A method, apparatus, computer equipment, and medium for downscaling soil physicochemical properties

Abstract

This application provides a method, apparatus, computer equipment, and medium for downscaling soil physicochemical properties, relating to the field of data processing. The method includes: preprocessing first environmental covariate data related to soil physicochemical properties and using it as an input variable; using soil physicochemical property data interpolated from a second soil census as labels to form a dataset; constructing a deep learning downscaling model that integrates local and global features; training the constructed deep learning downscaling model based on the constructed dataset to obtain a trained target deep learning downscaling model; and inputting second environmental covariate data into the trained target deep learning downscaling model to generate downscaled soil physicochemical property data.

Description

Technical Field

[0001] This application relates to the field of data processing, and in particular to a method, apparatus, computer equipment, and readable storage medium for downscaling soil physicochemical properties. Background Technology

[0002] Soil physicochemical property downscaling methods in related technologies typically rely on traditional statistical regression, geostatistical methods, or machine learning models. These methods often fail to effectively preserve spatial details or effectively fuse information at different scales when processing high-resolution environmental covariates and soil physicochemical property data. Therefore, existing technologies have significant limitations in high-resolution soil physicochemical property prediction and generalization ability under complex environmental conditions. Summary of the Invention

[0003] In view of this, this application provides a method, apparatus, computer equipment, and readable storage medium for downscaling soil physicochemical properties.

[0004] In a first aspect, embodiments of this application provide a method for downscaling soil physicochemical properties, including:

[0005] The data of the first environmental covariate related to soil physicochemical properties were preprocessed and used as input variables, and the data of soil physicochemical properties interpolated from the second soil survey were used as labels to form a dataset.

[0006] Build a deep learning downscaling model that integrates local and global features;

[0007] The deep learning downscaling model is trained based on the constructed dataset to obtain the trained target deep learning downscaling model.

[0008] The second environmental covariate data is input into the trained target deep learning downscaling model to generate downscaled data of soil physicochemical properties.

[0009] The method described in the embodiments of this application may also have the following additional technical features:

[0010] Optionally, in the above technical solution, the data of the first environmental covariate related to the soil physicochemical properties are preprocessed, including:

[0011] The first environmental covariate data related to soil physicochemical properties at different resolutions were resampled to the same resolution;

[0012] Normalize the first environmental covariate data at the same resolution;

[0013] The primary environmental covariate data related to soil physicochemical properties include at least one of the following: climate data, parent material data, topographic data, vegetation data, and human activity data.

[0014] In any of the above technical solutions, optionally, the trained target deep learning downscaling model includes a local feature extraction unit, a global feature extraction unit, and a feature concatenation unit;

[0015] The local feature extraction unit is used to capture the spatial detail features of the second environmental covariate data and obtain the local features of the second environmental covariate data.

[0016] The global feature extraction unit is used to calculate the relationship between any two points in the second environmental covariate data, capture global context information, obtain global features of the second environmental covariate data, and perform position and channel enhancement processing on the global features.

[0017] The feature stitching unit is used to stitch together local features and enhanced global features along the channel dimension to obtain downscaled data of soil physicochemical properties.

[0018] In any of the above technical solutions, optionally, the local feature extraction unit includes a feature embedding layer, a ResNet50 residual network, and multiple deconvolution layers, wherein the ResNet50 residual network includes multiple residual layers;

[0019] The second environmental covariate data is mapped into a high-dimensional feature map through a feature embedding layer. The high-dimensional feature map is then passed through multiple residual layers to extract detailed information layer by layer. The detailed information is then deconvolved through a deconvolution layer to increase the resolution of the detailed information and output local features.

[0020] In any of the above technical solutions, optionally, each residual layer includes multiple residual blocks, and each residual block includes multiple convolutional layers, batch normalization units, and ReLU activation function units;

[0021] Each deconvolutional layer includes a transposed convolutional layer, a ReLU activation function unit, and a batch normalization unit.

[0022] In any of the above technical solutions, optionally, the global feature extraction unit includes a Non-Local unit and a DAModule attention mechanism unit; wherein, the Non-Local unit is used to calculate the relationship between any two points in the input feature map, capture global context information, and obtain global features of the second environmental covariate data; the DAModule attention mechanism unit combines positional attention and channel attention to enhance the responsiveness of the spatially important regions and key feature channels of the global features, respectively.

[0023] The feature stitching unit includes a vector stitching unit and a fully connected layer. The vector stitching unit is used to stitch local features and enhanced global features along the channel dimension, while the fully connected layer is used to fuse and reduce the dimensionality of the stitched features to obtain downscaled data of soil physicochemical properties.

[0024] Optionally, in any of the above technical solutions, the method further includes:

[0025] During the training of the constructed deep learning downscaling model based on the built dataset, a Dropout layer is added after the feature embedding layer and the fully connected layer of the deep learning downscaling model to randomly discard some features; and / or,

[0026] The dataset includes a training set and a test set. During the training of the deep learning downscaling model based on the constructed dataset, the model is trained on different training and test sets using cross-validation. The average performance index of the model is calculated, and the performance of the deep learning downscaling model is evaluated based on the average performance index.

[0027] Secondly, embodiments of this application provide a soil physicochemical property downscaling device, comprising:

[0028] The dataset construction module is used to preprocess the first environmental covariate data related to soil physicochemical properties and use it as the input variable, and use the soil physicochemical property data interpolated from the second soil survey as the label to form the dataset;

[0029] The model building module is used to build deep learning downscaling models that integrate local and global features;

[0030] The model training module is used to train the constructed deep learning downscaling model based on the built dataset to obtain the trained target deep learning downscaling model.

[0031] The downscaling module is used to input the second environmental covariate data into the trained target deep learning downscaling model to generate downscaled data of soil physicochemical properties.

[0032] Thirdly, embodiments of this application provide a computer device including a processor and a memory, the memory storing programs or instructions that can run on the processor, the programs or instructions implementing the steps of the method as described in the first aspect when executed by the processor.

[0033] Fourthly, embodiments of this application provide a readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the method as described in the first aspect.

[0034] The soil physicochemical property downscaling method, apparatus, computer equipment, and readable storage medium of this application provide a scheme for downscaling soil physicochemical properties based on a deep learning model that effectively combines local and global features. This scheme is used for high-resolution downscaling of soil physicochemical properties, which can more comprehensively capture the spatial distribution characteristics of soil physicochemical properties and improve the efficiency and accuracy of soil physicochemical property downscaling.

[0035] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0036] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0037] Figure 1 A schematic flowchart of the soil physicochemical property downscaling method according to an embodiment of this application is shown;

[0038] Figure 2 A schematic diagram of the structure of a deep learning downscaling model according to an embodiment of this application is shown;

[0039] Figure 3 A comparison diagram of particle downsizing in an embodiment of this application is shown;

[0040] Figure 4 A comparison diagram of clay particle downscaling in embodiments of this application is shown;

[0041] Figure 5 A comparison diagram of sand particle size reduction in embodiments of this application is shown;

[0042] Figure 6 This application shows a comparison diagram of organic matter downscaling in an embodiment of the present application;

[0043] Figure 7 A structural block diagram of the soil physicochemical property downscaling device according to an embodiment of this application is shown;

[0044] Figure 8 A structural block diagram of a computer device according to an embodiment of this application is shown. Detailed Implementation

[0045] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0046] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0047] The method, apparatus, computer equipment, and readable storage medium for downscaling soil physicochemical properties provided in this application will be described in detail below with reference to the accompanying drawings and through specific embodiments and application scenarios. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0048] This application provides a method for downscaling soil physicochemical properties, such as... Figure 1 As shown, the method includes:

[0049] Step 101: Preprocess the first environmental covariate data related to soil physicochemical properties and use it as the input variable. Use the soil physicochemical property data interpolated from the second soil survey as the label to form a dataset.

[0050] In this step, firstly, environmental covariate data closely related to soil formation, i.e., first environmental covariate data, are collected, such as climate data, parent material data, topographic data, vegetation data, and human activity data. Since these data typically originate from different remote sensing platforms and data products, their spatial resolution and data format may be inconsistent; therefore, these data need to be processed uniformly. In one embodiment of this application, the first environmental covariate data related to soil physicochemical properties are preprocessed, including:

[0051] The first environmental covariate data related to soil physicochemical properties at different resolutions were resampled to the same resolution;

[0052] Normalize the first environmental covariate data at the same resolution.

[0053] In this embodiment, data preprocessing first involves resampling environmental covariate data from different resolutions to the same resolution. The resampling process employs bilinear interpolation to ensure spatial resolution consistency while preserving as much of the original data's characteristics as possible. After resampling, all environmental covariate data undergoes min-max normalization, using the following formula:

[0054]

[0055] in, It is the raw data. It is the smallest value in the dataset. It is the largest value in the dataset. These are normalized data, ranging from 0 to 1. This application eliminates feature imbalances caused by differences in units and numerical ranges by unifying the value range of all environmental covariate data to 0 to 1.

[0056] Then, low-resolution soil physicochemical property data generated by interpolation from the second national soil census were selected as label data. This data, based on interpolation from soil profile sampling points nationwide, accurately reflects the spatial distribution characteristics of soil physicochemical properties. To construct the deep learning dataset, normalized environmental covariate data were used as input variables, paired with the soil physicochemical property data, and cropped to a 7:3 ratio to form the training and test sets.

[0057] Step 102: Build a deep learning downscaling model that integrates local and global features.

[0058] In this step, the core design idea of ​​the deep learning downscaling model proposed in this application is to fuse local and global features to comprehensively capture the detailed characteristics and macroscopic patterns of soil physicochemical properties, generating high-resolution downscaling data. The deep learning downscaling model is as follows: Figure 2 As shown, the system includes a local feature extraction unit, a global feature extraction unit, and a feature concatenation unit. The local feature extraction unit comprises a feature embedding layer, a ResNet50 residual network, and multiple deconvolutional layers. The ResNet50 residual network includes multiple residual layers, each containing multiple residual blocks. Each residual block includes multiple convolutional layers, a batch normalization unit, and a ReLU activation function unit. Each deconvolutional layer includes a transposed convolutional layer, a ReLU activation function unit, and a batch normalization unit. The global feature extraction unit includes a Non-Local unit and a DAModule attention mechanism unit. The feature concatenation unit includes a vector concatenation unit and a fully connected layer.

[0059] In one embodiment, the distinction between local and global features in environmental covariate data analysis is based on spatial scale and information abstraction level. Local features reflect detailed changes within a small area (e.g., from meters to kilometers), possessing high spatial resolution and strong heterogeneity, and are typically extracted from high-resolution data (e.g., remote sensing imagery, DEM). Global features describe macroscopic patterns over a large area (e.g., from regional to global), possessing low spatial resolution and strong homogeneity, and are typically extracted from low-resolution data (e.g., climate zoning, parent material type distribution maps) or statistical summary information. Local features of climate data may include intraday or seasonal variations in temperature / precipitation (e.g., hourly temperature fluctuations), microclimate differences (e.g., urban heat island effect, temperature difference between valleys and ridges), and high-resolution climate model outputs (e.g., precipitation data at a 1-kilometer grid). Global features of climate data may include climate zone divisions (e.g., tropical, temperate, and polar), long-term statistical indicators such as annual mean temperature and annual precipitation, and large-scale climate models (e.g., El Niño phenomenon, monsoon systems). Local features of parent material data can include the particle size distribution of soil parent material (e.g., the proportion of sand, silt, and clay), the influence of local geological structures (e.g., faults, folds) on parent material distribution, and parent material type patches in high-resolution geological maps (e.g., parent material variation within small watersheds). Global features of parent material data can include regional parent material type distribution (e.g., alluvial parent material in the North China Plain, loess parent material in the Loess Plateau), parent material genetic types of large-scale geological units (e.g., platforms, geosynclines), and parent material age and genesis (e.g., Quaternary sediments, bedrock weathering). Local features of topographic data can include topographic detail indicators (e.g., slope, aspect, curvature, topographic moisture index TWI), micro-landforms (e.g., valleys, ridges, terraces), and topographic relief in high-resolution DEMs (e.g., 30-meter grids). Global features of topographic data can include macro-topographic types (e.g., mountains, plains, hills), large-scale topographic trends (e.g., the uplift of the Tibetan Plateau, the subsidence of coastal plains), and watershed boundaries and drainage system classification (e.g., first-order watersheds, second-order watersheds). Local features of vegetation data can include time-series variations in vegetation indices (e.g., seasonal fluctuations in NDVI), local vegetation cover types (e.g., patch distribution of farmland, woodland, and grassland), and vegetation structure in high-resolution remote sensing images (e.g., canopy density and leaf area index (LAI)). Global features of vegetation data can include regional vegetation types (e.g., tropical rainforest, temperate grassland, and tundra), large-scale vegetation zone distribution (e.g., taiga forest and sawanga forest), and vegetation productivity indices (e.g., regional mean of net primary productivity (NPP)). Local features of human activity data can include fine spatiotemporal dynamics of land use change (e.g., urban sprawl and farmland reclamation), local pollution source distribution (e.g., emission hotspots from factories and roads), and the intensity of human activity in high-resolution nighttime light remote sensing (e.g., light intensity in villages and towns).Global characteristics of human activity data can include regional human activity intensity (such as population density and spatial distribution of GDP), large-scale land use types (such as the regional proportion of agricultural areas, industrial areas, and protected areas), and long-term environmental impacts of human activities (such as changes in urbanization rate and forest coverage).

[0060] In one embodiment, replacing the model of this application with other deep learning models can also accomplish the task of high-resolution downscaling of soil physicochemical properties. For example, models such as U-Net, Transformer, and models combined with generative adversarial networks all have powerful feature extraction and image reconstruction capabilities, which can meet the basic requirements of downscaling. These models also perform excellently in extracting multi-scale features and restoring spatial details, and therefore can serve as alternative solutions to achieve similar inventive objectives.

[0061] Step 103: Train the constructed deep learning downscaling model based on the constructed dataset to obtain the trained target deep learning downscaling model.

[0062] In this step, the deep learning downscaling model built in step 102 is trained based on the dataset constructed in step 101 to obtain the target deep learning downscaling model.

[0063] During model training, the previously constructed dataset was used for supervised learning. The input variables were low-resolution environmental covariate data, and the labels were soil physicochemical properties data at the corresponding resolution. To measure the model's prediction error, mean squared error (MSE) was used as the loss function. The formula for calculating MSE is:

[0064] MSE=

[0065] in, It is the actual value. Here, n is the predicted value, and n is the sample size. MSE can effectively measure the deviation between the predicted value and the true value, and guide the optimization process of model parameters.

[0066] In one embodiment of this application, the trained target deep learning downscaling model includes a local feature extraction unit, a global feature extraction unit, and a feature concatenation unit. The following provides a detailed description of each unit:

[0067] (1) Local feature extraction unit: The local feature extraction unit is used to capture the spatial detail features of the second environmental covariate data and obtain the local features of the second environmental covariate data.

[0068] In one embodiment of this application, the local feature extraction unit includes a feature embedding layer, a ResNet50 residual network, and multiple deconvolutional layers. The ResNet50 residual network includes multiple residual layers. The second environmental covariate data is mapped into a high-dimensional feature map through the feature embedding layer. The high-dimensional feature map is then sequentially processed through multiple residual layers to extract detailed information layer by layer. The detailed information is then deconvolved through deconvolutional layers to increase the resolution of the detailed information and output local features. Each residual layer includes multiple residual blocks, and each residual block includes multiple convolutional layers, a batch normalization unit, and a ReLU activation function unit. Each deconvolutional layer includes a transposed convolutional layer, a ReLU activation function unit, and a batch normalization unit.

[0069] The local feature extraction module is built upon the classic ResNet50 residual network and is primarily used to capture the spatial detail features of the input data. The input 2D image first passes through a feature embedding layer, where a convolutional operation maps the raw data into a high-dimensional feature map, enhancing subsequent feature extraction capabilities. Subsequently, the feature map passes through four residual layers of the ResNet50 residual network. Each residual layer consists of multiple residual blocks, which address the gradient vanishing and network degradation problems faced by deep networks through skip connections. Each residual block contains three convolutional layers, combined with batch normalization and ReLU activation functions, enabling the extraction of local texture, edges, and other detailed information layer by layer. To further recover detailed information, the local feature extraction module adds three deconvolutional layers after the ResNet50 output. Deconvolution is an upsampling method that progressively recovers the original resolution or a higher resolution based on the feature map, capturing more details. Each deconvolutional layer typically includes a transposed convolutional layer, along with ReLU activation and batch normalization. Through this layer-by-layer upsampling, the model can effectively reconstruct high-resolution spatial detail features.

[0070] (2) Global feature extraction unit. The global feature extraction unit is used to calculate the relationship between any two points in the second environmental covariate data, capture global context information, obtain global features of the second environmental covariate data, and perform position and channel enhancement processing on the global features.

[0071] In one embodiment of this application, the global feature extraction unit includes a Non-Local unit and a DAModule attention mechanism unit. The Non-Local unit calculates the relationship between any two points in the input feature map, captures global context information, and obtains the global features of the second environmental covariate data. The DAModule attention mechanism unit combines positional attention and channel attention to enhance the responsiveness of spatially important regions and key feature channels of the global features, respectively.

[0072] In the global feature extraction part, the model uses a Non-Local module, which captures global contextual information by calculating the relationship between any two points in the input feature map. Compared to traditional convolutional operations, the Non-Local module can span a larger receptive field and extract the overall structural features of the image. Subsequently, the model further introduces the DAModule attention mechanism, which combines positional attention and channel attention to enhance the responsiveness of spatially important regions and key feature channels, respectively. This dual attention mechanism improves the expressive power of global features, enabling the model to more accurately capture complex spatial dependencies.

[0073] (3) Feature splicing unit: The feature splicing unit is used to splice local features and enhanced global features in the channel dimension to obtain downscaled data of soil physicochemical properties.

[0074] In one embodiment of this application, the feature splicing unit includes a vector splicing unit and a fully connected layer; wherein, the vector splicing unit is used to splice local features and enhanced global features in the channel dimension, and the fully connected layer is used to fuse and reduce the dimension of the spliced ​​features to obtain downscaled data of soil physicochemical properties.

[0075] In this embodiment, after the local and global features are extracted, they are concatenated along the channel dimension. Then, a fully connected layer operation is used to achieve feature fusion and dimensionality reduction. The output feature map retains both detailed information and global patterns, laying the foundation for generating high-resolution soil physicochemical property data.

[0076] In one embodiment of this application, the method further includes: during the training of the constructed deep learning downscaling model based on the constructed dataset, adding a Dropout layer after the feature embedding layer and the fully connected layer of the deep learning downscaling model to randomly discard some features.

[0077] During training, various regularization techniques, such as Dropout and weight decay, were employed to avoid overfitting and improve the model's generalization ability.

[0078] On the one hand, a Dropout layer is added after the feature embedding layer and the fully connected layer to randomly discard some features, reduce the co-adaptation between features, and enhance the generalization ability of the model.

[0079] On the other hand, weight decay can be implemented by adding a penalty term (L2 norm) related to the model weights to the loss function to limit model complexity and prevent overfitting. The weight decay parameter can be set to 0.001. Weight decay can make the model parameters tend to smaller values, enhancing the smoothness of the model.

[0080] To accelerate convergence and improve training efficiency, the AdamW optimizer was introduced, employing a dynamic learning rate adjustment strategy. At each training epoch, the model's performance on the validation set was evaluated, and the trends in loss and accuracy were observed to determine the model's convergence status. Weight decay parameters were passed during the AdamW optimizer configuration initialization.

[0081] In one embodiment, the dynamic learning adjustment strategy uses the ReduceLROnPlateau scheduler, passing the AdamW optimizer as a parameter. The ReduceLROnPlateau scheduler is configured during model training such that the learning rate is halved when the training loss decreases by less than 1e-4 over four consecutive epochs, to help the model converge more stably.

[0082] In one embodiment of this application, the method further includes: the dataset includes a training set and a test set; during the training of the constructed deep learning downscaling model based on the constructed dataset, the model is trained on different training sets and test sets using cross-validation, the average performance index of the model is calculated, and the performance of the deep learning downscaling model is evaluated based on the average performance index of the model.

[0083] In this embodiment, cross-validation is used to further evaluate the robustness of the model. Training is repeated under different training and test set partitions, and the model's average performance metrics, such as root mean square error (RMSE) and coefficient of determination (R²), are calculated. 2 This ensures that the model's performance on the test set is stable and reliable.

[0084] The formula for calculating RMSE is as follows:

[0085] RMSE=

[0086] in, It is the actual value. 'n' is the predicted value, and 'n' is the sample size. RMSE is the square root of the mean squared error (MSE), which measures the square root of the average of the squared differences between the predicted and actual values.

[0087] The calculation formula is as follows:

[0088]

[0089] in, It is the actual value. It is a predicted value. It is the average of all true values, and n is the sample size. This represents the proportion of variance of the dependent variable explained by the model, and its value ranges from 0 to 1. The closer the value is to 1, the better the model fit, meaning a higher proportion of variance is explained by the model. Experiments showed that this model achieved an RMSE of 0.5135. It reached 0.9174.

[0090] Step 104: Input the second environmental covariate data into the trained target deep learning downscaling model to generate downscaling data of soil physicochemical properties.

[0091] In this embodiment, after model training is completed, preprocessed high-resolution second environmental covariate data is input. First, the data is resampled to a uniform high resolution and subjected to min-max normalization to adapt to the model input. Subsequently, the high-resolution second environmental covariate data is used to generate corresponding higher-resolution downscaled data of soil physicochemical properties through a trained target deep learning downscaling model.

[0092] In one embodiment, the high-resolution second environmental covariate data is segmented into 32×32 pixel blocks, and the output soil physicochemical property downscaling data is also segmented into blocks. These blocks are then stitched together to obtain the final high-resolution soil physicochemical property downscaling data. For example, the downscaling result is as follows: Figure 3 , Figure 4 , Figure 5 , Figure 6 As shown, where, Figure 3 (a) in the image is a low-resolution image of the powder particles before downscaling. Figure 3 (b) in the image is a high-resolution image of the powder particles after downscaling. Figure 4 (a) in the image is a low-resolution image of the clay particles before downscaling. Figure 4 (b) in the image is a high-resolution image of the clay particles after downscaling. Figure 5 (a) in the image is a low-resolution image of the sand grains before downscaling. Figure 5 (b) in the image is a high-resolution image after downscaling of the sand grains. Figure 6 (a) in the image is a low-resolution image of organic matter before downscaling. Figure 6 (b) in the image is a high-resolution image of the organic matter after downscaling.

[0093] This application proposes a deep learning-based downscaling method for soil physicochemical properties that combines local and global features. By integrating local features extracted by ResNet50 with global features obtained through the Non-Local module and the DAModule attention mechanism, a novel solution for downscaling soil organic matter is provided. Compared with traditional ensemble machine learning techniques, this method not only improves the accuracy of soil physicochemical property prediction but also enhances the model's ability to simulate complex relationships between environmental covariates.

[0094] As a specific implementation of the above-mentioned method for downscaling soil physicochemical properties, this application provides a soil physicochemical property downscaling device. For example... Figure 7 As shown, the soil physicochemical property downscaling device 700 includes: a dataset construction module 701, a model building module 702, a model training module 703, and a downscaling module 704.

[0095] Among them, the dataset construction module 701 is used to preprocess the first environmental covariate data related to soil physicochemical properties and use it as input variables, and use the soil physicochemical property data interpolated from the second soil survey as labels to form a dataset;

[0096] Model building module 702 is used to build a deep learning downscaling model that integrates local and global features;

[0097] The model training module 703 is used to train the constructed deep learning downscaling model based on the constructed dataset to obtain the trained target deep learning downscaling model.

[0098] The downscaling module 704 is used to input the second environmental covariate data into the trained target deep learning downscaling model to generate downscaled data of soil physicochemical properties.

[0099] Furthermore, the dataset construction module 701 is specifically used for:

[0100] The first environmental covariate data related to soil physicochemical properties at different resolutions were resampled to the same resolution;

[0101] Normalize the first environmental covariate data at the same resolution;

[0102] The primary environmental covariate data related to soil physicochemical properties include at least one of the following: climate data, parent material data, topographic data, vegetation data, and human activity data.

[0103] Furthermore, the trained target deep learning downscaling model includes a local feature extraction unit, a global feature extraction unit, and a feature concatenation unit;

[0104] The local feature extraction unit is used to capture the spatial detail features of the second environmental covariate data and obtain the local features of the second environmental covariate data.

[0105] The global feature extraction unit is used to calculate the relationship between any two points in the second environmental covariate data, capture global context information, obtain global features of the second environmental covariate data, and perform position and channel enhancement processing on the global features.

[0106] The feature stitching unit is used to stitch together local features and enhanced global features along the channel dimension to obtain downscaled data of soil physicochemical properties.

[0107] Furthermore, the local feature extraction unit includes a feature embedding layer, a ResNet50 residual network, and multiple deconvolutional layers. The ResNet50 residual network includes multiple residual layers.

[0108] The second environmental covariate data is mapped into a high-dimensional feature map through a feature embedding layer. The high-dimensional feature map is then passed through multiple residual layers to extract detailed information layer by layer. The detailed information is then deconvolved through a deconvolution layer to increase the resolution of the detailed information and output local features.

[0109] Furthermore, each residual layer includes multiple residual blocks, and each residual block includes multiple convolutional layers, batch normalization units, and ReLU activation function units;

[0110] Each deconvolutional layer includes a transposed convolutional layer, a ReLU activation function unit, and a batch normalization unit.

[0111] Furthermore, the global feature extraction unit includes a Non-Local unit and a DAModule attention mechanism unit; wherein, the Non-Local unit is used to calculate the relationship between any two points in the input feature map, capture global context information, and obtain global features of the second environmental covariate data; the DAModule attention mechanism unit combines positional attention and channel attention to enhance the responsiveness of the spatially important regions and key feature channels of the global features, respectively.

[0112] The feature stitching unit includes a vector stitching unit and a fully connected layer. The vector stitching unit is used to stitch local features and enhanced global features along the channel dimension, while the fully connected layer is used to fuse and reduce the dimensionality of the stitched features to obtain downscaled data of soil physicochemical properties.

[0113] Furthermore, the model training module 703 is also used for:

[0114] During the training of the constructed deep learning downscaling model based on the built dataset, a Dropout layer is added after the feature embedding layer and the fully connected layer of the deep learning downscaling model to randomly discard some features; and / or,

[0115] The dataset includes a training set and a test set. During the training of the deep learning downscaling model based on the constructed dataset, the model is trained on different training and test sets using cross-validation. The average performance index of the model is calculated, and the performance of the deep learning downscaling model is evaluated based on the average performance index.

[0116] The soil physicochemical property downscaling device 700 in this application embodiment can be a computer device or a component within a computer device, such as an integrated circuit or a chip. The computer device can be a terminal or other device besides a terminal. The soil physicochemical property downscaling device 700 provided in this application embodiment can achieve… Figure 1 The various processes implemented in the soil physicochemical property downscaling method implementation example will not be repeated here to avoid repetition.

[0117] This application also provides a computer device, such as... Figure 8 As shown, the computer device 800 includes a processor 801 and a memory 802. The memory 802 stores a program or instruction that can run on the processor 801. When the program or instruction is executed by the processor 801, it implements the various steps of the above-described soil physicochemical property downscaling method embodiment and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0118] The memory 802 can be used to store software programs and various data. The memory 802 may primarily include a first storage area for storing programs or instructions and a second storage area for storing data. The first storage area may store the operating system, application programs or instructions required for at least one function (such as sound playback, image playback, etc.). Furthermore, the memory 802 may include volatile memory or non-volatile memory, or both. The non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct memory bus RAM (DRRAM). The memory 802 in the embodiments of this application includes, but is not limited to, these and any other suitable types of memory.

[0119] Processor 801 may include one or more processing units; optionally, processor 801 integrates an application processor and a modem processor, wherein the application processor mainly handles operations involving the operating system, user interface, and applications, and the modem processor mainly handles wireless communication signals, such as a baseband processor. It is understood that the aforementioned modem processor may also not be integrated into processor 801.

[0120] This application also provides a readable storage medium storing a program or instructions. When the program or instructions are executed by a processor, they implement the various processes of the above-described soil physicochemical property downscaling method embodiments and achieve the same technical effect. To avoid repetition, these will not be described again here.

[0121] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0122] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method of downscaling soil physico-chemical properties, characterized in that, The method comprises the following steps: Preprocessing first environmental covariate data related to soil physical and chemical properties as input variables, and using soil physical and chemical property data interpolated from the second soil survey as labels to form a data set; Building a deep learning dimension reduction model that integrates local features and global features; Training the built deep learning dimension reduction model based on the constructed data set to obtain a trained target deep learning dimension reduction model; Inputting second environmental covariate data into the trained target deep learning dimension reduction model to generate soil physical and chemical property dimension reduction data; The target deep learning dimension reduction model comprises a local feature extraction unit, a global feature extraction unit, and a feature splicing unit; The local feature extraction unit comprises a feature embedding layer, a ResNet50 residual network, and a plurality of deconvolution layers, the ResNet50 residual network comprises a plurality of residual layers, the second environmental covariate data is mapped to a high-dimensional feature map through the feature embedding layer, the high-dimensional feature map is sequentially subjected to layer-by-layer detail information extraction through a plurality of residual layers to obtain detail information, and the detail information is subjected to deconvolution operation through the deconvolution layers to increase the resolution of the detail information and output the local feature; The global feature extraction unit comprises a Non-Local unit and a DAModule attention mechanism unit; wherein the Non-Local unit is used to calculate the relationship between any two points in the input feature map to capture global context information and obtain the global feature of the second environmental covariate data; and the DAModule attention mechanism unit combines position attention and channel attention to respectively enhance the response ability of the spatial important area and the key feature channel of the global feature; The feature splicing unit comprises a vector splicing unit and a fully connected layer; wherein the vector splicing unit is used to splice the local feature and the enhanced global feature in the channel dimension, and the fully connected layer is used to fuse and reduce the dimension of the spliced feature to obtain the soil physical and chemical property dimension reduction data.

2. The method of downscaling soil physico-chemical properties according to claim 1, characterized in that, The preprocessing of the first environmental covariate data related to soil physical and chemical properties comprises the following steps: Resampling the first environmental covariate data related to soil physical and chemical properties of different resolutions to the same resolution; Normalizing the first environmental covariate data of the same resolution; The first environmental covariate data related to soil physical and chemical properties comprises at least one of the following: climate data, parent material data, terrain data, vegetation data, and human activity data.

3. The soil physical and chemical property dimension reduction method according to claim 1, wherein The local feature extraction unit is used to capture the spatial detail features of the second environmental covariate data to obtain the local feature of the second environmental covariate data; The global feature extraction unit is used to calculate the relationship between any two points in the second environmental covariate data to capture global context information and obtain the global feature of the second environmental covariate data, and the global feature is subjected to position and channel enhancement processing. The feature splicing unit is configured to splice the local features and the enhanced global features in a channel dimension to obtain the soil physicochemical property downscaling data.

4. The soil physicochemical property downscaling method of claim 1, wherein, Each of the residual layers comprises a plurality of residual blocks, each of the residual blocks comprises a plurality of convolution layers, a batch normalization unit and a ReLU activation function unit; Each of the deconvolution layers comprises a transpose convolution layer, a ReLU activation function unit and a batch normalization unit.

5. The method of downscaling soil physico-chemical properties according to any one of claims 1 to 4, characterized in that, The method further comprises: In the process of training the built deep learning downscaling model based on the constructed data set, a Dropout layer is added after the feature embedding layer and the fully connected layer of the deep learning downscaling model to randomly discard part of the features; and / or, The data set comprises a training set and a test set; in the process of training the built deep learning downscaling model based on the constructed data set, the model is trained under different training sets and test sets by a cross-validation method, the average performance index of the model is calculated, and the performance of the deep learning downscaling model is evaluated according to the average performance index of the model.

6. A soil physico-chemical property downscaling device, characterized in that, Comprise: a data set construction module configured to pre-process first environmental covariate data related to soil physicochemical properties as input variables, and to construct a data set by taking soil physicochemical property data interpolated from the second soil survey as labels; a model construction module configured to construct a deep learning downscaling model that fuses local features and global features; a model training module configured to train the built deep learning downscaling model based on the constructed data set to obtain a trained target deep learning downscaling model; a downscaling module configured to input second environmental covariate data into the trained target deep learning downscaling model to generate soil physicochemical property downscaling data; The target deep learning downscaling model comprises a local feature extraction unit, a global feature extraction unit and a feature splicing unit. The local feature extraction unit comprises a feature embedding layer, a ResNet50 residual network and a plurality of deconvolution layers, the ResNet50 residual network comprises a plurality of residual layers, the second environmental covariate data is mapped to a high-dimensional feature map through the feature embedding layer, the high-dimensional feature map is sequentially subjected to layer-by-layer detail information extraction through the plurality of residual layers to obtain detail information, and the detail information is subjected to deconvolution operation through the deconvolution layers to increase the resolution of the detail information, and the local features are output; The global feature extraction unit comprises a Non-Local unit and a DAModule attention mechanism unit; wherein the Non-Local unit is configured to calculate the relationship between any two points in the input feature map to capture global context information and obtain global features of the second environmental covariate data; and the DAModule attention mechanism unit combines position attention and channel attention to respectively enhance the response capability of spatial important regions and key feature channels of the global features. The feature splicing unit comprises a vector splicing unit and a full connection layer; wherein the vector splicing unit is used for splicing the local features and the enhanced global features in the channel dimension, and the full connection layer is used for fusing and reducing the dimension of the spliced features to obtain the soil physicochemical property reduced scale data.

7. A computer device, comprising: The application further provides a computer readable storage medium storing a program or instruction, wherein the program or instruction is executed by a processor to implement the steps of the soil physicochemical property reduced scale method according to any one of claims 1 to 5.

8. A readable storage medium, on which a program or instructions are stored, characterized in that, The program or instruction is executed by the processor to implement the steps of the soil physicochemical property reduced scale method according to any one of claims 1 to 5.

Images