An artificial intelligence driven method for predicting soil organic carbon in forestry

Abstract

This invention relates to the technical field of organic carbon prediction and discloses an artificial intelligence-driven method for predicting forestry soil organic carbon. The method includes: constructing spatiotemporally aligned remote sensing environmental features of a regional grid in a forestry area; calculating the temporary litter cover response index of the regional grid; extracting remote sensing covariates highly correlated with the temporary litter cover response index as litter pollution covariates; using an extreme random tree regression model to remove the litter pollution covariates, generating soil organic carbon prediction features and purification correction feature generation rules, and constructing a forestry soil organic carbon prediction model for soil organic carbon prediction. This invention eliminates remote sensing signal pollution caused by litter cover by constructing purification correction feature generation rules, and uses a gated residual structure to achieve adaptive interference suppression and nonlinear mapping of soil organic carbon, improving prediction accuracy and enabling rapid and accurate monitoring of soil organic carbon in large-scale forestry areas.

Description

Technical Field

[0001] This invention relates to the field of organic carbon prediction using big data processing, and more particularly to an artificial intelligence-driven method for predicting organic carbon in forestry soils. Background Technology

[0002] Forest soil organic carbon (SOC) is a core component of the carbon pool in terrestrial ecosystems, playing a crucial role in global carbon cycling, forest ecosystem service function assessment, and carbon sink measurement. In deciduous broad-leaved forests and mixed coniferous and broad-leaved forest ecosystems, the 0-10cm mineral soil layer is the most active interface layer for organic matter input, decomposition and transformation, and microbial activity, and is also the most representative observation layer for forest carbon pool monitoring, climate change response research, and carbon accounting. Traditional forest soil organic carbon acquisition mainly relies on field quadrat sampling and indoor chemical determination. Although the accuracy at single points is reliable, it has limitations such as high cost, long processing time, sparse spatial coverage, and difficulty in high-frequency dynamic updates, failing to meet the needs of large-area, high-resolution, multi-temporal dynamic monitoring and refined management of forest soil organic carbon. With the rapid development of satellite remote sensing, multi-source environmental data fusion, and artificial intelligence algorithms, spatial prediction of soil organic carbon based on remote sensing inversion and machine learning models has become the mainstream technical direction for improving regional monitoring efficiency and achieving large-scale spatial mapping.

[0003] In forest ecosystems, deciduous broad-leaved forests and mixed coniferous and broad-leaved forests have a long-term litter cover layer on the ground. The thickness, moisture content, decomposition degree, and spatial distribution of this layer dynamically change seasonally, significantly obscuring and interfering with remote sensing signals. The reflection and absorption characteristics of optical bands and the backscattering information of microwave bands often preferentially reflect the state of the litter layer in forest areas, rather than the physicochemical properties of the 0-10cm mineral soil itself. Existing prediction methods generally do not identify, quantify, or remove litter interference signals, directly using mixed remote sensing signals for modeling. This leads to an observational mismatch between predicted features and measured labels of organic carbon in mineral soils, making the models prone to learning spurious associations, exhibiting poor stability, and weak cross-temporal generalization ability. Furthermore, existing methods lack standardized background grouping, interference component stripping mechanisms, and adaptive feature purification processes, making it difficult to accurately remove litter pollution covariates. This results in large prediction biases and distorted spatial heterogeneity representations, failing to meet the demand for high-precision, highly robust intelligent prediction of forestry soil organic carbon.

[0004] Existing research on remote sensing prediction of soil organic carbon largely revolves around environmental factor screening, spatial regression modeling, and machine learning fitting, focusing on the statistical correlation between soil organic carbon and variables such as topography, vegetation, climate, and soil type. For example, patent CN119001060B discloses a multi-scale correlation analysis and prediction method for analyzing the spatial variability of soil organic carbon. By collecting soil samples and environmental variables, it conducts Spearman global correlation analysis and geographically weighted local correlation analysis. After eliminating multicollinear variables, multi-scale geographically weighted regression fitting is used to analyze the spatial heterogeneity of soil organic carbon, revealing its complex correlation with environmental variables at multiple spatial scales. This type of method has good applicability in areas with homogeneous surfaces such as farmland and grassland, and can improve the continuity and interpretability of soil organic carbon spatial prediction to a certain extent. However, such methods mostly focus on the global or local statistical relationship between static environmental factors and soil organic carbon, without addressing interference separation and dynamic correction for the unique surface cover structure of forests, thus limiting their applicability under complex forest stand conditions.

[0005] To address this issue, this invention proposes an artificial intelligence-driven method for predicting forestry soil organic carbon, which significantly improves the accuracy and cross-regional generalization ability of soil organic carbon prediction, enabling efficient and accurate monitoring of forestry mineral soil organic carbon across large areas, multiple time phases, and high resolution, providing scientific and reliable technical support for forest carbon sink assessment and ecosystem management. Summary of the Invention

[0006] This invention proposes an AI-driven method for predicting forestry soil organic carbon. Step S1 involves regular rasterization of the forestry area, simultaneously collecting measured data on mineral soil organic carbon, multi-temporal remote sensing data, and forestry environmental data, and performing multi-source feature extraction and spatiotemporal alignment integration. This constructs standardized, spatially and temporally consistent remote sensing environmental features, addressing technical issues such as spatiotemporal misalignment of multi-source data, inconsistent spatial units, and chaotic feature dimensions. Step S2 extracts background response sequences and calculates the temporary cover response index of surface litter, while simultaneously selecting highly correlated remote sensing covariates as litter pollution covariates. This allows for precise location of the source and intensity of remote sensing disturbances caused by litter. Step S3 utilizes extreme random tree... The regression model fits the nonlinear mapping relationship and removes the interference components corresponding to the litter pollution covariate, generating pure soil organic carbon prediction features and solidified purification correction feature generation rules. This can remove temporary interference signals that are not intrinsic to soil, solving the technical problems of remote sensing signal distortion caused by litter pollution and the inability of linear models to fit nonlinear interference. Step S4 constructs a forestry soil organic carbon prediction model that integrates gating and residual structures based on purification prediction features and correction rules, achieving high-precision end-to-end prediction. This can improve the robustness and accuracy of the model, solving the technical problems of traditional models being severely affected by environmental and temporary cover interference, having weak nonlinear fitting ability, and being unable to accurately and dynamically predict forestry soil organic carbon over a large area.

[0007] To achieve the above objectives, this invention provides an artificial intelligence-driven method for predicting forestry soil organic carbon, comprising the following steps: S1: The forestry area is rasterized, and the measured data of mineral soil organic carbon, multi-temporal remote sensing data and forestry environment data of the regional raster are collected. The collected data are processed by feature extraction to form the spatiotemporal aligned remote sensing environment features of the regional raster in the forestry area. S2: Extract the background response sequence from the spatiotemporally aligned remote sensing environmental features, calculate the temporary surface litter cover response index of the regional grid, and extract remote sensing covariates that are highly correlated with the temporary surface litter cover response index from the spatiotemporally aligned remote sensing environmental features as litter pollution covariates. S3: The extreme random tree regression model is used to remove the litter pollution covariate from the temporary litter cover response index and generate soil organic carbon prediction features and purification correction feature generation rules for each regional grid. S4: Based on the soil organic carbon prediction features and the purification correction feature generation rules, construct a forestry soil organic carbon prediction model. Use the forestry soil organic carbon prediction model to receive the spatiotemporally aligned remote sensing environmental features of regional grids in the forestry area, and generate the predicted soil organic carbon values ​​of regional grids in the forestry area.

[0008] As a further improvement of the present invention: Furthermore, in step S1, the forestry area is rasterized, and measured data of mineral soil organic carbon, multi-temporal remote sensing data, and forestry environmental data are collected from the raster area, including: S11: The forestry area is rasterized according to a regular square grid, dividing the forestry area into multiple non-overlapping area grids; S12: Soil sampling points are set up at the center of the area grid. Mineral soil samples are collected periodically at the soil sampling points according to a preset fixed period. The soil organic carbon content of the mineral soil samples is measured as the measured data of mineral soil organic carbon of the area grid. S13: Periodically collect remote sensing images of the area grid in different optical bands and microwave bands according to a preset fixed period, as multi-temporal remote sensing data of the area grid; S14: Collect the digital elevation model of the area grid and the vegetation type in the area grid as the forestry environment data of the area grid.

[0009] Furthermore, step S1 involves feature extraction processing of the collected data to form spatiotemporally aligned remote sensing environmental features of the forestry area using regional raster data. This also includes: S15: Arrange the measured data of mineral soil organic carbon of the area grid in the order of collection time to obtain the measured data sequence of mineral soil organic carbon of the area grid, which is used as the measured organic carbon feature of the area grid. S16: Calculate the average pixel value of the remote sensing images of the region grid in different optical bands, and arrange them according to the acquisition time of the remote sensing images to obtain the sequence of the average pixel value of the region grid in different optical bands, which is used as the optical band feature of the region grid. S17: Calculate the average pixel value of the remote sensing image of the region grid in the microwave band, and arrange them according to the acquisition time of the remote sensing image to obtain the pixel average value sequence of the region grid in the microwave band, which is used as the microwave band feature of the region grid. S18: Based on the digital elevation model of the regional grid, extract the elevation value of the center of the regional grid, calculate the slope feature of the regional grid, classify and encode the vegetation type of the regional grid to obtain the vegetation coding feature of the regional grid, and splice the slope feature and vegetation coding feature of the regional grid as the forestry environment feature of the regional grid. S19: The measured organic carbon features, optical band features, microwave band features, and forestry environment features of the regional grid are spliced ​​together to form the spatiotemporally aligned remote sensing environment features of the regional grid.

[0010] Further, in step S2, the background response sequence in the spatiotemporally aligned remote sensing environmental features is extracted, and the temporary land cover response index of the regional grid is calculated, including: S21: Extract forestry environment features from the spatiotemporally aligned remote sensing environment features, and divide the region grid into multiple background groups based on the forestry environment features; Specifically, the process of dividing the region grid into multiple background groups is as follows: S211: Based on forestry environmental characteristics, calculate the inter-group similarity between any two regional grids. The higher the inter-group similarity, the higher the probability that the two regional grids are classified into the same background group. ; in, Representation of area grid Inter-group similarity This represents an exponential function with the natural constant as its base. These represent the area grids in sequence. The slope characteristics and vegetation coding characteristics, These represent the area grids in sequence. The slope characteristics and vegetation coding characteristics, This represents the XOR operator. Indicates the slope control weight. Indicates the vegetation type penalty coefficient; S212: A background group discrimination method based on intra-group environmental entropy is adopted to generate background group numbers for the region grid, and the region grid is automatically divided into background groups. The discrimination process of the background group discrimination method is as follows: Initialize the number of background groups K and generate an initial feature vector for each background group. Use the initial feature vector as the forestry environment feature of the first region grid within the background group. Based on the inter-group similarity calculation formula, the discriminant index between the region raster to be classified into background groups and any background group is calculated, and the background group code of the region raster is generated: S212: A background group discrimination method based on intra-group environmental entropy is adopted to generate background group numbers for the region grid, and the region grid is automatically divided into background groups. The discrimination process of the background group discrimination method is as follows: Initialize the number of background groups K and generate an initial feature vector for each background group. Use the initial feature vector as the forestry environment feature of the first region grid within the background group. Based on the inter-group similarity calculation formula, the discriminant index between the region raster to be classified into background groups and any background group is calculated, and the background group code of the region raster is generated: ; ; in, Representation of area grid Background group coding, , This represents the set of raster regions for the current s-th background group (including the first region raster). Represents a set of raster regions Any region of the raster, Representation of area grid Inter-group similarity This represents the entropy penalty coefficient. Indicates joining the region grid The environmental entropy of the s-th background group This represents the number of raster cells in the s-th background group, where the lower the environmental entropy, the more homogeneous the vegetation coding features of the raster cells within the background group. This represents the proportion of the vegetation-encoded feature of the c-th plant within the s-th background group. This represents the control parameters, where the first and second vegetation coding features are [1,0] and [0,1] respectively. Represents a logarithmic function with the natural constant as its base; Indicates the selection that makes The minimum value of s is taken as the output, where ; Divide the raster regions with the same background group code into the same background group until all region raster regions have been divided, and then delete the first region raster region in the background group. S22: Calculate the mean optical band features and mean microwave band features of all regional grids in the background group to form the background response sequence of the background group; S23: Based on the background response sequence of the background group to which the regional grid belongs, calculate the difference sequence between the optical band features, microwave band features and background response sequence of the regional grid, and calculate the temporary surface cover response index of the regional grid based on the difference sequence.

[0011] Furthermore, step S2, which extracts remote sensing covariates highly correlated with the temporary cover response index of surface litter from the spatiotemporally aligned remote sensing environmental features as litter pollution covariates, also includes: The optical band features and microwave band features in the spatiotemporally aligned remote sensing environment features are extracted, and the optical band features are split into pixel mean sequences of different optical bands, and the microwave band features are used as the pixel mean sequences of the microwave band. The Pearson correlation coefficient between the pixel mean sequence and the temporary cover response index of surface litter was calculated. Pixel mean sequences with Pearson correlation coefficients higher than a preset similarity threshold were selected as remote sensing covariates that are highly correlated with the temporary cover response index of surface litter. All extracted remote sensing covariates were used to form the litter pollution covariates of the regional raster.

[0012] Furthermore, in step S3, an extreme random tree regression model is used to remove the litter pollution covariate from the temporary litter cover response index, including: S31: The extreme random tree regression model is obtained by parallel integration of multiple independent and unrelated extreme decision trees, wherein the extreme random tree regression model includes an input layer, a decision tree integration layer and an output layer; S32: The input layer receives the surface litter temporary cover response index and litter pollution covariate of the area grid; S33: The extreme decision tree in the decision tree ensemble layer adopts random selection of splitting features and random generation of splitting thresholds, and recursively divides the sample space layer by layer to adaptively fit the nonlinear and non-monotonic variation law of the litter pollution covariate with the change of the temporary litter cover response index. The variation patterns obtained from fitting each extreme decision tree are fused by mean fusion to obtain a global nonlinear mapping function with litter pollution covariate as input; S34: The output layer inputs the litter pollution covariate into the global nonlinear mapping function to obtain the interference component corresponding to the litter pollution covariate, calculates the difference between the temporary litter cover response index and the interference component, and obtains the temporary litter cover response index after removing the litter pollution covariate.

[0013] Furthermore, the rules for generating soil organic carbon prediction features and remediation correction features for each region's raster in step S3 also include: The temporary cover response index of surface litter, after removing litter pollution covariates, is used as the soil organic carbon prediction feature of the regional raster, and the global nonlinear mapping function with litter pollution covariates as input is used as the purification correction feature generation rule of the regional raster.

[0014] Further, in step S4, the forestry soil organic carbon prediction model is used to receive the spatiotemporally aligned remote sensing environmental features of regional grids in the forestry area, and to generate predicted soil organic carbon values ​​for regional grids in the forestry area, including: S41: The forestry soil organic carbon prediction model includes an input layer, a background group identification layer, an exponential variable extraction layer, a purification correction layer, a gated adaptive screening layer, a residual mapping layer, and a regression prediction layer. S42: Spatiotemporal alignment of remote sensing environment characteristics of the input layer receiving area grid; S43: The background group identification layer extracts the forestry environment features from the spatiotemporally aligned remote sensing environment features, and divides the area grid into the corresponding background group based on the forestry environment features; S44: The index variable extraction layer extracts the background response sequence of the background group to which the regional raster belongs, and calculates the temporary cover response index of surface litter and the litter pollution covariate. S45: The purification correction layer uses the purification correction feature generation rules of the regional grid to generate the interference component corresponding to the litter pollution covariate, calculates the difference between the temporary cover response index of the surface litter and the interference component, and uses it as the temporary cover response index of the surface litter after removing the litter pollution covariate. S46: The gated adaptive screening layer receives the surface litter temporary cover response index and interference components after removing the litter pollution covariate. It uses a dual-branch, opposing gating weight method to generate interference gating weights for the residual litter pollution characteristic component and intrinsic gating weights for the effective soil organic carbon characteristic component. The formula for generating the dual-branch, opposing gating weights is as follows: ; ; in, Indicates intrinsic gating weights, Indicates the interference gating weight, This represents the response index of temporary land cover after removing the covariate of land cover pollution. Indicates interference components, This represents the activation function. All represent the trainable convolutional weight matrix parameters in the gated adaptive filtering layer. All of these represent trainable bias parameters in the gated adaptive filtering layer; S47: The residual mapping layer uses the interference gating weights and intrinsic gating weights to perform residual mapping on the surface litter temporary cover response index after removing the litter pollution covariate, obtaining the residual mapping features, wherein the residual mapping formula is: ; in, Represents the residual mapping characteristics. This represents the trainable residual convolution matrix in the residual mapping layer. This represents the trainable residual bias in the residual mapping layer. This represents the element-wise multiplication operator. Indicates the weight control parameters; S48: The regression prediction layer adopts a fully connected layer structure, with residual mapping features as input and soil organic carbon prediction value as output.

[0015] Compared with existing technologies, this invention proposes an artificial intelligence-driven method for predicting forestry soil organic carbon, which has the following beneficial effects: First, this invention calculates inter-group similarity based on slope and vegetation feature encoding, introduces vegetation XOR penalty and environmental entropy constraint to achieve high-precision background group division and ensure that the environment within the same group is homogeneous. Specifically, a background response sequence is constructed by the mean within the group to eliminate the influence of the fixed environmental background. Then, the difference between the band features and the background response sequence is calculated, and the surface litter temporary cover response index is obtained by integrating multi-band offset and temporal jump. This index can sensitively reflect the thickness, coverage and temporal abrupt change of litter, and realize independent characterization of interference signals.

[0016] Meanwhile, this invention constructs a complete prediction model that includes input, background recognition, index extraction, purification correction, gating adaptive screening, residual mapping, and regression prediction. This model achieves high-precision prediction of forestry soil organic carbon over a large area and in a gridded manner, meeting the needs of actual ecological monitoring. Specifically, the gating layer adopts a dual-branch, opposing weight structure to enhance effective soil features and suppress litter residue interference, thereby achieving adaptive feature screening. The residual layer introduces gating weighted residual connections to retain intrinsic information and compensate for nonlinear biases, avoiding gradient vanishing and improving fitting ability and generalization. The forestry soil organic carbon prediction model deeply integrates interference removal and intelligent prediction, solving the problems of traditional models being greatly affected by litter interference, having low accuracy, and poor stability. Attached Figure Description

[0017] Figure 1 A flowchart illustrating an artificial intelligence-driven method for predicting forestry soil organic carbon, provided in an embodiment of the present invention. Figure 2 This is a structural diagram of a forestry soil organic carbon prediction model provided in an embodiment of the present invention; Figure 3 This is an experimental comparison diagram provided for one embodiment of the present invention. Detailed Implementation

[0018] The realization of the objectives, functional characteristics, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] This invention provides an artificial intelligence-driven method for predicting forestry soil organic carbon. The executing entity of this AI-driven forestry soil organic carbon prediction method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this invention: a server, a terminal, etc. In other words, the AI-driven forestry soil organic carbon prediction method can be executed by software or hardware installed on a terminal device or a server device, and the software can be a blockchain platform. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster.

[0020] Reference Figure 1 as well as Figure 2 Embodiment 1 of the present invention is as follows: An artificial intelligence-driven method for predicting forestry soil organic carbon, the method comprising: S1: The forestry area is rasterized, and measured data of mineral soil organic carbon, multi-temporal remote sensing data and forestry environmental data are collected from the regional raster. Feature extraction processing is performed on the collected data to form the spatiotemporally aligned remote sensing environmental features of the regional raster in the forestry area.

[0021] Specifically, in step S1, the forestry area is rasterized, and measured data of mineral soil organic carbon, multi-temporal remote sensing data, and forestry environmental data are collected from the raster area, including: S11: The forestry area is rasterized according to a regular square grid, dividing the forestry area into multiple non-overlapping grid areas; specifically, the grid resolution of the regular square grid is 10 meters. 10 meters; S12: Soil sampling points are set up at the center of the area grid. Mineral soil samples are collected periodically at the soil sampling points according to a preset fixed period. The soil organic carbon content of the mineral soil samples is measured as the measured data of mineral soil organic carbon of the area grid. Specifically, the mineral soil sample is 100 grams of soil at a depth of 10 cm at the soil sampling point, and the soil organic carbon content is determined in a laboratory environment using the potassium dichromate oxidation-external heating method. The potassium dichromate oxidation-external heating method specifically involves: oxidizing the organic carbon in the mineral soil sample with excess potassium dichromate standard solution, then titrating the remaining potassium dichromate with ferrous ammonium sulfate standard solution, and calculating the soil organic carbon content based on the consumption. The preset fixed period is set to 1 day by default; S13: Periodically collect remote sensing images of the area grid in different optical bands and microwave bands according to a preset fixed period, as multi-temporal remote sensing data of the area grid; Specifically, the optical band includes visible light, near-infrared and short-wave infrared bands, which can sensitively reflect the spectral absorption and reflection characteristics of surface materials and are significantly affected by litter cover; the microwave band uses the VH / HV cross-polarized band of Sentinel-1, which has penetrability and sensitivity to surface structure, and can reflect structural information such as surface roughness and cover thickness. S14: Collect the digital elevation model of the area grid and the vegetation type in the area grid as the forestry environment data of the area grid.

[0022] Specifically, the vegetation types include two types: deciduous broad-leaved forests and mixed coniferous and broad-leaved forests.

[0023] Step S1 involves feature extraction processing of the collected data to form spatiotemporally aligned remote sensing environmental features of the forestry area using regional raster data. This also includes: S15: Arrange the measured data of mineral soil organic carbon of the area grid in the order of collection time to obtain the measured data sequence of mineral soil organic carbon of the area grid, which is used as the measured organic carbon feature of the area grid. S16: Calculate the average pixel value of the remote sensing images of the region grid in different optical bands, and arrange them according to the acquisition time of the remote sensing images to obtain the sequence of the average pixel value of the region grid in different optical bands, which is used as the optical band feature of the region grid. S17: Calculate the average pixel value of the remote sensing image of the region grid in the microwave band, and arrange them according to the acquisition time of the remote sensing image to obtain the pixel average value sequence of the region grid in the microwave band, which is used as the microwave band feature of the region grid. S18: Based on the digital elevation model of the regional grid, extract the elevation value of the center of the regional grid, calculate the slope feature of the regional grid, classify and encode the vegetation type of the regional grid to obtain the vegetation coding feature of the regional grid, and splice the slope feature and vegetation coding feature of the regional grid as the forestry environment feature of the regional grid. Specifically, the formula for calculating the slope characteristics of the area grid is: ; in, This represents the slope characteristics of the raster cell in the nth row and mth column of a forestry region. N represents the total number of rows in the forestry region grid obtained by dividing it in the north-south direction, and M represents the total number of columns in the forestry region grid obtained by dividing it in the east-west direction. This indicates the unit raster resolution for the area rasterization process (set to 10 meters). This represents the elevation value of the center of the raster cell in the nth row and m+1th column of the forestry area. This represents the elevation value of the center of the raster cell in the nth row and m-1th column of the forestry area. This represents the elevation value of the center of the raster cell in the (n+1)th row and mth column of the forestry area. This represents the elevation value of the center of the raster cell in the (n-1)th row and mth column of the forestry area. Represents pi; S19: The measured organic carbon features, optical band features, microwave band features, and forestry environment features of the regional grid are spliced ​​together to form the spatiotemporally aligned remote sensing environment features of the regional grid.

[0024] As an embodiment of the present invention, the spatiotemporal aligned remote sensing environmental features of the region grid are represented as follows: ; ; ; ; ; ; in, This represents the spatiotemporal aligned remote sensing environmental characteristics of the raster grid in the nth row and mth column of a forestry area. These are, in order, the spatiotemporal aligned remote sensing environmental features. The measured characteristics of organic carbon, optical band characteristics, microwave band characteristics, and forestry environmental characteristics; These represent the pixel mean sequences for the visible light, near-infrared, and short-wave infrared bands in the optical band features, respectively. This represents the sequence of pixel mean values ​​for a region grid in the microwave band. Indicates forestry environmental characteristics The slope characteristics in Indicates forestry environmental characteristics The vegetation coding features in the data; specifically, the vegetation coding features of deciduous broad-leaved forest are [1,0], and the vegetation coding features of mixed coniferous and broad-leaved forest are [0,1]. Indicates the measured characteristics of organic carbon Measured data of organic carbon in group D mineral soils. Indicates the measured characteristics of organic carbon The measured data of organic carbon in the dth group of mineral soils, where D represents the sequence length; Represents the pixel mean sequence The mean of pixels in group D, Represents the pixel mean sequence The average pixel value of group D in the dataset.

[0025] It should be noted that this invention constructs forestry environmental features by temporally arranging measured organic carbon features, optical / microwave band pixel mean sequences, and splicing slope and vegetation codes, ultimately forming spatiotemporally aligned remote sensing environmental features. This achieves standardized integration of multi-dimensional features. Specifically, slope is precisely calculated using a differential operator, and vegetation type is binary encoded to ensure that environmental features are calculable and comparable. Overall, the constructed spatiotemporally aligned remote sensing environmental features have clear feature dimensions and a unified structure. They retain the true label of soil organic carbon while fully utilizing the sensitivity of optical and microwave sensors to land cover, and incorporating topographic and vegetation base information. This feature construction method achieves effective fusion and noise reduction of high-dimensional information, highlighting the separability of intrinsic soil information and interference information. It provides high-quality input for background group classification, pollution covariate screening, and subsequent remediation prediction, improving the reliability and interpretability of the entire process.

[0026] S2: Extract the background response sequence from the spatiotemporally aligned remote sensing environmental features, and calculate the temporary surface litter cover response index of the regional grid. Extract remote sensing covariates that are highly correlated with the temporary surface litter cover response index from the spatiotemporally aligned remote sensing environmental features as litter pollution covariates.

[0027] Specifically, step S2 involves extracting the background response sequence from the spatiotemporally aligned remote sensing environmental features and calculating the temporary land cover response index of the regional raster, including: S21: Extract forestry environment features from the spatiotemporally aligned remote sensing environment features, and divide the region grid into multiple background groups based on the forestry environment features; Specifically, the process of dividing the region grid into multiple background groups is as follows: S211: Based on forestry environmental characteristics, calculate the inter-group similarity between any two regional grids. The higher the inter-group similarity, the higher the probability that the two regional grids are classified into the same background group. ; in, Representation of area grid Inter-group similarity This represents an exponential function with the natural constant as its base. These represent the area grids in sequence. The slope characteristics and vegetation coding characteristics, These represent the area grids in sequence. The slope characteristics and vegetation coding characteristics, This represents the XOR operator. This indicates the slope control weight (default setting is 0.1). This represents the vegetation type penalty coefficient (default setting is 100). S212: A background group discrimination method based on intra-group environmental entropy is adopted to generate background group numbers for the region grid, and the region grid is automatically divided into background groups. The discrimination process of the background group discrimination method is as follows: Initialize the number of background groups K and generate an initial feature vector for each background group. Use the initial feature vector as the forestry environment feature of the first area grid in the background group. The default number of background groups is 4. The initial feature vectors of the 1st to 4th groups are [5,1,0], [25,1,0], [5,0,1], and [25,0,1], respectively. Based on the inter-group similarity calculation formula, the discriminant index between the region raster to be classified into background groups and any background group is calculated, and the background group code of the region raster is generated: ; ; in, Representation of area grid Background group coding, , This represents the set of raster regions for the current s-th background group (including the first region raster). Represents a set of raster regions Any region of the raster, Representation of area grid Inter-group similarity This represents the entropy penalty coefficient (default setting is 0.05). Indicates joining the region grid The environmental entropy of the s-th background group This represents the number of raster cells in the s-th background group, where the lower the environmental entropy, the more homogeneous the vegetation coding features of the raster cells within the background group. This represents the proportion of the vegetation-encoded feature of the c-th plant within the s-th background group. This represents the control parameter (default setting is 0.0001), where the first and second vegetation encoding features are [1,0] and [0,1] respectively. Represents a logarithmic function with the natural constant as its base; Indicates the selection that makes The minimum value of s is taken as the output, where ; Divide the raster regions with the same background group code into the same background group until all region raster regions have been divided, and then delete the first region raster region in the background group. S22: Calculate the mean optical band features and mean microwave band features of all regional grids in the background group to form the background response sequence of the background group; S23: Based on the background response sequence of the background group to which the regional grid belongs, calculate the difference sequence between the optical band features, microwave band features and background response sequence of the regional grid, and calculate the temporary surface cover response index of the regional grid based on the difference sequence.

[0028] As an embodiment of the present invention, the background response sequence includes the mean values ​​of the pixel mean values ​​of all regions within the background group in the visible light, near infrared, short-wave infrared, and microwave bands, and are respectively represented as the visible light background sequence, near infrared background sequence, short-wave infrared band background sequence, and microwave band background sequence. The pixel mean sequences of the visible light, near-infrared, short-wave infrared, and microwave bands in the region grid are calculated respectively, and the sequence differences between them and the visible light background sequence, near-infrared background sequence, short-wave infrared background sequence, and microwave background sequence of the background group of the region grid are calculated respectively, to form visible light difference sequences, near-infrared difference sequences, short-wave infrared difference sequences, and microwave difference sequences. Specifically, the difference between the pixel mean sequence of visible light and the visible light background sequence is calculated to form a visible light difference sequence of length D. Specifically, the formula for calculating the temporary surface cover response index of the area grid is as follows: ; ; ; in, This represents the temporary land cover response index of the nth row and mth column of the regional raster in a forestry area. Indicator of Temporary Cover Response Index for Ground Litter The response exponential component corresponds to the data collected in the d-th period. These represent the d-th sequence values ​​of the visible light difference sequence, near-infrared difference sequence, short-wave infrared band difference sequence, and microwave band difference sequence of the regional raster in the n-th row and m-th column of the forestry area, respectively. This represents the adjacent time-series jump variable of the nth row and mth column raster in the forestry region during the dth period. This represents the sequence value of the raster cell in the nth row and mth column of the forestry area during the (d-1)th period. .

[0029] Furthermore, the background grouping method and environmental entropy constraints ensure the consistency of the environment within the same group, and the difference calculation effectively removes the contribution of the static environment, so that the index only responds to temporary changes in litter, providing an accurate benchmark for subsequent screening of pollution covariates and removal of interference, and improving the pollution separation effect.

[0030] Step S2, which extracts remote sensing covariates highly correlated with the temporary cover response index of surface litter from the spatiotemporally aligned remote sensing environmental features as litter pollution covariates, also includes: The optical band features and microwave band features in the spatiotemporally aligned remote sensing environment features are extracted, and the optical band features are split into pixel mean sequences of different optical bands, and the microwave band features are used as the pixel mean sequences of the microwave band. The Pearson correlation coefficient between the pixel mean sequence and the temporary cover response index of surface litter was calculated. Pixel mean sequences with Pearson correlation coefficients higher than a preset similarity threshold were selected as remote sensing covariates that are highly correlated with the temporary cover response index of surface litter. All extracted remote sensing covariates were used to form the litter pollution covariates of the regional raster.

[0031] It should be noted that this invention extracts pixel mean sequences from optical and microwave bands, and uses Pearson correlation coefficients to screen sequences highly correlated with the temporary cover response index of surface litter as litter pollution covariates, thereby achieving precise localization of interference features. The correlation coefficient screening objectively quantifies the correlation strength between band features and litter interference, ensuring the purity and effectiveness of covariates. Optical multi-band and microwave band complement each other, comprehensively covering spectral and structural interference sources. Therefore, this screening method avoids feature redundancy and interference from irrelevant variables, accurately identifying the key feature components that cause remote sensing signal pollution.

[0032] S3: The extreme random tree regression model is used to remove the litter pollution covariate from the temporary litter cover response index and generate soil organic carbon prediction features and purification correction feature generation rules for each regional grid.

[0033] Specifically, step S3 uses an extreme random tree regression model to remove the litter pollution covariate from the temporary litter cover response index, including: S31: The extreme random tree regression model is obtained by parallel integration of multiple independent and unrelated extreme decision trees, wherein the extreme random tree regression model includes an input layer, a decision tree integration layer and an output layer; S32: The input layer receives the surface litter temporary cover response index and litter pollution covariate of the area grid; S33: The extreme decision tree in the decision tree ensemble layer adopts random selection of splitting features and random generation of splitting thresholds, and recursively divides the sample space layer by layer to adaptively fit the nonlinear and non-monotonic variation law of the litter pollution covariate with the change of the temporary litter cover response index. The variation patterns obtained from fitting each extreme decision tree are fused by mean fusion to obtain a global nonlinear mapping function with litter pollution covariate as input; S34: The output layer inputs the litter pollution covariate into the global nonlinear mapping function to obtain the interference component corresponding to the litter pollution covariate, calculates the difference between the temporary litter cover response index and the interference component, and obtains the temporary litter cover response index after removing the litter pollution covariate.

[0034] As an embodiment of the present invention, the present invention uses the temporary cover response index of surface litter and litter pollution covariate of all regional grids under the same background group as the same sample space. The extreme decision tree randomly selects a set of pixel mean sequences from the litter pollution covariate as the current splitting feature and randomly generates a splitting threshold to divide the current sample space into two sample sets. The above operation is repeated to recursively divide the sample space until the preset termination condition is reached (the number of divisions reaches 4 times). The mean of the output values ​​of all leaf nodes is taken as the extreme tree interference component corresponding to the litter pollution covariate in the sample space. The recursive division form of each extreme decision tree is converted into a function form by using a global nonlinear mapping function. The extreme tree interference components of each extreme decision tree are integrated by using the global nonlinear mapping function to form a global nonlinear mapping function for all regional grids under the background group. Furthermore, the actual surface litter temporary cover response index of the regional grid is collected, and the recursive partitioning rule of the extreme decision tree is optimized with the goal of minimizing the difference between the actual surface litter temporary cover response index and (surface litter temporary cover response index - interference component). Specifically, the litter thickness at the center of the regional grid is collected at a preset fixed period and normalized to form the actual surface litter temporary cover response index of the regional grid.

[0035] It should be noted that this invention employs an extreme random tree regression model, integrating multiple independent extreme decision trees in parallel. The sample space is recursively partitioned through random features and random threshold splitting, adaptively fitting the nonlinear, non-monotonic relationship between the litter pollution covariate and the temporary litter cover response index. This approach requires no pre-defined function form, exhibiting strong fitting ability and good noise resistance. A global nonlinear mapping function is obtained through multi-tree mean fusion, outputting the interference components corresponding to the litter pollution covariate. The model is then optimized using the true values ​​to ensure the accuracy and reliability of the interference components. Pollution removal is achieved by subtracting the interference components from the original temporary litter cover response index, resulting in clean intrinsic features. This effectively separates nonlinear interference, solving the problems of insufficient fitting and incomplete removal by traditional linear methods, significantly improving feature purification quality, and providing a highly reliable input for forestry soil organic carbon prediction models.

[0036] Step S3, which generates soil organic carbon prediction features and remediation correction features for each region's raster, also includes: The temporary cover response index of surface litter, after removing litter pollution covariates, is used as the soil organic carbon prediction feature of the regional raster, and the global nonlinear mapping function obtained in step S33 with litter pollution covariates as input is used as the purification correction feature generation rule of the regional raster.

[0037] S4: Based on the soil organic carbon prediction features and the purification correction feature generation rules, construct a forestry soil organic carbon prediction model. Use the forestry soil organic carbon prediction model to receive the spatiotemporally aligned remote sensing environmental features of regional grids in the forestry area, and generate the predicted soil organic carbon values ​​of regional grids in the forestry area.

[0038] Specifically, step S4 involves using the forestry soil organic carbon prediction model to receive the spatiotemporally aligned remote sensing environmental features of regional grids within a forestry area, and generating predicted soil organic carbon values ​​for these regional grids. This includes: S41: See below Figure 2 The diagram shown illustrates the structure of a forestry soil organic carbon prediction model. This model includes an input layer, a background group identification layer, an exponential variable extraction layer, a purification and correction layer, a gated adaptive screening layer, a residual mapping layer, and a regression prediction layer. S42: Spatiotemporal alignment of remote sensing environment characteristics of the input layer receiving area grid; S43: The background group identification layer extracts the forestry environment features from the spatiotemporally aligned remote sensing environment features, and divides the area grid into the corresponding background group based on the forestry environment features; S44: The index variable extraction layer extracts the background response sequence of the background group to which the regional raster belongs, and calculates the temporary cover response index of surface litter and the litter pollution covariate. S45: The purification correction layer uses the purification correction feature generation rules of the regional grid to generate the interference component corresponding to the litter pollution covariate, calculates the difference between the temporary cover response index of the surface litter and the interference component, and uses it as the temporary cover response index of the surface litter after removing the litter pollution covariate. S46: The gated adaptive screening layer receives the surface litter temporary cover response index and interference components after removing the litter pollution covariate. It uses a dual-branch, opposing gating weight method to generate interference gating weights for the residual litter pollution characteristic component and intrinsic gating weights for the effective soil organic carbon characteristic component. The formula for generating the dual-branch, opposing gating weights is as follows: ; ; in, Indicates intrinsic gating weights, Indicates the interference gating weight, This represents the response index of temporary land cover after removing the covariate of land cover pollution. Indicates interference components, This indicates the activation function; the default activation function is the Sigmoid function. All represent the trainable convolutional weight matrix parameters in the gated adaptive filtering layer. All of these represent trainable bias parameters in the gated adaptive filtering layer; S47: The residual mapping layer uses the interference gating weights and intrinsic gating weights to perform residual mapping on the surface litter temporary cover response index after removing the litter pollution covariate, obtaining the residual mapping features, wherein the residual mapping formula is: ; in, Represents the residual mapping characteristics. This represents the trainable residual convolution matrix in the residual mapping layer. This represents the trainable residual bias in the residual mapping layer. This represents the element-wise multiplication operator. This represents the weight control parameter (default setting is 0.2). S48: The regression prediction layer adopts a fully connected layer structure, with residual mapping features as input and soil organic carbon prediction value as output.

[0039] As an embodiment of the present invention, multiple training datasets are formed by extracting time-series data from multi-temporal remote sensing data and forestry environmental data. A training loss function is constructed with the goal of minimizing the difference between the measured data of mineral soil organic carbon and the predicted value of soil organic carbon. The gradient descent algorithm is used to optimize and solve the trainable parameters in the forestry soil organic carbon prediction model.

[0040] Example 2: As an embodiment of the present invention, this invention collects multi-temporal Sentinel-1 / 2 remote sensing data of a 320 km² area in a southern forestry demonstration zone, measured data of mineral soil organic carbon from 1260 regional raster cells, DEM topographic data, and vegetation type data to construct a complete prediction dataset including litter cover period, non-cover period, and seasonal transition period. The performance of the artificial intelligence-driven forestry soil organic carbon prediction method described in this invention is compared with that of traditional multiple linear regression models, random forest models, and in four indicators: coefficient of determination, root mean square error, mean absolute error, and prediction efficiency improvement factor. (Refer to...) Figure 3 The experimental comparison diagram shows that, compared with traditional models, this invention, through precise removal of surface litter interference, dual-branch opposition gating feature screening, and dual-path residual nonlinear fitting mechanism, achieves a 28.6% increase in prediction determination coefficient, a 41.3% reduction in root mean square error, a 35.7% reduction in mean absolute error, and a 3.2-fold increase in prediction efficiency for single-region grids. It has significant technical advantages and practical value in large-scale, highly interfered, and multi-temporal forestry soil organic carbon monitoring scenarios.

[0041] It should be noted that the terms "comprising," "including," or any other variations thereof used herein are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method 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, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0042] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0043] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. An artificial intelligence-driven method for predicting forestry soil organic carbon, characterized in that, The method includes: S1: The forestry area is rasterized, and the measured data of mineral soil organic carbon, multi-temporal remote sensing data and forestry environment data of the regional raster are collected. The collected data are processed by feature extraction to form the spatiotemporal aligned remote sensing environment features of the regional raster in the forestry area. S2: Extract the background response sequence from the spatiotemporally aligned remote sensing environmental features, calculate the temporary surface litter cover response index of the regional grid, and extract remote sensing covariates that are highly correlated with the temporary surface litter cover response index from the spatiotemporally aligned remote sensing environmental features as litter pollution covariates. S3: The extreme random tree regression model is used to remove the litter pollution covariate from the temporary litter cover response index and generate soil organic carbon prediction features and purification correction feature generation rules for each regional grid. S4: Based on the soil organic carbon prediction features and the purification correction feature generation rules, construct a forestry soil organic carbon prediction model. Use the forestry soil organic carbon prediction model to receive the spatiotemporally aligned remote sensing environmental features of regional grids in the forestry area, and generate the predicted soil organic carbon values ​​of regional grids in the forestry area.

2. The artificial intelligence-driven method for predicting forestry soil organic carbon as described in claim 1, characterized in that, In step S1, the forestry area is rasterized, and measured data of mineral soil organic carbon, multi-temporal remote sensing data, and forestry environmental data are collected from the raster area, including: S11: The forestry area is rasterized according to a regular square grid, dividing the forestry area into multiple non-overlapping area grids; S12: Soil sampling points are set up at the center of the area grid. Mineral soil samples are collected periodically at the soil sampling points according to a preset fixed period. The soil organic carbon content of the mineral soil samples is measured as the measured data of mineral soil organic carbon of the area grid. S13: Periodically collect remote sensing images of the area grid in different optical bands and microwave bands according to a preset fixed period, as multi-temporal remote sensing data of the area grid; S14: Collect the digital elevation model of the area grid and the vegetation type in the area grid as the forestry environment data of the area grid.

3. The artificial intelligence-driven method for predicting forestry soil organic carbon as described in claim 2, characterized in that, Step S1 involves feature extraction processing of the collected data to form spatiotemporally aligned remote sensing environmental features of the forestry area using regional raster data. This also includes: S15: Arrange the measured data of mineral soil organic carbon of the area grid in the order of collection time to obtain the measured data sequence of mineral soil organic carbon of the area grid, which is used as the measured organic carbon feature of the area grid. S16: Calculate the average pixel value of the remote sensing images of the region grid in different optical bands, and arrange them according to the acquisition time of the remote sensing images to obtain the sequence of the average pixel value of the region grid in different optical bands, which is used as the optical band feature of the region grid. S17: Calculate the average pixel value of the remote sensing image of the region grid in the microwave band, and arrange them according to the acquisition time of the remote sensing image to obtain the pixel average value sequence of the region grid in the microwave band, which is used as the microwave band feature of the region grid. S18: Based on the digital elevation model of the regional grid, extract the elevation value of the center of the regional grid, calculate the slope feature of the regional grid, classify and encode the vegetation type of the regional grid to obtain the vegetation coding feature of the regional grid, and splice the slope feature and vegetation coding feature of the regional grid as the forestry environment feature of the regional grid. S19: The measured organic carbon features, optical band features, microwave band features, and forestry environment features of the regional grid are spliced ​​together to form the spatiotemporally aligned remote sensing environment features of the regional grid.

4. The artificial intelligence-driven method for predicting forestry soil organic carbon as described in claim 1, characterized in that, Step S2 extracts the background response sequence from the spatiotemporally aligned remote sensing environmental features and calculates the temporary land cover response index of the regional raster, including: S21: Extract forestry environment features from the spatiotemporally aligned remote sensing environment features, and divide the region grid into multiple background groups based on the forestry environment features; S22: Calculate the mean optical band features and mean microwave band features of all regional grids in the background group to form the background response sequence of the background group; S23: Based on the background response sequence of the background group to which the regional grid belongs, calculate the difference sequence between the optical band features, microwave band features and background response sequence of the regional grid, and calculate the temporary surface cover response index of the regional grid based on the difference sequence.

5. The artificial intelligence-driven method for predicting forestry soil organic carbon as described in claim 4, characterized in that, Step S2, which extracts remote sensing covariates highly correlated with the temporary cover response index of surface litter from the spatiotemporally aligned remote sensing environmental features as litter pollution covariates, also includes: The optical band features and microwave band features in the spatiotemporally aligned remote sensing environment features are extracted, and the optical band features are split into pixel mean sequences of different optical bands, and the microwave band features are used as the pixel mean sequences of the microwave band. The Pearson correlation coefficient between the pixel mean sequence and the temporary cover response index of surface litter was calculated. Pixel mean sequences with Pearson correlation coefficients higher than a preset similarity threshold were selected as remote sensing covariates that are highly correlated with the temporary cover response index of surface litter. All extracted remote sensing covariates were used to form the litter pollution covariates of the regional raster.

6. The artificial intelligence-driven method for predicting forestry soil organic carbon as described in claim 1, characterized in that, In step S3, an extreme random tree regression model is used to remove the litter pollution covariate from the temporary litter cover response index, including: S31: The extreme random tree regression model is obtained by parallel integration of multiple independent and unrelated extreme decision trees, wherein the extreme random tree regression model includes an input layer, a decision tree integration layer and an output layer; S32: The input layer receives the surface litter temporary cover response index and litter pollution covariate of the area grid; S33: The extreme decision tree in the decision tree ensemble layer adopts random selection of splitting features and random generation of splitting thresholds, and recursively divides the sample space layer by layer to adaptively fit the nonlinear and non-monotonic variation law of the litter pollution covariate with the change of the temporary litter cover response index. The variation patterns obtained from fitting each extreme decision tree are fused by mean fusion to obtain a global nonlinear mapping function with litter pollution covariate as input; S34: The output layer inputs the litter pollution covariate into the global nonlinear mapping function to obtain the interference component corresponding to the litter pollution covariate, calculates the difference between the temporary litter cover response index and the interference component, and obtains the temporary litter cover response index after removing the litter pollution covariate.

7. The artificial intelligence-driven method for predicting forestry soil organic carbon as described in claim 6, characterized in that, Step S3, which generates soil organic carbon prediction features and remediation correction features for each region's raster, also includes: The temporary cover response index of surface litter, after removing litter pollution covariates, is used as the soil organic carbon prediction feature of the regional raster, and the global nonlinear mapping function obtained in step S33 with litter pollution covariates as input is used as the purification correction feature generation rule of the regional raster.

8. The artificial intelligence-driven method for predicting forestry soil organic carbon as described in claim 1, characterized in that, In step S4, the forestry soil organic carbon prediction model is used to receive the spatiotemporally aligned remote sensing environmental features of regional grids in the forestry area, and to generate predicted soil organic carbon values ​​for regional grids in the forestry area, including: S41: The forestry soil organic carbon prediction model includes an input layer, a background group identification layer, an exponential variable extraction layer, a purification correction layer, a gated adaptive screening layer, a residual mapping layer, and a regression prediction layer. S42: Spatiotemporal alignment of remote sensing environment characteristics of the input layer receiving area grid; S43: The background group identification layer extracts the forestry environment features from the spatiotemporally aligned remote sensing environment features, and divides the area grid into the corresponding background group based on the forestry environment features; S44: The index variable extraction layer extracts the background response sequence of the background group to which the regional raster belongs, and calculates the temporary cover response index of surface litter and the litter pollution covariate. S45: The purification correction layer uses the purification correction feature generation rules of the regional grid to generate the interference component corresponding to the litter pollution covariate, calculates the difference between the temporary cover response index of the surface litter and the interference component, and uses it as the temporary cover response index of the surface litter after removing the litter pollution covariate. S46: The gated adaptive screening layer receives the surface litter temporary cover response index and interference components after removing the litter pollution covariate. It uses a dual-branch, opposing gating weight method to generate interference gating weights for the residual litter pollution characteristic component and intrinsic gating weights for the effective soil organic carbon characteristic component. The formula for generating the dual-branch, opposing gating weights is as follows: ； ； in, Indicates intrinsic gating weights, Indicates the interference gating weight, This represents the response index of temporary land cover after removing the covariate of land cover pollution. Indicates interference components, This represents the activation function. All represent the trainable convolutional weight matrix parameters in the gated adaptive filtering layer. All of these represent trainable bias parameters in the gated adaptive filtering layer; S47: The residual mapping layer uses the interference gating weights and intrinsic gating weights to perform residual mapping on the surface litter temporary cover response index after removing the litter pollution covariate, obtaining the residual mapping features, wherein the residual mapping formula is: ； in, Represents the residual mapping characteristics. This represents the trainable residual convolution matrix in the residual mapping layer. This represents the trainable residual bias in the residual mapping layer. This represents the element-wise multiplication operator. Indicates the weight control parameters; S48: The regression prediction layer adopts a fully connected layer structure, with residual mapping features as input and soil organic carbon prediction value as output.

Images