learning. Background Technology

[0002] Soil environmental data is the core foundational data for soil resource management, precision agriculture, non-point source pollution control, and ecological environmental protection. Traditional soil environmental monitoring mainly relies on fixed-point field sampling combined with laboratory physicochemical analysis. Although this method offers high measurement accuracy, it suffers from inherent drawbacks such as high sampling costs, long monitoring cycles, and limited spatial coverage, making it difficult to meet the current demand for large-scale, high spatiotemporal resolution dynamic monitoring of the soil environment.

[0003] With the rapid development of remote sensing technology, Internet of Things sensor technology, and geographic information system technology, the ability to acquire multi-source heterogeneous soil environmental data has been significantly improved, forming a multi-dimensional data system that includes satellite remote sensing data, ground in-situ sensor data, historical soil survey data, and laboratory analysis data.

[0004] Existing multi-source soil environmental data fusion technologies mostly employ simple weighted averaging or feature splicing strategies, failing to fully explore the intrinsic interaction between static soil properties and dynamic environmental changes, resulting in insufficient utilization of complementary information from multi-source data. Deep learning-based fusion methods driven purely by data lack effective constraints from soil science knowledge, leading to some prediction results that violate fundamental soil physicochemical laws and have unclear physical meanings. Most existing uncertainty assessment methods can only provide global accuracy indicators, failing to quantify uncertainty at the pixel level, and cannot distinguish the uncertainty contributions from different data sources and model structures, severely limiting the application value of fusion results in practical decision-making. Summary of the Invention

[0005] The present invention aims to provide a deep learning-based multi-source soil environmental data fusion method to solve the problems of insufficient utilization of complementary information in multi-source soil data fusion, pure data-driven prediction violating soil physicochemical laws, and inability to achieve pixel-level multi-dimensional uncertainty quantification in the existing technology.

[0006] To achieve the above objectives, the present invention provides the following method:

[0007] This invention provides a method for fusing multi-source soil environmental data based on deep learning:

[0008] S1: Collect raw soil environmental data from multiple sources;

[0009] S2: Perform spatiotemporal alignment preprocessing on the original multi-source soil environment data to obtain spatiotemporally aligned multi-source soil environment data;

[0010] S3: Construct a dual-branch feature extraction network, input the spatiotemporally aligned multi-source soil environmental data into the dual-branch feature extraction network, and extract the static global feature vector and the dynamic spatiotemporal feature tensor respectively;

[0011] S4: Construct a cross-modal interactive attention fusion module, input the static global feature vector and the dynamic spatiotemporal feature tensor into the cross-modal interactive attention fusion module, perform bidirectional information interaction, and generate a multimodal fusion feature vector;

[0012] S5: Construct a deep learning training framework with knowledge constraints in the field of soil science, input the multimodal fusion feature vector into the deep learning training framework with knowledge constraints in the field of soil science for iterative training, and obtain the trained multi-source soil environment data fusion model.

[0013] S6: Obtain multi-source soil environmental data of the area to be fused, input the multi-source soil environmental data of the area to be fused into the trained multi-source soil environmental data fusion model, and output high spatiotemporal resolution soil environmental fusion data and pixel-level uncertainty assessment results.

[0014] Furthermore, the multi-source soil environmental raw data includes static soil property data, dynamic remote sensing monitoring data, ground in-situ sensor data, and laboratory analysis data; the static soil property data includes soil type, soil texture, topographic slope, and altitude; the dynamic remote sensing monitoring data includes Landsat series satellite multispectral data, Sentinel-2 satellite multispectral data, and MODIS vegetation index data; the ground in-situ sensor data includes real-time monitoring data of soil moisture content, soil temperature, and soil electrical conductivity; and the laboratory analysis data includes soil organic matter content, soil pH value, soil cation exchange capacity, and soil heavy metal content data.

[0015] Furthermore, the step of performing spatiotemporal alignment preprocessing on the original multi-source soil environmental data to obtain spatiotemporally aligned multi-source soil environmental data includes: uniformly converting all original multi-source soil environmental data into the WGS84 coordinate system and UTM projection; resampling all data with different spatial resolutions to a uniform 30m×30m spatial resolution using bilinear interpolation; interpolating all data with different temporal resolutions to a uniform daily temporal resolution using cubic spline interpolation; imputing missing values ​​in all data using the K-nearest neighbor interpolation method based on spatial Euclidean distance, with K set to 5; and individually identifying and removing outliers using the 3σ criterion for each feature dimension to obtain spatiotemporally aligned multi-source soil environmental data.

[0016] Furthermore, the step of constructing the dual-branch feature extraction network includes: constructing a static feature extraction branch, which consists of three fully connected layers stacked sequentially, each followed by a ReLU activation function and a batch normalization layer, with the number of neurons in the three fully connected layers being 128, 64, and 32 respectively, and outputting a static global feature vector with a dimension of 32; constructing a dynamic spatiotemporal feature extraction branch, which consists of three ConvLSTM layers stacked sequentially, each with a 3×3 kernel size, a stride of 1, and padding of 1, followed by a ReLU activation function and a batch normalization layer, and outputting a dynamic spatiotemporal feature tensor with a dimension of 32×H×W, where H and W are the height and width of the feature map, respectively; and connecting the static feature extraction branch and the dynamic spatiotemporal feature extraction branch in parallel to form a dual-branch feature extraction network.

[0017] Furthermore, the number of input layer neurons in the static feature extraction branch is consistent with the dimension of the static soil attribute data; the number of input channels in the dynamic spatiotemporal feature extraction branch is consistent with the total number of bands in the dynamic remote sensing monitoring data and the ground in-situ sensor data.

[0018] Furthermore, the step of constructing the cross-modal interactive attention fusion module includes: constructing a static feature-guided dynamic attention submodule, calculating the cosine similarity between the static global feature vector and the features at each time step in the dynamic spatiotemporal feature tensor, normalizing the cosine similarity using a softmax function to obtain dynamic attention weights, and multiplying the dynamic attention weights element-wise with the dynamic spatiotemporal feature tensor to obtain weighted dynamic features; constructing a dynamic feature-enhanced static attention submodule, performing global average pooling on the weighted dynamic features to obtain dynamic global features, calculating the cosine similarity between the dynamic global features and the features at each dimension in the static global feature vector, normalizing the cosine similarity using a softmax function to obtain static attention weights, and multiplying the static attention weights element-wise with the static global feature vector to obtain weighted static features; and constructing a feature concatenation layer, concatenating the weighted static features with the global average features of the weighted dynamic features to generate a multimodal fusion feature vector with a dimension of 64.

[0019] Furthermore, the calculation formula for the static feature-guided dynamic attention submodule is as follows:

[0020] ;

[0021] ;

[0022] in, For the first Dynamic attention weights at each time step. It is a static global feature vector. For the first The dynamic spatiotemporal characteristics of each time step The cosine similarity function is used. This is a weighted dynamic feature.

[0023] Furthermore, the step of obtaining the trained multi-source soil environment data fusion model includes: randomly dividing the spatiotemporally aligned multi-source soil environment data into a training set, a validation set, and a test set in a 7:2:1 ratio; initializing all parameters of the deep learning training framework with soil science domain knowledge constraints, using the Adam optimizer, setting the initial learning rate to 0.001, the batch size to 32, and the weight decay coefficient to 1e-5; inputting the multimodal fusion feature vectors from the training set into the training framework, using a composite loss function for iterative training, and validating on the validation set after each epoch; stopping training when the validation set loss no longer decreases for 10 consecutive epochs, and saving the current model parameters; using the test set to perform accuracy testing on the saved model, and obtaining the trained multi-source soil environment data fusion model when the test set determination coefficient R² is greater than or equal to 0.85.

[0024] Furthermore, the formula for calculating the composite loss function is as follows:

[0025] ;

[0026] ;

[0027] in, This is the total loss value. Lp is the mean squared error loss value, and Lp is the soil physics prior knowledge constraint loss value. and The overall weight coefficients of the loss function and + =1, This is the predicted value of soil moisture content. This is the predicted value of soil bulk density. This is a predicted value for soil organic matter content. This is a predicted value for soil cation exchange capacity. , , , The normalization coefficient is... , These are the weight coefficients for the constraint terms.

[0028] Furthermore, the steps of acquiring multi-source soil environmental data of the region to be fused, inputting the multi-source soil environmental data of the region to be fused into the trained multi-source soil environmental data fusion model, and outputting high spatiotemporal resolution soil environmental fusion data and pixel-level uncertainty assessment results include: acquiring original multi-source soil environmental data of the region to be fused; performing spatiotemporal alignment preprocessing on the original multi-source soil environmental data of the region to be fused to obtain spatiotemporally aligned multi-source soil environmental data to be fused; inputting the spatiotemporally aligned multi-source soil environmental data to be fused into the trained multi-source soil environmental data fusion model, adding dropout layers to the last two fully connected layers of the model, and setting the dropout rate to 0.2; performing 50 independent forward inferences, calculating the average of the 50 inference results as the high spatiotemporal resolution soil environmental fusion data, and calculating the standard deviation of the 50 inference results as the uncertainty assessment value of the corresponding pixel; outputting the high spatiotemporal resolution soil environmental fusion data and pixel-level uncertainty assessment results in GeoTIFF format, and generating a corresponding spatial distribution map.

[0029] The beneficial effects of this invention are reflected in the following aspects: This invention fully mines complementary information from multi-source data through a dual-branch feature extraction network, introduces a cross-modal interactive attention mechanism to improve feature fusion quality, constructs a composite loss function with knowledge constraints from the soil science domain to ensure the physical rationality of the prediction results, and employs the Monte Carlo dropout method to achieve pixel-level uncertainty assessment. The newly added dynamic weight adaptive adjustment step of the composite loss function further improves the model's generalization ability, and the pixel-level multi-source uncertainty decomposition step can clearly identify the sources of error, providing targeted guidance for subsequent work. This invention achieves high spatiotemporal resolution soil environmental data fusion with a spatial resolution of 30m and a daily temporal resolution, which can meet the needs of large-scale, refined soil environmental monitoring. Attached Figure Description

[0030] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0031] Figure 1 A flowchart illustrating a deep learning-based multi-source soil environmental data fusion method provided in an embodiment of the present invention;

[0032] Figure 2 This is a schematic diagram of the cross-modal interactive attention fusion module structure provided in an embodiment of the present invention;

[0033] Figure 3A flowchart illustrating the dynamic weight adaptive adjustment of the composite loss function provided in this embodiment of the invention. Detailed Implementation

[0034] To enable those skilled in the art to better understand the present invention, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or end that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or ends.

[0036] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0037] Existing multi-source soil environmental data fusion technologies mostly employ simple weighted averaging or feature splicing strategies, failing to fully explore the intrinsic interaction between static soil properties and dynamic environmental changes, resulting in insufficient utilization of complementary information from multi-source data. Deep learning-based fusion methods driven purely by data lack effective constraints from soil science knowledge, leading to some prediction results that violate fundamental soil physicochemical laws and have unclear physical meanings. Most existing uncertainty assessment methods can only provide global accuracy indicators, failing to quantify uncertainty at the pixel level, and cannot distinguish the uncertainty contributions from different data sources and model structures, severely limiting the application value of fusion results in practical decision-making.

[0038] The present invention aims to provide a deep learning-based multi-source soil environmental data fusion method to solve the problems of insufficient utilization of complementary information in multi-source soil data fusion, pure data-driven prediction violating soil physicochemical laws, and inability to achieve pixel-level multi-dimensional uncertainty quantification in the existing technology.

[0039] like Figure 1 As shown in the figure, a specific embodiment of the present invention provides a multi-source soil environmental data fusion method based on deep learning, including the following steps:

[0040] S1: Collect raw soil environmental data from multiple sources.

[0041] In this embodiment of the invention, the multi-source soil environmental raw data includes static soil property data, dynamic remote sensing monitoring data, ground in-situ sensor data, and laboratory analysis data; the static soil property data includes soil type, soil texture, topographic slope, and altitude; the dynamic remote sensing monitoring data includes Landsat series satellite multispectral data, Sentinel-2 satellite multispectral data, and MODIS vegetation index data; the ground in-situ sensor data includes real-time monitoring data of soil moisture content, soil temperature, and soil electrical conductivity; and the laboratory analysis data includes soil organic matter content, soil pH value, soil cation exchange capacity, and soil heavy metal content data.

[0042] S2: Perform spatiotemporal alignment preprocessing on the original multi-source soil environment data to obtain spatiotemporally aligned multi-source soil environment data.

[0043] In this embodiment of the invention, all original multi-source soil environmental data are uniformly converted into the WGS84 coordinate system and UTM projection; bilinear interpolation is used to resample all data with different spatial resolutions to a uniform 30m×30m spatial resolution; cubic spline interpolation is used to interpolate all data with different temporal resolutions to a uniform daily temporal resolution; K-nearest neighbor interpolation based on spatial Euclidean distance is used to interpolate missing values ​​in all data, with K set to 5; and outliers are identified and removed individually for each feature dimension using the 3σ criterion, resulting in spatiotemporally aligned multi-source soil environmental data.

[0044] S3: Construct a dual-branch feature extraction network. Input spatiotemporally aligned multi-source soil environmental data into the dual-branch feature extraction network to extract static global feature vectors and dynamic spatiotemporal feature tensors respectively.

[0045] In this embodiment of the invention, a static feature extraction branch is constructed by stacking three fully connected layers sequentially, each followed by a ReLU activation function and a batch normalization layer. The number of neurons in the three fully connected layers are 128, 64, and 32 respectively, and the output is a static global feature vector with a dimension of 32. A dynamic spatiotemporal feature extraction branch is constructed by stacking three ConvLSTM layers sequentially, each with a 3×3 kernel size, a stride of 1, and padding of 1, followed by a ReLU activation function and a batch normalization layer, and the output is a dynamic spatiotemporal feature tensor with a dimension of 32×H×W, where H and W are the height and width of the feature map, respectively. The static feature extraction branch and the dynamic spatiotemporal feature extraction branch are connected in parallel to form a dual-branch feature extraction network. The number of neurons in the input layer of the static feature extraction branch is consistent with the dimension of the static soil attribute data. The number of input channels of the dynamic spatiotemporal feature extraction branch is consistent with the total number of bands of the dynamic remote sensing monitoring data and the ground in-situ sensor data.

[0046] S4: As Figure 2 As shown, a cross-modal interactive attention fusion module is constructed. The static global feature vector and the dynamic spatiotemporal feature tensor are input into the cross-modal interactive attention fusion module to perform bidirectional information interaction and generate a multimodal fusion feature vector.

[0047] In this embodiment of the invention, a static feature-guided dynamic attention submodule is constructed. The cosine similarity between the static global feature vector and the features at each time step in the dynamic spatiotemporal feature tensor is calculated. After normalization using the softmax function, dynamic attention weights are obtained. The dynamic attention weights are then multiplied element-wise by the dynamic spatiotemporal feature tensor to obtain weighted dynamic features. A static attention submodule for dynamic feature enhancement is constructed. Global average pooling is performed on the weighted dynamic features to obtain dynamic global features. The cosine similarity between the dynamic global features and the features in each dimension of the static global feature vector is calculated. After normalization using the softmax function, static attention weights are obtained. The static attention weights are then multiplied element-wise by the static global feature vector to obtain weighted static features. A feature concatenation layer is constructed. The global average features of the weighted static features and the weighted dynamic features are concatenated dimensionally to generate a multimodal fusion feature vector with dimension 64. The calculation formula for the static feature-guided dynamic attention submodule is as follows:

[0048] ;

[0049] ;

[0050] in, For the first Dynamic attention weights at each time step. It is a static global feature vector. For the first The dynamic spatiotemporal characteristics of each time step The cosine similarity function is used. This is a weighted dynamic feature.

[0051] S5: Construct a deep learning training framework with knowledge constraints in the soil science domain, input the multimodal fusion feature vector into the deep learning training framework with knowledge constraints in the soil science domain for iterative training, and obtain the trained multi-source soil environment data fusion model.

[0052] In this embodiment of the invention, spatiotemporally aligned multi-source soil environmental data are randomly divided into training, validation, and test sets in a 7:2:1 ratio. All parameters of the deep learning training framework with soil science domain knowledge constraints are initialized using the Adam optimizer, with an initial learning rate of 0.001, a batch size of 32, and a weight decay coefficient of 1e-5. The multimodal fusion feature vectors from the training set are input into the training framework, and iterative training is performed using a composite loss function. Validation is performed on the validation set after each epoch. Training stops when the validation set loss no longer decreases after 10 consecutive epochs, and the current model parameters are saved. The accuracy of the saved model is tested using the test set. When the test set determination coefficient R² is greater than or equal to 0.85, the trained multi-source soil environmental data fusion model is obtained. The formula for calculating the composite loss function is:

[0053] ;

[0054] ;

[0055] in, This is the total loss value. Lp is the mean squared error loss value, and Lp is the soil physics prior knowledge constraint loss value. and The overall weight coefficients of the loss function and + =1, This is the predicted value of soil moisture content. This is the predicted value of soil bulk density. This is a predicted value for soil organic matter content. This is a predicted value for soil cation exchange capacity. , , , The normalization coefficient is... , These are the constraint term weight coefficients; a dynamic weight adaptive adjustment step based on the composite loss function is performed every 5 epochs of training, such as... Figure 3 As shown, calculate the mean squared error loss value Lm_val and the domain knowledge constraint loss value Lp_val on the current validation set. If Lm_val / Lp_val > 2, then... Decrease by 0.05 Increase by 0.05; if Lm_val / Lp_val < 0.5, then... Increase by 0.05 Decrease by 0.05; otherwise, keep the weight unchanged, and always ensure that + =1.

[0056] S6: Obtain multi-source soil environmental data of the area to be fused, input the multi-source soil environmental data of the area to be fused into the trained multi-source soil environmental data fusion model, and output high spatiotemporal resolution soil environmental fusion data and pixel-level uncertainty assessment results.

[0057] In this embodiment of the invention, multi-source soil environmental raw data of the region to be fused are acquired, and spatiotemporal alignment preprocessing is performed on the multi-source soil environmental raw data of the region to be fused to obtain spatiotemporally aligned multi-source soil environmental data to be fused; the spatiotemporally aligned multi-source soil environmental data to be fused is input into the trained multi-source soil environmental data fusion model, and dropout layers are added to the last two fully connected layers of the model with a dropout rate of 0.2; 50 independent forward inferences are performed, and the average value of the 50 inference results is calculated as the high spatiotemporal resolution soil environmental fusion data, and the standard deviation of the 50 inference results is calculated as the uncertainty evaluation value of the corresponding pixel; a pixel-level multi-source uncertainty decomposition step is performed: keeping the model parameters unchanged, 50 forward inferences are performed using noise-free ideal input data, and the standard deviation of the inference results is calculated as the model uncertainty; the total uncertainty is subtracted from the model uncertainty to obtain the data uncertainty; the high spatiotemporal resolution soil environmental fusion data and the pixel-level uncertainty evaluation results are output in GeoTIFF format, and a corresponding spatial distribution map is generated.

[0058] The beneficial effects of this invention are reflected in the following aspects: This invention fully mines complementary information from multi-source data through a dual-branch feature extraction network, introduces a cross-modal interactive attention mechanism to improve feature fusion quality, constructs a composite loss function with knowledge constraints from the soil science domain to ensure the physical rationality of the prediction results, and employs the Monte Carlo dropout method to achieve pixel-level uncertainty assessment. The newly added dynamic weight adaptive adjustment step of the composite loss function further improves the model's generalization ability, and the pixel-level multi-source uncertainty decomposition step can clearly identify the sources of error, providing targeted guidance for subsequent work. This invention achieves high spatiotemporal resolution soil environmental data fusion with a spatial resolution of 30m and a daily temporal resolution, which can meet the needs of large-scale, refined soil environmental monitoring.

[0059] The above descriptions are merely embodiments of the present invention. Commonly known technical solutions or characteristics are not described in detail here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A method for fusing multi-source soil environmental data based on deep learning, characterized in that, The method includes: S1: Collect raw soil environmental data from multiple sources; S2: Perform spatiotemporal alignment preprocessing on the original multi-source soil environment data to obtain spatiotemporally aligned multi-source soil environment data; S3: Construct a dual-branch feature extraction network, input the spatiotemporally aligned multi-source soil environmental data into the dual-branch feature extraction network, and extract the static global feature vector and the dynamic spatiotemporal feature tensor respectively; S4: Construct a cross-modal interactive attention fusion module, input the static global feature vector and the dynamic spatiotemporal feature tensor into the cross-modal interactive attention fusion module, perform bidirectional information interaction, and generate a multimodal fusion feature vector; S5: Construct a deep learning training framework with knowledge constraints in the field of soil science, input the multimodal fusion feature vector into the deep learning training framework with knowledge constraints in the field of soil science for iterative training, and obtain the trained multi-source soil environment data fusion model. S6: Obtain multi-source soil environmental data of the area to be fused, input the multi-source soil environmental data of the area to be fused into the trained multi-source soil environmental data fusion model, and output high spatiotemporal resolution soil environmental fusion data and pixel-level uncertainty assessment results.

2. The method for fusing multi-source soil environmental data based on deep learning according to claim 1, characterized in that: The multi-source soil environment raw data includes static soil property data, dynamic remote sensing monitoring data, ground in-situ sensor data, and laboratory test analysis data; The static soil attribute data includes soil type, soil texture, terrain slope, and altitude; The dynamic remote sensing monitoring data includes Landsat series satellite multispectral data, Sentinel-2 satellite multispectral data, and MODIS vegetation index data; The ground-based in-situ sensor data includes real-time monitoring data of soil moisture content, soil temperature, and soil electrical conductivity. The laboratory analysis data include soil organic matter content, soil pH value, soil cation exchange capacity, and soil heavy metal content.

3. The method for fusing multi-source soil environmental data based on deep learning according to claim 1, characterized in that, The step of performing spatiotemporal alignment preprocessing on the original multi-source soil environmental data to obtain spatiotemporally aligned multi-source soil environmental data includes: All original multi-source soil environmental data were uniformly converted into the WGS84 coordinate system and UTM projection. Bilinear interpolation was used to resample all data with different spatial resolutions to a uniform 30m×30m spatial resolution; The cubic spline interpolation method was used to interpolate all data with different time resolutions to a uniform daily time resolution; Missing values ​​in all data were imputed using the K-nearest neighbor imputation method based on spatial Euclidean distance, with the K value set to 5. Outliers were identified and removed individually for each feature dimension using the 3σ criterion, resulting in spatiotemporally aligned multi-source soil environmental data.

4. The method for fusing multi-source soil environmental data based on deep learning according to claim 1, characterized in that, The step of constructing the dual-branch feature extraction network includes: A static feature extraction branch is constructed, and three fully connected layers are stacked in sequence. Each layer is followed by a ReLU activation function and a batch normalization layer. The number of neurons in the three fully connected layers are 128, 64 and 32 respectively, and the output is a static global feature vector with a dimension of 32. A dynamic spatiotemporal feature extraction branch is constructed, and three ConvLSTM layers are stacked sequentially. Each layer has a 3×3 kernel size, a stride of 1, and padding of 1. It is followed by a ReLU activation function and a batch normalization layer, and the output is a dynamic spatiotemporal feature tensor with a dimension of 32×H×W, where H and W are the height and width of the feature map, respectively. The static feature extraction branch and the dynamic spatiotemporal feature extraction branch are connected in parallel to form a dual-branch feature extraction network.

5. The method for fusing multi-source soil environmental data based on deep learning according to claim 4, characterized in that: The number of neurons in the input layer of the static feature extraction branch is consistent with the dimension of the static soil attribute data; The number of input channels in the dynamic spatiotemporal feature extraction branch is consistent with the total number of bands in the dynamic remote sensing monitoring data and the ground in-situ sensor data.

6. The method for fusing multi-source soil environmental data based on deep learning according to claim 1, characterized in that, The steps for constructing the cross-modal interactive attention fusion module include: A static feature-guided dynamic attention submodule is constructed. The cosine similarity between the static global feature vector and the features at each time step in the dynamic spatiotemporal feature tensor is calculated. The dynamic attention weights are obtained by normalization using the softmax function. The dynamic attention weights are then multiplied element-wise with the dynamic spatiotemporal feature tensor to obtain the weighted dynamic features. A static attention submodule for dynamic feature enhancement is constructed. The weighted dynamic features are subjected to global average pooling to obtain dynamic global features. The cosine similarity between the dynamic global features and the features of each dimension in the static global feature vector is calculated. The static attention weights are obtained by normalization using the softmax function. The static attention weights are multiplied element-wise with the static global feature vector to obtain weighted static features. A feature concatenation layer is constructed to concatenate the weighted static features with the global average features of the weighted dynamic features to generate a multimodal fusion feature vector with a dimension of 64.

7. The method for fusing multi-source soil environmental data based on deep learning according to claim 6, characterized in that, The calculation formula for the static feature-guided dynamic attention submodule is as follows: ； ； in, For the first Dynamic attention weights at each time step. It is a static global feature vector. For the first The dynamic spatiotemporal characteristics of each time step The cosine similarity function is used. This is a weighted dynamic feature.

8. The method for fusing multi-source soil environmental data based on deep learning according to claim 1, characterized in that, The steps for obtaining the trained multi-source soil environmental data fusion model include: The spatiotemporally aligned multi-source soil environmental data were randomly divided into training set, validation set and test set in a ratio of 7:2:

1. Initialize all parameters of the deep learning training framework with soil science domain knowledge constraints, using the Adam optimizer, with an initial learning rate of 0.001, a batch size of 32, and a weight decay coefficient of 1e-5; The multimodal fusion feature vectors in the training set are input into the training framework, and iterative training is performed using a composite loss function. Validation is performed on the validation set after each training epoch. Training stops when the validation set loss no longer decreases after 10 consecutive epochs, and the current model parameters are saved. The accuracy of the saved model is tested using the test set. When the coefficient of determination R² of the test set is greater than or equal to 0.85, the trained multi-source soil environmental data fusion model is obtained.

9. A method for fusing multi-source soil environmental data based on deep learning according to claim 8, characterized in that, The formula for calculating the composite loss function is as follows: ； ； in, This is the total loss value. Lp is the mean squared error loss value, and Lp is the soil physics prior knowledge constraint loss value. and The overall weight coefficients of the loss function and + =1, This is the predicted value of soil moisture content. This is the predicted value of soil bulk density. This is a predicted value for soil organic matter content. This is a predicted value for soil cation exchange capacity. , , , The normalization coefficient is... , These are the weight coefficients for the constraint terms.

10. A method for fusing multi-source soil environmental data based on deep learning according to claim 1, characterized in that, The steps of acquiring multi-source soil environmental data of the region to be fused, inputting the multi-source soil environmental data of the region to be fused into the trained multi-source soil environmental data fusion model, and outputting high spatiotemporal resolution soil environmental fusion data and pixel-level uncertainty assessment results include: Obtain the original multi-source soil environment data of the region to be merged, and perform spatiotemporal alignment preprocessing on the original multi-source soil environment data of the region to be merged to obtain the spatiotemporal aligned multi-source soil environment data to be merged. The spatiotemporally aligned multi-source soil environmental data to be fused is input into the trained multi-source soil environmental data fusion model. Dropout layers are added to the last two fully connected layers of the model, and the dropout rate is set to 0.

2. Perform 50 independent forward inferences, calculate the average of the 50 inference results as high spatiotemporal resolution soil environment fusion data, and calculate the standard deviation of the 50 inference results as the uncertainty assessment value of the corresponding pixel. The high spatiotemporal resolution soil environment fusion data and pixel-level uncertainty assessment results are output in GeoTIFF format, and a corresponding spatial distribution map is generated.

Images