Black soil thickness historical trajectory reconstruction system and method

Abstract

The present application relates to the technical field of soil erosion model prediction, and particularly relates to a black soil thickness historical trajectory reconstruction system and method, a multi-source data time sequence construction module calculates the prior of RUSLE thickness loss according to the black soil thickness at the start and end time and the RUSLE model, constructs the prior sequence of the black soil thickness from the initial time to the end time, and extracts relevant dynamic and static environmental variables; a time sequence token coding module codes the prior sequence of the black soil thickness and the dynamic and static variables step by step, and generates a time sequence coding sequence; a thickness trajectory prediction module processes the coding sequence, outputs the thickness prediction change, synchronously learns the black soil prior correction term and the reliability, and fuses the three to generate a preliminary prediction trajectory of the black soil thickness step by step. A trajectory prediction correction module determines the correction amount according to the thickness at the end of the preliminary trajectory, corrects the whole trajectory, and outputs the final prediction trajectory of the black soil thickness step by step. The present application fuses the loss prior and coding, and accurately predicts the time sequence of the black soil thickness.

Description

Technical Field

[0001] This invention belongs to the field of soil erosion model prediction technology, and in particular relates to a system and method for reconstructing the historical trajectory of black soil thickness. Background Technology

[0002] Existing research on black soil thickness can be broadly categorized into three types: one is static thickness mapping methods based on field profile surveys, statistical interpolation, or empirical regression; another is thickness change estimation methods represented by empirical process erosion models such as RUSLE (Modified General Soil Loss Equation); and the third is machine learning or deep learning reconstruction methods that have gradually emerged in recent years. The first two types of methods have significant limitations in long-term continuous trajectory recovery, complex driving coupling representation, and regional-scale applications; while the latter type of methods, although possessing strong nonlinear fitting capabilities, relies solely on supervision at the beginning and end of the time points, often lacking reliable prior constraints, and is prone to problems such as trajectory drift, distortion of intermediate years, and insufficient interpretability of results.

[0003] In summary, the existing technology has at least the following problems: First, relying solely on empirical process models such as RUSLE usually only provides empirical estimates of soil loss, which often leads to underestimation and ignores widespread gully erosion, freeze-thaw erosion, and wind erosion in Northeast China. This makes it difficult to fully express the nonlinear time-series response of the black soil thickness evolution process, which is affected by both natural and complex human activities. Secondly, when relying solely on deep learning methods, although it is possible to fit the data at the beginning and end points using multiple source variables, the lack of physical priors can easily lead to unstable and discontinuous trajectories in the middle years. Third, many methods only output a single prediction result, lacking multi-level results such as prior trajectory, original trajectory, endpoint correction trajectory, and diagnostic layer, which is not conducive to subsequent mechanism analysis and interpretability assessment. Summary of the Invention

[0004] In view of this, the present invention aims to provide a system and method for reconstructing the historical trajectory of black soil thickness. First, a priori trajectory of annual thickness changes is provided. Then, the temporal nonlinear relationship of multi-source variables is extracted, and adaptive correction priors are determined to determine correction values. Finally, the correction values ​​are redistributed to the entire time trajectory, thereby obtaining a more stable, more interpretable, and closer to the real process temporal reconstruction result of black soil thickness.

[0005] To achieve the above objectives, the technical solution created by this invention is implemented as follows: A system for reconstructing the historical trajectory of black soil thickness includes: a multi-source data time series construction module for acquiring the black soil thickness at the initial and final times, as well as dynamic and static variables related to the black soil thickness; calculating soil loss step by step using the RUSLE model, and then converting the soil loss step by step into a RUSLE thickness loss prior; constructing a black soil thickness prior sequence from the initial time to the final time step by step using the RUSLE thickness loss prior; and a time series token encoding module for performing time-step token encoding on the obtained black soil thickness prior sequence, dynamic variables, and static variables to obtain a time-step encoded sequence. The thickness trajectory prediction module processes the encoded sequence to obtain the predicted thickness change at each time step, and learns the prior correction term, prior confidence weight, and gating coefficient corresponding to the RUSLE thickness loss prior. It then fuses the predicted thickness change, prior correction term, prior confidence weight, and gating coefficient at each time step to obtain the preliminary predicted trajectory of the black soil thickness at each time step. The trajectory prediction correction module determines the endpoint correction amount based on the black soil thickness at the last time step, the black soil thickness at the endpoint time, and prior information in the preliminary predicted trajectory, and allocates the endpoint correction amount to the preliminary predicted trajectory to obtain the final predicted trajectory of the black soil thickness after calibration at each time step.

[0006] Furthermore, the multi-source data time series construction module includes a data preprocessing submodule and a RUSLE prior generation submodule. In the data preprocessing submodule, the input initial time black soil thickness, final time black soil thickness, time-step RUSLE model factors, dynamic variables, and static variables are processed by raster alignment, coordinate unification, invalid value handling, and standardization. In the RUSLE prior generation submodule, soil loss is calculated time-step using RUSLE model factors, and the soil loss is converted into the corresponding time-step RUSLE thickness loss prior using bulk density data. Taking the initial time black soil thickness as the trajectory starting point, the corresponding time-step RUSLE thickness loss prior is subtracted time-step according to the time sequence to obtain the time-step black soil thickness prior sequence from the initial time to the final time.

[0007] Furthermore, the temporal token encoding module includes a token input composition submodule, a token position encoding submodule, and a Transformer encoding submodule. In the token input composition submodule: the prior RUSLE thickness loss for all time steps is summed, and then the proportion of the prior RUSLE thickness loss for each time step to the total prior RUSLE thickness loss is calculated. The token features for the current time step are composed of static variables, dynamic variables, initial black soil thickness, the prior cumulative RUSLE thickness change from the initial time to the current time step, the prior RUSLE thickness loss for the current time step, the loss proportion for the current time step, and the time information for the current time step. In the token position encoding submodule, after linear mapping of the token features for each time step, the processed token features are added to the position encoding for the corresponding time step to obtain the token position encoding for each time step. In the Transformer encoding submodule, the token position encoding for each time step is processed using a Transformer network to obtain a time-step encoded sequence.

[0008] Furthermore, the thickness trajectory prediction module includes a data-driven incremental branch, a residual correction branch, a priori reliability estimation branch, a priori trust weight branch, a gating branch, and a fusion branch; In the data-driven incremental branch: the encoded sequence is processed step-by-step with the following first mapping to obtain the predicted thickness change for each time step: ; in, f represents the predicted thickness change at time step t. data Let h represent the first mapping function. t This represents the encoded feature at time step t in the encoded sequence; In the residual correction branch, the encoded sequence undergoes a second mapping process step-by-step as follows to obtain the prior correction term for the RUSLE thickness loss prior at each time step: ; in, f represents the prior correction term at time step t. corr Indicates the second mapping function; In the prior reliability estimation branch: the magnitude confidence of the total prior RUSLE thickness loss is calculated using the following formula: ; Among them, c mag S represents the magnitude confidence level, σ represents the sigmoid function, and S refThe reference scale represents the total prior amount of RUSLE thickness loss, and a and b represent adjustment parameters; The entropy confidence of the loss proportion over the time distribution is calculated using the following formula: ; Among them, c ent Entropy confidence is used to characterize the degree of concentration of the prior time distribution. Multiplying the entropy confidence score by the amplitude confidence score yields the prior confidence score, i.e.: c prior =c mag ×c ent ; Among them, c prior Indicates prior confidence level; The prior reliability estimate is obtained through the third mapping operation of the following formula: ; Where r represents the prior reliability estimate, f rel H represents the third mapping function. start This indicates the initial thickness of the black soil. Represents the average coding characteristics of the entire time sequence; In the prior trust weight branch: for the encoded features at each time step in the encoded sequence, the prior RUSLE thickness loss at each time step, the prior cumulative RUSLE thickness change from the initial time to the current time step, and the loss ratio at each time step, the following time-step-wise fourth mapping process is performed to obtain the prior trust weight for the prior RUSLE thickness loss at each time step: ; in, Let D represent the prior trust weight at time step t. t Let C represent the prior Rusle thickness loss at time step t. t Let q represent the prior cumulative Rusle thickness change at time step t. t f represents the loss percentage at time step t. trust This represents the fourth mapping function; In the gated branch, the encoding features at each time step in the encoded sequence, the prior Rusle thickness loss at each time step, and the loss ratio at each time step are subjected to the following time-step-wise fifth mapping process to obtain the gate coefficients of the prior correction term: g t =σ(f gate (h t D t ,q t )); Among them, gt f represents the gating coefficient at time step t. gate Represents the fifth mapping function; In the fusion branch, the thickness change at each time step after correction is calculated using the following formula: ; in, This represents the thickness change at the t-th time step after correction. The preliminary predicted trajectory is obtained using the following formula: ; in, This represents the thickness of the black soil at time step t in the preliminary predicted trajectory. This represents the final thickness change at the k-th time step.

[0009] Furthermore, in the trajectory prediction and correction module: the prior Rusle thickness loss for all time steps is subtracted step-by-step from the initial time black soil thickness according to the following formula to obtain the prior end-time black soil thickness: ; Among them, H prior The term represents the prior black soil thickness at the endpoint time. Based on the full-time average coding features, the total prior RUSLE thickness loss, prior confidence, prior reliability estimate, endpoint black soil thickness, prior endpoint black soil thickness, the black soil thickness at the last time step in the preliminary predicted trajectory, and the initial black soil thickness, the endpoint correction mapping is performed using the following formula to obtain the endpoint correction amount: ; Where e represents the endpoint correction amount, f end Represents the endpoint correction mapping function. H represents the thickness of the black soil at the last time step in the preliminary predicted trajectory. end Indicates the thickness of the black soil at the end of the time; The cumulative normalized weight for adding the prediction correction e to the black soil thickness in the initial prediction trajectory is determined by the following formula: ; Where, ω t α represents the cumulative normalized weight at time t; α represents the influence coefficient of the prior cumulative Rusle thickness change in the weight calculation. The final predicted trajectory is obtained using the following formula: ; in, This represents the thickness of the black soil at time step t in the final predicted trajectory.

[0010] Furthermore, the first mapping function, the second mapping function, the third mapping function, the fourth mapping function, the fifth mapping function, and the endpoint correction mapping function are all multilayer perceptron mapping functions; the multilayer perceptron mapping function includes at least one cascaded fully connected layer and a nonlinear activation function.

[0011] A method for reconstructing the historical trajectory of black soil thickness includes: S1: Obtain the initial time black soil thickness data, the final time black soil thickness data, the time-step RUSLE model factors, dynamic variables, and static variables, and perform spatial alignment and preprocessing on the data; S2: Construct a historical trajectory reconstruction system for black soil thickness as provided in this invention; S3: Using the initial time black soil thickness data in step S1 as the trajectory starting point and the end time black soil thickness data as the endpoint supervision constraint, combined with the time-step RUSLE model factors, dynamic variables and static variables, the reconstructed system obtained in step S2 is trained to obtain the reconstructed model. S4: Input the black soil thickness to be analyzed into the reconstruction model obtained in step S3 to obtain the predicted black soil thickness trajectory with the black soil thickness to be analyzed as the time starting point.

[0012] Furthermore, in step S3, the model is trained using the joint loss function of the following formula: ; Among them, L total L represents the joint loss function. raw L represents the original terminal loss. cal L represents the corrected end loss. smooth Smoothing loss based on the rate of change of adjacent years ， L corr Represents the residual modified regularized loss ， L wp Represents the prior trust weight constraint loss ， L rel L represents the reliability branch constraint loss. endcorr Indicates endpoint correction regularization loss ， L bias Indicates overall deviation penalty ， L nonneg λ1 to λ9 represent nonnegativity constraints, and λ1 to λ9 represent loss weights. The original terminal loss is: ; in, This represents the mean squared error loss function; The corrected end loss is: ; in, This indicates the thickness of the black soil at the last time step in the final predicted trajectory; The adjacent-year rate of change smoothing loss is used to constrain the difference in thickness rate of change between adjacent time steps, and is as follows: ; Where, d t This represents the time length corresponding to the t-th time step. In this embodiment of the invention, the time step is specifically a year, therefore the time length d t =1; The residual correction regularization loss is used to constrain the magnitude of the prior correction term, as follows: ; The prior trust weight constraint loss is used to constrain the consistency between the prior trust weights and the prior reliability, and is as follows: ; Where, r target This represents the prior reliability target value determined based on the error between the RUSLE direct prior end thickness and the black soil thickness at the end time. The reliability branch constraint loss can be expressed as: ; The endpoint correction regularization loss is used to constrain the magnitude of the endpoint correction, and is as follows: L endcorr =e 2 ; The overall deviation penalty can be expressed as: ; Where N represents the total number of samples, H end,i This represents the thickness of the black soil at the last time step in the i-th sample. This represents the thickness of the black soil at the last time step in the final predicted trajectory of the i-th sample. Nonnegativity constraints are used to suppress unreasonable negative thickness values, as follows: .

[0013] Compared with the prior art, the present invention can achieve the following beneficial effects: This invention creates a system and method for reconstructing the historical trajectory of black soil thickness. By converting the annual RUSLE results into annual thickness increment priors, it avoids the complete free drift of pure deep learning under limited supervision at the beginning and end. It constructs prior credibility by jointly using the information entropy of the prior total amount and the prior time distribution, and then dynamically adjusts the fusion ratio of physical priors and data-driven results by combining prior trust weights and reliability branches, avoiding mechanical dependence on low-quality priors. Through the combined action of data-driven increments, residual correction, and gating mechanisms, the model can capture nonlinear thickness changes caused by factors such as topography, climate, vegetation, and cultivation intensity while preserving physical rationality. Through endpoint correction heads and cumulative weight backpropagation mechanisms, the thickness error of the endpoint year is reasonably distributed across the entire historical trajectory, thereby balancing the accuracy of the endpoint with the smoothness and consistency of the trajectory in intermediate years. Attached Figure Description

[0014] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments and descriptions of the invention are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 A schematic diagram of the overall system for reconstructing the historical trajectory of black soil thickness as described in an embodiment of the present invention; Figure 2 A schematic diagram of the data processing process of the data preprocessing submodule described in the embodiments of the present invention; Figure 3 A schematic diagram illustrating the feature processing of the timing token encoding module as described in an embodiment of the present invention; Figure 4 A schematic diagram of the data processing process of the thickness trajectory prediction module described in the embodiment of the present invention; Figure 5 A schematic flowchart of the method for reconstructing the historical trajectory of black soil thickness according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the joint loss function described in an embodiment of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not constitute a limitation thereof. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0016] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] like Figure 1 As shown in the embodiment of the present invention, the black soil thickness historical trajectory reconstruction system includes a multi-source data time series construction module, a time series token encoding module, a thickness trajectory prediction module, and a trajectory prediction correction module.

[0018] The multi-source data time series construction module is used to obtain the black soil thickness at the initial and final times, as well as the dynamic and static variables related to the black soil thickness; the soil loss is calculated step by step using the RUSLE model, and then the soil loss at each time step is converted into RUSLE thickness loss prior; using the RUSLE thickness loss prior, the black soil thickness prior sequence from the initial time to the final time is constructed step by step.

[0019] In this invention, static variables are spatially fixed environmental variables used to characterize the background environment. In this embodiment, the static variables are 6-channel variables, including DEM, slope, aspect, K-factor, TWI, and TPI / curvature derivative. Dynamic variables are spatially variable environmental variables used to characterize environmental changes over time. In this embodiment, the dynamic variables are 5-channel variables, including annual R-factor, annual NDVI, annual cultivated land intensity, surface temperature, and average wind speed. Furthermore, in this embodiment, the time step is specifically in years, meaning the black soil thickness sequence includes the black soil thickness for each year from the initial time to the final time.

[0020] In some embodiments, the multi-source data time series construction module includes a data preprocessing submodule and a RUSLE prior generation submodule.

[0021] In the data preprocessing submodule, the initial time black soil thickness, the final time black soil thickness, the time-step RUSLE model factors, dynamic variables, and static variables are processed by raster alignment, coordinate unification, invalid value handling, and standardization.

[0022] In the RUSLE prior generation submodule, the soil loss is calculated step by step using the RUSLE model factor, and the soil loss is converted into the RUSLE thickness loss prior for the corresponding time step using the bulk density data. Taking the initial black soil thickness as the trajectory starting point, the RUSLE thickness loss prior for the corresponding time step is subtracted step by step according to the time sequence to obtain the black soil thickness prior sequence step by step from the initial time to the end time.

[0023] In this embodiment of the invention, the data processing procedure of the data preprocessing submodule is as follows: Figure 2As shown, in the data preprocessing submodule: using the initial time black soil thickness raster as the alignment standard, the final time black soil thickness raster, dynamic variables, and static variables are raster aligned. The preprocessing submodule provides a consistent data foundation for subsequent prior generation and Transformer time series modeling by strictly unifying the spatial framework and numerical representation. In this embodiment, the initial time black soil thickness raster is specifically used as the alignment standard, requiring strict alignment of the number of rows and columns, and coordinate reference systems of the final time black soil thickness raster, dynamic variables, and static variables. In this embodiment, the slope aspect in the static variables is not presented in angular form, but is converted from angle to slope aspect sine and cosine components to avoid the problem that 0° and 360° are numerically discontinuous but spatially aligned. Furthermore, in this embodiment, invalid values ​​in the black soil thickness sequence, dynamic variables, and static variables are uniformly denoted as NODATA=-9999.0, and replaced with NaN during subsequent standardization. After standardization, invalid results are set to 0. The training or inference process is only initiated when the initial black soil thickness, the final black soil thickness, the RUSLE thickness loss prior, the static variables, and the dynamic variables are all valid.

[0024] In this embodiment of the invention, the RUSLE prior generation submodule uses the RUSLE model to calculate the soil loss at each time step based on the rainfall erosivity factor, soil erodibility factor, slope length and gradient factor, vegetation cover factor, and soil and water conservation measure factor. It then combines the soil bulk density or thickness conversion parameters to convert the soil loss into a time-step RUSLE thickness loss prior. The black soil thickness at each time step in the black soil thickness prior sequence is obtained through the loss calculation at each time step.

[0025] The time-series token encoding module performs time-step token encoding on the obtained black soil thickness prior sequence, dynamic variables, and static variables to obtain the time-step encoded sequence.

[0026] In some embodiments, the feature processing procedure of the timing token encoding module is as follows: Figure 3 As shown, it includes a token input composition submodule, a token position encoding submodule, and a Transformer encoding submodule.

[0027] In the token input component submodule: sum the prior RUSLE thickness loss for all time steps, and then calculate the proportion of the prior RUSLE thickness loss for each time step to the total prior RUSLE thickness loss; combine the static variables, dynamic variables, initial black soil thickness, prior RUSLE thickness loss for the current time step, prior cumulative RUSLE thickness change from the initial time to the current time step, the proportion of loss for the current time step, and the time information for the current time step to form the token feature for the current time step.

[0028] Specifically, the process for obtaining the prior total of RUSLE thickness loss is as follows: ; Where S represents the prior total amount of Rusle thickness loss, and L t Let T represent the prior RUSLE thickness loss at time step t, where T represents the total number of time steps. The loss ratio reflects the prior time distribution, and its acquisition process is as follows: ; Where, q t This represents the loss percentage at time step t.

[0029] In this embodiment of the invention, the static variables are 9-channel variables and the dynamic variables are 7-channel variables. The initial black soil thickness, the total prior RUSLE thickness loss, the prior RUSLE thickness loss at the current time step, the loss ratio at the current time step, and the time information at the current time step are all 1-dimensional data. At this time, the token feature of each time step is 21-dimensional data. In addition, the time information at the current time step is the time span from the initial time to the current time step.

[0030] In the token location encoding submodule, after linearly mapping the token features at each time step, the processed token features are added to the location encoding of the corresponding time step to obtain the token location encoding for each time step.

[0031] In this embodiment of the invention, after linearly mapping the token at each time step in the token location encoding submodule, layer normalization, GELU activation, and Dropout operations are performed, and then the result is added to the learnable location encoding to obtain the token location encoding for each time step. The location encoding is a matrix with the same size as the processed token features. "Learnable" means that the location encoding is not a fixed sine code, but rather participates in training as a model parameter matrix and is updated together with the model parameters during training.

[0032] In the Transformer encoding submodule, the token position encoding at each time step is processed using the Transformer network to obtain a time-step-by-time encoded sequence. In this embodiment of the invention, the Transformer network adopts a 4-layer, 8-head multi-head self-attention structure, with a hidden layer dimension of 256, a feedforward network dimension of 1024, and a dropout value of 0.1. The Transformer network uniformly maps the token position encoding at each time step to the same temporal feature space, providing a foundation for subsequent time-step-by-time loss fusion.

[0033] The thickness trajectory prediction module processes the encoded sequence to obtain the predicted thickness change at each time step, and learns the prior correction term, prior confidence weight and gating coefficient corresponding to the RUSLE thickness loss prior. The predicted thickness change, prior correction term, prior confidence weight and gating coefficient are fused according to the time step to obtain the preliminary predicted trajectory of the black soil thickness at each time step.

[0034] In some embodiments, the data processing procedure of the thickness trajectory prediction module is as follows: Figure 4 As shown, it includes a data-driven incremental branch, a residual correction branch, a priori reliability estimation branch, a priori trust weight branch, a gating branch, and a fusion branch.

[0035] In the data-driven incremental branch, the encoded sequence is processed step-by-step by the following first mapping to obtain the predicted thickness change for each time step: ; in, f represents the predicted thickness change at time step t. data Let h represent the first mapping function. t This represents the coding feature at time step t in the coding sequence.

[0036] In the residual correction branch, the encoded sequence undergoes a second mapping process step-by-step as follows to obtain the prior correction term for the RUSLE thickness loss prior at each time step: ; in, f represents the prior correction term at time step t. corr Indicates the second mapping function; In the prior reliability estimation branch, the magnitude confidence of the total prior RUSLE thickness loss is calculated using the following formula: ; Among them, c mag S represents the magnitude confidence level, σ represents the sigmoid function, and S ref The reference scale representing the total prior amount of RUSLE thickness loss is used to normalize the total prior amount, which is obtained by statistical analysis of the total prior amount in the training samples. b represents the adjustment parameter, ε represents a very small positive number to prevent the denominator from being 0 or the logarithmic term from being invalid, and the magnitude confidence is used to characterize the prior significance reflected by the magnitude of the total prior amount of RUSLE thickness loss. The larger the total prior amount of RUSLE thickness loss, the more obvious the prior of RUSLE thickness loss of the pixels in the raster over the entire time range, and the higher the reference value of its prior information for trajectory reconstruction. The entropy confidence of the loss proportion over the time distribution is calculated using the following formula: ; Among them, c ent Entropy confidence is used to characterize the degree of concentration of the prior time distribution. Multiplying the entropy confidence score by the amplitude confidence score yields the prior confidence score, i.e.: c prior =c mag ×c ent ; Among them, c prior This represents the prior confidence level. The calculation of the prior confidence level enables the system to dynamically distinguish between high-confidence prior pixels and low-confidence prior pixels in the raster. The prior reliability estimate is obtained through the third mapping operation of the following formula: ; Where r represents the prior reliability estimate, f rel H represents the third mapping function. start This indicates the initial thickness of the black soil. This represents the average coding feature across the entire time sequence; the value of r ranges from 0 to 1. The larger the prior reliability estimate r, the more reliable the prior of the RUSLE thickness loss at the sample level; the smaller the prior reliability estimate r, the less the model should rely on the prior of the RUSLE thickness loss. In the prior trust weight branch, the encoding features at each time step in the encoded sequence, the prior RUSLE thickness loss at each time step, the prior cumulative RUSLE thickness change from the initial time to the current time step, and the loss ratio at each time step are processed by the following time-step fourth mapping to obtain the prior trust weight for the RUSLE thickness loss prior at each time step: ; in, Let D represent the prior trust weight at time step t. t Let D represent the prior RUSLE thickness loss at time step t. t ≥0, C t Let q represent the prior cumulative Rusle thickness change at time step t. t f represents the loss percentage at time step t. trust This represents the fourth mapping function; In the gated branch, the encoding features at each time step in the encoded sequence, the prior Rusle thickness loss at each time step, and the loss ratio at each time step are subjected to the following time-step-wise fifth mapping process to obtain the gate coefficients of the prior correction term: g t =σ(f gate (ht D t ,q t )); Among them, g t f represents the gating coefficient at time step t. gate Represents the fifth mapping function; In the fusion branch, the thickness change at each time step after correction is calculated using the following formula: ; in, This represents the thickness change at the t-th time step after correction. The preliminary predicted trajectory is obtained using the following formula: ; in, This represents the thickness of the black soil at time step t in the preliminary predicted trajectory. This represents the thickness loss at the k-th time step.

[0037] The trajectory prediction and correction module determines the endpoint correction amount based on the black soil thickness at the last time step in the preliminary predicted trajectory, the black soil thickness at the end time, and prior information, and allocates the endpoint correction amount to the preliminary predicted trajectory to obtain the final predicted trajectory after time step calibration of the black soil thickness.

[0038] In some embodiments, in the trajectory prediction and correction module, the prior Rusle thickness loss for all time steps is subtracted step-by-step from the initial time-based black soil thickness according to the following formula to obtain the prior final time-based black soil thickness: ; Among them, H prior The term represents the prior black soil thickness at the endpoint time. Based on the full-time average coding features, the total prior RUSLE thickness loss, prior confidence, prior reliability estimate, endpoint black soil thickness, prior endpoint black soil thickness, the black soil thickness at the last time step in the preliminary predicted trajectory, and the initial black soil thickness, the endpoint correction mapping is performed using the following formula to obtain the endpoint correction amount: ; Where e represents the endpoint correction amount, f end Represents the endpoint correction mapping function. H represents the thickness of the black soil at the last time step in the preliminary predicted trajectory. end Indicates the thickness of the black soil at the end of the time; The cumulative normalized weight for adding the prediction correction e to the black soil thickness in the initial prediction trajectory is determined by the following formula: ; Where, ω tD represents the cumulative normalized weight at time t, α represents the influence coefficient of the prior cumulative Rusle thickness change in the weight calculation, and D k This represents the prior RUSLE thickness loss at the k-th time step; The final predicted trajectory is obtained using the following formula: ; in, This represents the thickness of the black soil at time step t in the final predicted trajectory.

[0039] In some embodiments, the first mapping function, the second mapping function, the third mapping function, the fourth mapping function, the fifth mapping function, and the endpoint correction mapping function are all multilayer perceptron mapping functions; the multilayer perceptron mapping function includes at least one cascaded fully connected layer and one nonlinear activation function. In the embodiments of the present invention, the multilayer perceptron mapping function is a cascaded fully connected layer and one nonlinear activation function.

[0040] Based on the proposed black soil thickness historical trajectory reconstruction system, this invention also provides a method for reconstructing the black soil thickness historical trajectory, such as... Figure 5 As shown, it includes: S1: Obtain the initial time black soil thickness data, the final time black soil thickness data, time-step RUSLE model factors, dynamic variables, and static variables, and perform spatial alignment and preprocessing on the data. In this embodiment of the invention, black soil thickness sequence data from the paper "The thickness and variation rate of mollic epipedon in the drylands of northern China over the past 40 years" is obtained. The black soil thickness sequence includes year-by-year black soil thickness data for a specific region. Furthermore, a combination of bank sampling and weighted sampling is used. During sample sampling, it is preferable to first subtract the initial time black soil thickness from the final time black soil thickness in the sequence. Since both the initial time black soil thickness and the final time black soil thickness are presented in raster form, to balance computational efficiency and sample representativeness, this embodiment of the invention uses a combination of bank sampling and weighted sampling. First, a certain number of pixels are randomly selected from the effective pixels to form a training sample library. Then, the pixel samples are stratified according to the difference between the black soil thickness at the initial time and the black soil thickness at the end time, so that pixels with large thickness variations, small thickness variations, and near-stable thickness variations can all be sampled. Furthermore, a portion of the samples are drawn according to the original natural distribution, while another portion is drawn according to a stratified balanced distribution. This reduces the impact of uneven thickness variations on model training and enhances the model's ability to learn from regions with significant thickness variations.

[0041] S2: Construct a historical trajectory reconstruction system for black soil thickness as provided in this invention.

[0042] S3: Using the initial time black soil thickness data in step S1 as the trajectory starting point and the end time black soil thickness data as the endpoint supervision constraint, the reconstructed system obtained in step S2 is trained by combining the time-step RUSLE model factors, dynamic variables and static variables to obtain the reconstructed model.

[0043] In some embodiments, step S3 uses a joint loss function of the following formula (e.g., Figure 6 (As shown) Model training is performed: ; Among them, L total L represents the joint loss function. raw L represents the original terminal loss. cal L represents the corrected end loss. smooth Smoothing loss based on the rate of change of adjacent years ， L corr Represents the residual modified regularized loss ， L wp Represents the prior trust weight constraint loss ， L rel L represents the reliability branch constraint loss. endcorr Indicates endpoint correction regularization loss ， L bias Indicates overall deviation penalty ， L nonneg λ1 represents the non-negativity constraint, and λ1 to λ9 represent the loss weights.

[0044] The original terminal loss is: ; in, This represents the mean squared error loss function; The corrected end loss is: ; in, This indicates the thickness of the black soil at the last time step in the final predicted trajectory; The adjacent-year rate of change smoothing loss is used to constrain the difference in thickness rate of change between adjacent time steps, and is as follows: ; Where, d t This represents the time length corresponding to the t-th time step. In this embodiment of the invention, the time step is specifically a year, therefore the time length d t =1; The residual correction regularization loss is used to constrain the magnitude of the prior correction term, as follows: ; The prior trust weight constraint loss is used to constrain the consistency between the prior trust weights and the prior reliability, and is as follows: ; Where, r target This represents the prior reliability target value determined based on the error between the RUSLE direct prior end thickness and the black soil thickness at the end time. The reliability branch constraint loss can be expressed as: ; The endpoint correction canonical loss is used to constrain the magnitude of the endpoint correction, and can be expressed as: L endcorr =e 2 ; The overall deviation penalty can be expressed as: ; Where N represents the total number of samples, H end,i This represents the thickness of the black soil at the last time step in the i-th sample. This represents the thickness of the black soil at the last time step in the final predicted trajectory of the i-th sample. Nonnegativity constraints are used to suppress unreasonable negative thickness values ​​and can be expressed as: .

[0045] In this embodiment of the invention, the optimizer used during training is AdamW, which is used in conjunction with gradient clipping and a learning rate decay scheduler.

[0046] S4: Input the black soil thickness to be analyzed into the reconstruction model obtained in step S3 to obtain the predicted black soil thickness trajectory with the black soil thickness to be analyzed as the time starting point.

[0047] In this embodiment of the invention, after model training is completed, the trained model parameters, standardized parameters, year sequence, time span, and model structure configuration are saved. During inference, the above parameters and configuration are read, and inference is performed block by block across the entire map according to the grid window to obtain the final predicted trajectory of the calibrated time-step black soil thickness. The output includes the RUSLE direct prior year-by-year trajectory, the original year-by-year trajectory of the model, the endpoint-corrected year-by-year trajectory, and diagnostic layers such as prior confidence, prior confidence weight, gating coefficient, and endpoint correction amount.

[0048] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0049] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A system for reconstructing the historical trajectory of black soil thickness, characterized in that, include: A multi-source data time series construction module is used to obtain the black soil thickness at the initial and final times, as well as dynamic and static variables related to the black soil thickness; The soil loss was calculated step by step using the RUSLE model, and then the step-by-step soil loss was converted into RUSLE thickness loss prior. Using the RUSLE thickness loss prior, a black soil thickness prior sequence from the initial time to the end time was constructed step by step. The time-series token encoding module performs time-step token encoding on the obtained black soil thickness prior sequence, dynamic variables, and static variables to obtain the time-step encoded sequence. The thickness trajectory prediction module processes the encoded sequence to obtain the predicted thickness change at each time step, and learns the prior correction term, prior confidence weight and gating coefficient corresponding to the RUSLE thickness loss prior. The predicted thickness change, prior correction term, prior confidence weight and gating coefficient are fused according to the time step to obtain the preliminary predicted trajectory of the black soil thickness at each time step. The trajectory prediction and correction module determines the endpoint correction amount based on the black soil thickness at the last time step in the preliminary predicted trajectory, the black soil thickness at the end time, and prior information, and allocates the endpoint correction amount to the preliminary predicted trajectory to obtain the final predicted trajectory after time step calibration of the black soil thickness. The thickness trajectory prediction module includes a data-driven incremental branch, a residual correction branch, a priori reliability estimation branch, a priori trust weight branch, a gating branch, and a fusion branch. In the data-driven incremental branch: the encoded sequence is processed step-by-step with the following first mapping to obtain the predicted thickness change for each time step: ； in, f represents the predicted thickness change at time step t. data Let h represent the first mapping function. t This represents the encoded feature at time step t in the encoded sequence; In the residual correction branch, the encoded sequence undergoes a second mapping process step-by-step as follows to obtain the prior correction term for the RUSLE thickness loss prior at each time step: ； in, f represents the prior correction term at time step t. corr Indicates the second mapping function; In the prior reliability estimation branch: the magnitude confidence of the total prior RUSLE thickness loss is calculated using the following formula: ； Among them, c mag S represents the magnitude confidence level, σ represents the sigmoid function, and S ref The reference scale represents the total prior amount of RUSLE thickness loss, b represents the adjustment parameter, ε represents the smallest positive number to prevent the denominator from being zero or the logarithmic term from being invalid, and S represents the total prior amount of RUSLE thickness loss. The entropy confidence of the loss proportion over the time distribution is calculated using the following formula: ； Among them, c ent The entropy confidence level is used to characterize the concentration of the prior time distribution, and T represents the total time steps. Multiplying the entropy confidence score by the amplitude confidence score yields the prior confidence score, i.e.: c prior =c mag ×c ent ; Among them, c prior Indicates prior confidence level; The prior reliability estimate is obtained through the third mapping operation of the following formula: ； Where r represents the prior reliability estimate, f rel H represents the third mapping function. start This indicates the initial thickness of the black soil. Represents the average coding characteristics of the entire time sequence; In the prior trust weight branch: for the encoded features at each time step in the encoded sequence, the prior RUSLE thickness loss at each time step, the prior cumulative RUSLE thickness change from the initial time to the current time step, and the loss ratio at each time step, the following time-step-wise fourth mapping process is performed to obtain the prior trust weight for the prior RUSLE thickness loss at each time step: ； in, Let D represent the prior trust weight at time step t. t Let C represent the prior Rusle thickness loss at time step t. t Let q represent the prior cumulative Rusle thickness change at time step t. t f represents the loss percentage at time step t. trust This represents the fourth mapping function; In the gated branch, the encoding features at each time step in the encoded sequence, the prior Rusle thickness loss at each time step, and the loss ratio at each time step are subjected to the following time-step-wise fifth mapping process to obtain the gate coefficients of the prior correction term: g t =σ(f gate (h t ,D t ,q t )); Among them, g t f represents the gating coefficient at time step t. gate Represents the fifth mapping function; In the fusion branch, the thickness change at each time step after correction is calculated using the following formula: ； in, This represents the thickness change at the t-th time step after correction. The preliminary predicted trajectory is obtained using the following formula: ； in, This represents the thickness of the black soil at time step t in the preliminary predicted trajectory. This represents the final thickness change at the k-th time step.

2. The black soil thickness historical trajectory reconstruction system according to claim 1, characterized in that, The multi-source data time series construction module includes a data preprocessing submodule and a RUSLE prior generation submodule; In the data preprocessing submodule, the initial time black soil thickness, the final time black soil thickness, the time-step RUSLE model factors, dynamic variables and static variables are processed by raster alignment, coordinate unification, invalid value handling and standardization. In the RUSLE prior generation submodule, the soil loss is calculated step by step using the RUSLE model factor, and the soil loss is converted into the RUSLE thickness loss prior for the corresponding time step using the bulk density data. Taking the initial black soil thickness as the trajectory starting point, the RUSLE thickness loss prior for the corresponding time step is subtracted step by step according to the time sequence to obtain the black soil thickness prior sequence step by step from the initial time to the end time.

3. The black soil thickness historical trajectory reconstruction system according to claim 2, characterized in that, The timing token encoding module includes a token input composition submodule, a token position encoding submodule, and a Transformer encoding submodule; In the token input component submodule: sum the prior RUSLE thickness loss for all time steps, and then calculate the proportion of the prior RUSLE thickness loss for each time step to the total prior RUSLE thickness loss; combine the static variables, dynamic variables, initial black soil thickness, prior RUSLE thickness loss for the current time step, prior cumulative RUSLE thickness change from the initial time to the current time step, the proportion of loss for the current time step, and the time information for the current time step to form the token feature for the current time step; In the token location encoding submodule, after linear mapping of the token features at each time step, the processed token features are added to the location encoding of the corresponding time step to obtain the token location encoding for each time step. In the Transformer encoding submodule, the token position encoding at each time step is processed using the Transformer network to obtain the time-step encoded sequence.

4. The black soil thickness historical trajectory reconstruction system according to claim 3, characterized in that, In the trajectory prediction and correction module: The prior Rusle thickness loss for all time steps is subtracted step-by-step from the initial time-based black soil thickness using the following formula to obtain the prior final time-based black soil thickness: ； Among them, H prior The term represents the prior black soil thickness at the endpoint time. Based on the full-time average coding features, the total prior RUSLE thickness loss, prior confidence, prior reliability estimate, endpoint black soil thickness, prior endpoint black soil thickness, the black soil thickness at the last time step in the preliminary predicted trajectory, and the initial black soil thickness, the endpoint correction mapping is performed using the following formula to obtain the endpoint correction amount: ； Where e represents the endpoint correction amount, f end Represents the endpoint correction mapping function. H represents the thickness of the black soil at the last time step in the preliminary predicted trajectory. end Indicates the thickness of the black soil at the end of the time; The cumulative normalized weight for adding the prediction correction e to the black soil thickness in the initial prediction trajectory is determined by the following formula: ； Where, ω t denoted as the cumulative normalized weight at time t, and α represents the influence coefficient of the prior cumulative Rusle thickness change in the weight calculation; The final predicted trajectory is obtained using the following formula: ； in, This represents the thickness of the black soil at time step t in the final predicted trajectory.

5. The black soil thickness historical trajectory reconstruction system according to claim 4, characterized in that, The first mapping function, the second mapping function, the third mapping function, the fourth mapping function, the fifth mapping function, and the endpoint correction mapping function are all multilayer perceptron mapping functions; the multilayer perceptron mapping function includes at least one cascaded fully connected layer and a nonlinear activation function.

6. A method for reconstructing the historical trajectory of black soil thickness, characterized in that, include: S1: Obtain the initial time black soil thickness data, the final time black soil thickness data, the time-step RUSLE model factors, dynamic variables, and static variables, and perform spatial alignment and preprocessing on the data; S2: Construct the historical trajectory reconstruction system for black soil thickness as described in claim 4 or 5; S3: Using the initial time black soil thickness data in step S1 as the trajectory starting point and the end time black soil thickness data as the endpoint supervision constraint, combined with the time-step RUSLE model factors, dynamic variables and static variables, the reconstructed system obtained in step S2 is trained to obtain the reconstructed model. S4: Input the black soil thickness to be analyzed into the reconstruction model obtained in step S3 to obtain the predicted black soil thickness trajectory with the black soil thickness to be analyzed as the time starting point.

7. The method for reconstructing the historical trajectory of black soil thickness according to claim 6, characterized in that, In step S3, the model is trained using the joint loss function of the following formula: ； Among them, L total L represents the joint loss function. raw L represents the original terminal loss. cal L represents the corrected end loss. smooth L represents the smoothing loss based on the rate of change between adjacent years. corr L represents the residual modified regularized loss. wp L represents the prior trust weight constraint loss. rel L represents the reliability branch constraint loss. endcorr L represents the endpoint correction regularization loss. bias L represents the overall deviation penalty. nonneg λ1 to λ9 represent nonnegativity constraints, and λ1 to λ9 represent loss weights. The original terminal loss is: ； in, This represents the mean squared error loss function; The corrected end loss is: ； in, This indicates the thickness of the black soil at the last time step in the final predicted trajectory; The adjacent-year rate of change smoothing loss is used to constrain the difference in thickness rate of change between adjacent time steps, and is as follows: ； Where, d t This represents the time length corresponding to the t-th time step, where the time step is specifically in years; therefore, the time length d is... t =1; The residual correction regularization loss is used to constrain the magnitude of the prior correction term, as follows: ； The prior trust weight constraint loss is used to constrain the consistency between the prior trust weights and the prior reliability, and is as follows: ； Where, r target This represents the prior reliability target value determined based on the error between the RUSLE direct prior end thickness and the black soil thickness at the end time. The reliability branch constraint loss is expressed as: ； The endpoint correction regularization loss is used to constrain the magnitude of the endpoint correction, and is as follows: THE endcorr =and 2 ; The overall deviation penalty is expressed as: ； Where N represents the total number of samples, H end,i This represents the thickness of the black soil at the last time step in the i-th sample. This represents the thickness of the black soil at the last time step in the final predicted trajectory of the i-th sample. Nonnegativity constraints are used to suppress unreasonable negative thickness values, as follows: 。

Images