Machine learning and multi-source variable-based high-precision prediction method for organic carbon in soil in arid region

Abstract

The invention discloses a machine learning and multi-source variable-based high-precision prediction method for soil organic carbon in an arid region, and solves the problems of high spatial heterogeneity of soil organic carbon in the arid region and high traditional detection cost. The method comprises the following steps: collecting a soil sample by adopting layered random sampling, and integrating seven types of environmental variables such as climate, vegetation, terrain and remote sensing to construct a feature set; a Boruta algorithm is adopted for screening, an optimized feature set is obtained, and redundant data are eliminated; the method comprises the following steps: dividing a training set and a test set through Kolmogorov-Smirnov test; the method comprises the following steps of: generating 100 base models by using parameter optimized random forest and extreme gradient lifting double-model collaborative prediction and combining a Bootstrap integration strategy; inputting the target area optimization feature set to obtain an organic carbon prediction result; according to the method, the monitoring cost is effectively reduced, spatial-temporal dynamic inversion of soil organic carbon based on historical environment variables is supported, and a reliable technical scheme is provided for dynamic monitoring and ecological management of soil carbon.

Description

Technical Field

[0001] The present invention relates to a method for accurately predicting soil organic carbon in arid regions, particularly to a method for accurately predicting soil organic carbon in arid regions based on machine learning and multi-source variables, belonging to the technical field of methods for accurately predicting soil organic carbon in arid regions. Background Art

[0002] Soil organic carbon, as a key link in the global carbon cycle, plays a crucial role in terrestrial ecosystems.

[0003] It is not only essential for maintaining soil fertility and promoting plant growth but also plays an irreplaceable role in regulating global climate change.

[0004] By fixing and storing carbon elements, soil organic carbon can effectively slow down the rising rate of carbon dioxide concentration in the atmosphere, thereby playing a certain inhibitory role in global warming.

[0005] Arid regions occupy a relatively large proportion of the earth's land surface, covering approximately 45% of the global land area.

[0006] Although the ecological systems in arid regions are relatively fragile, the soil organic carbon reserves in this region cannot be ignored, accounting for 32% of the global soil organic carbon pool (to a depth of 2 meters).

[0007] However, the soil organic carbon in arid regions is extremely sensitive to climate change. With the continuous warming of the global climate, climate model predictions show that by 2100, the area of arid regions will further expand, and it is expected to exceed 50% of the land area.

[0008] This change will lead to the aggravation of soil desertification, causing a large amount of the originally stored organic carbon in the soil to be released and converted into carbon dioxide and enter the atmosphere, thereby further promoting global warming and forming a vicious cycle.

[0009] Therefore, accurately grasping the spatial distribution of soil organic carbon in arid regions has become a key prerequisite for analyzing the response mechanism of the global carbon cycle and is of great significance for formulating scientific and effective climate change response strategies.

[0010] Traditional soil organic carbon detection methods mainly rely on field sampling and laboratory analysis. When sampling in the field, staff need to select representative sample points in the vast research area, collect soil samples, and then bring them back to the laboratory to measure the organic carbon content using various chemical analysis methods.

[0011] Although this method can obtain relatively accurate detection results, it has many limitations.

[0012] On the one hand, field sampling work requires a large amount of human, material and time costs.

[0013] Staff not only have to operate in complex terrains and harsh climate conditions, but also need to carry professional sampling equipment and tools, which makes the sampling process extremely difficult.

[0014] On the other hand, due to the limited sampling range, it is difficult to achieve comprehensive monitoring of large areas and cannot meet the needs of large-scale research.

[0015] Especially in arid regions, their special geographical environments, such as vast deserts, complex landforms and harsh climate conditions, make the sampling difficulty further increase, and the sampling cost also increases significantly.

[0016] With the continuous progress of technology, machine learning methods have gradually been applied in the field of soil organic carbon prediction.

[0017] Compared with traditional spatial interpolation and geostatistical methods, machine learning methods can more accurately capture complex patterns and laws in the data by constructing non-linear relationships between soil organic carbon and environmental variables, thus significantly improving the prediction accuracy.

[0018] At the same time, the rapid development of remote sensing technology, geographic information system and big data analysis provides rich data sources for soil organic carbon prediction.

[0019] Multi-source data fusion enables researchers to obtain information on multiple aspects including climate, vegetation, terrain, soil properties, etc., providing strong support for constructing a more perfect prediction model.

[0020] However, although the global-scale soil organic carbon prediction methods are constantly developing, due to the unique geographical conditions and significant spatial heterogeneity of soil organic carbon distribution in arid regions, the existing prediction methods still face many challenges in their application in this region.

[0021] The geographical environment in arid regions is complex and diverse, with huge differences in climate conditions, landforms, vegetation cover and soil types in different regions. These factors are intertwined, resulting in a highly uneven distribution of soil organic carbon.

[0022] This spatial heterogeneity increases the prediction difficulty, making it difficult for existing prediction models to accurately reflect the true distribution of soil organic carbon in arid regions, and the prediction accuracy cannot meet the actual needs.

[0023] Therefore, it is urgent to develop a high-precision soil organic carbon prediction technology suitable for arid regions. Summary of the Invention

[0024] The main object of the present invention is to provide a high-precision prediction method for soil organic carbon in arid areas based on machine learning and multi-source variables.

[0025] The object of the present invention can be achieved by adopting the following technical solutions:

[0026] A high-precision prediction method for soil organic carbon in arid areas based on machine learning and multi-source variables, comprising the following steps:

[0027] Step 1: Collect surface soil samples in arid areas, and use a high-frequency combustion-infrared carbon-sulfur analyzer to measure the organic carbon content to construct an actual measurement data set, and the sample spatial distribution needs to cover all land use types and soil types and be relatively uniform in space;

[0028] Step 2: Based on the geographical coordinates of the soil samples, obtain seven types of environmental variables including climate, vegetation, terrain, soil properties, soil type, land use type, and remote sensing, and construct an initial feature set;

[0029] Step 3: Adopt the Boruta algorithm, input the measured soil organic carbon data and the initial feature set into the random forest model, and retain the key features marked as Confirmed in terms of feature importance to form an optimized feature set;

[0030] Step 4: Adopt the Kolmogorov-Smirnov test method based on the assumption of the same distribution, and generate candidate division schemes through 100 random samplings of the measured organic carbon data according to a ratio of 9:1, and select the scheme with the smallest K-S statistic to divide the training set and the test set;

[0031] Step 5: Select the random forest and extreme gradient boosting algorithms, input the training set and the optimized feature set, determine the optimal parameter combination through grid search and ten-fold cross-validation, and train 100 base models by Bootstrap sampling;

[0032] Step 6: Input the optimized feature set of the target area into the 100 base models, and output the prediction results of the soil organic carbon in the target area; Step 7: Calculate the decision coefficient, mean absolute error, and root mean square error based on the test set for accuracy evaluation, and calculate the total prediction variance and average prediction interval for uncertainty evaluation.

[0033] Preferably, in the step 1, collect surface soil samples of 0-20 cm, and obtain samples by the five-point mixing method.

[0034] Preferably, in the step 2, the climate variables include annual average rainfall, annual average temperature, and potential evapotranspiration;

[0035] The vegetation variables include vegetation normalized difference index, net primary productivity, and enhanced vegetation index; the terrain variables include elevation, slope, aspect, and terrain wetness index;

[0036] Soil property variables include clay content, gravel content, and silt content;

[0037] Remote sensing variables include Bands 1-7 of Landsat8 OLI and the calculated soil salinity index and normalized humidity index.

[0038] Preferably, in step three, Boruta feature screening is performed in the R4.1.2 environment, and the number of random forest iterations is set to 200 times.

[0039] Preferably, in step five, when optimizing the random forest model parameters, the parameter search range is n_estimators ∈ [100, 1000], max_depth ∈ [5, 20], max_features ∈ [0.3, 0.8];

[0040] When optimizing the extreme gradient boosting model parameters, the parameter search range is n_estimators ∈ [100, 1000], max_depth ∈ [3, 10], learning_rate ∈ [0.01, 0.3].

[0041] Preferably, in step six, the accuracy evaluation formula is:

[0042]

[0043] where y i is the measured organic carbon value of the i-th sample in the test set;

[0044] is the predicted organic carbon value of the i-th sample in the test set;

[0045] is the average value of the measured organic carbon in the test set;

[0046] n is the number of samples in the test set;

[0047] R 2 is the coefficient of determination, which measures the ability of the model to explain the variability of the target variable. The closer R 2 is to 1, the better the model fitting effect;

[0048] MAE measures the average absolute deviation between the model predicted value and the measured value. The closer MAE is to 0, the higher the model prediction accuracy;

[0049] RMSE measures the average deviation between the model predicted value and the measured value. The closer RMSE is to 0, the higher the model prediction accuracy.

[0050] The formula for the total prediction variance and average prediction interval for uncertainty evaluation is:

[0051]

[0052] VAR total = VAR bootstrap + MSE test

[0053]

[0054] PIR i = UPL i - LPL i

[0055]

[0056] where y i is the measured organic carbon value of the i-th sample in the test set;

[0057] is the predicted organic carbon value of the i-th sample in the test set;

[0058] n is the number of samples in the test set;

[0059] MSE test is a measure of the mean squared deviation between the predicted and measured values in the test set;

[0060] is the variance of the predicted results of 100 iterations for the i-th sample, calculated based on the Bootstrap sampling method and training 100 base models;

[0061] is the total predicted variance of the i-th sample;

[0062] is the average value of the predicted results of the i-th sample;

[0063] 1.645 is the z-value for a 90% confidence interval;

[0064] UPL i is the upper limit value of the 90% confidence interval for the i-th sample;

[0065] LPL i is the lower limit value of the 90% confidence interval for the i-th sample;

[0066] PIR i is the range value of the 90% confidence interval for the i-th sample. The narrower the interval range, the more stable the prediction result;

[0067] m is the number of predicted samples in the target area;

[0068] MPI is the average prediction interval value of the prediction results, reflecting the average uncertainty range of the model's prediction results in the target area. The smaller the value, the more stable the model's prediction results.

[0069] Preferably, it further includes a high-precision prediction system for soil organic carbon in arid areas for the high-precision prediction method of soil organic carbon in arid areas, and the system includes the following:

[0070] Sample collection and dataset construction module, which is used to collect surface soil samples in arid areas, measure the organic carbon content, and construct an actual measurement dataset;

[0071] Environmental variable acquisition and feature set construction module, which is used to acquire seven types of environmental variables and construct an initial feature set;

[0072] Feature screening module, which uses the Boruta algorithm to screen key features and generate an optimized feature set;

[0073] Dataset division module, which uses the Kolmogorov-Smirnov test method to divide the training set and the test set;

[0074] Model construction and optimization module, which selects the random forest and extreme gradient boosting algorithms, optimizes the parameters through grid search and ten-fold cross-validation, and trains the base model;

[0075] Model prediction module, which inputs the optimized feature set of the target area into the base model to obtain the prediction results of soil organic carbon;

[0076] Model evaluation and application module, which evaluates the accuracy and uncertainty of the model prediction results, and applies the model to predict and map soil organic carbon.

[0077] Preferably, the sample collection and dataset construction module collects 0-20 cm surface soil samples and obtains samples by the five-point mixing method;

[0078] The environmental variables obtained by the environmental variable acquisition and feature set construction module include seven types of variables: climate, vegetation, terrain, soil properties, soil type, land use type, and remote sensing.

[0079] Preferably, the parameter search ranges of the random forest model and the extreme gradient boosting model in the model construction and optimization module.

[0080] The beneficial technical effects of the present invention:

[0081] The high-precision prediction method of soil organic carbon in arid areas based on machine learning and multi-source variables provided by the present invention selects the best prediction model by comparing the accuracy indexes of the test set to ensure the reliability of the prediction results;

[0082] Multi-source data fusion and Boruta feature screening eliminate redundant features and improve the model prediction accuracy;

[0083] Adopt a dual model of random forest and extreme gradient boosting, and select the best prediction model by comparing the accuracy of the test set to ensure the reliability of the prediction results;

[0084] Construct an error evaluation system through the Bootstrap integration strategy to visualize the error boundary of the predicted results of soil organic carbon content;

[0085] Significantly reduce the cost of soil organic carbon monitoring at the regional scale in arid areas, and support the carbon dynamic inversion based on historical environmental variable data. Brief Description of the Drawings

[0086] Figure 1 This is the flow chart of the method of the present invention, specifically showing the full-process technical framework of sample collection, feature screening, model training and verification;

[0087] Figure 2 It is the spatial distribution and elevation characteristics of the sampling points in the target area (Tarim River Basin) of the embodiment;

[0088] Figure 3 It is the scatter plot of the predicted values and measured values of the random forest and extreme gradient boosting models, and shows the goodness of fit and the distribution range of soil organic carbon content;

[0089] Figure 4 It is the soil organic carbon content, upper limit, lower limit and interval of the 90% confidence interval at a 1-kilometer resolution in the Tarim River Basin predicted based on extreme gradient boosting. Detailed Embodiment

[0090] To make the technical solutions of the present invention clearer and more definite for those skilled in the art, the present invention will be further described in detail below in conjunction with the embodiments and the drawings, but the embodiments of the present invention are not limited thereto.

[0091] Taking the Tarim River Basin as the research area, the implementation steps are as follows:

[0092] (1) Soil sample collection and dataset construction. According to the 1:250,000 land use type map and soil type map of the Tarim River Basin, sampling points are arranged by the stratified random sampling method, and the uniformity of the spatial distribution of the sample points is considered.

[0093] Ensure coverage of 8 major land use types such as cultivated land, forest land, grassland, wetland, desert, saline-alkali land, bare land, and construction land;

[0094] At each sampling point, surface soil samples of 0-20 cm are collected by the five-point mixing method, and a total of 861 samples are obtained;

[0095] Use a high-frequency combustion-infrared carbon-sulfur analyzer to measure the organic carbon content of soil samples and construct a dataset of measured soil organic carbon content.

[0096] (2) Acquisition and preprocessing of multi-source environmental variables. The multi-source environmental variables are obtained through the following data sources:

[0097] The data of annual average rainfall and annual average temperature in climate variables are from the National Tibetan Plateau Scientific Data Center (http: / / data.tpdc.ac.cn), and the data of potential evapotranspiration are from the Earth Resource Data Cloud Platform (http: / / www.gis5g.com); the data of vegetation normalized difference index, net primary productivity and enhanced vegetation index in vegetation variables are from the Earth Resource Data Cloud Platform (http: / / www.gis5g.com); the elevation data in terrain variables are from the Earth Resource Data Cloud Platform (http: / / www.gis5g.com), and the derived slope, aspect and terrain wetness index are calculated by ArcGIS 10.8 software. The soil type variable data are from the Resource Science Data Platform (https: / / www.resdc.cn); the land use type variable data are from the National Glaciology, Cryopedology and Desert Science Data Center (http: / / www.ncdc.ac.cn); the soil property variable data are from the Earth Resource Data Cloud Platform (http: / / www.gis5g.com), including clay content, gravel content and sand content; the remote sensing variables obtain Landsat 8 OLI images through the Google Earth Engine platform, including Band 1-7 spectral bands, and calculate the soil salinity index (SI) and normalized difference water index (NDWI) based on the public formula.

[0098]

[0099] NDWI = (Band 5 - Band 6) / (Band 5 + Band 6)

[0100] (3) Feature screening and construction of optimized feature set. Perform Boruta feature screening in the R 4.1.2 environment:

[0101] Input the measured soil organic carbon content data and the initial feature set, and set the number of random forest iterations to 200 times; according to the feature importance ranking of the output results, retain the key features marked as "Confirmed", and screen out 14 key features from 24 initial features, namely annual average rainfall, annual average temperature, potential evapotranspiration, vegetation normalized difference index, net primary productivity, enhanced vegetation index, elevation, land use type, Band 3, Band 5, Band 6, Band 7, soil salinity index, normalized difference water index; thus construct an optimized feature set for subsequent modeling.

[0102] (4) Dataset partitioning and validation. Adopt a data partitioning strategy based on distribution consistency:

[0103] The measured soil organic carbon content dataset generates candidate partitioning schemes through 100 random samplings, with a ratio of 9:1 (775 samples in the training set and 86 samples in the test set); calculate the Kolmogorov-Smirnov (K-S) test statistic D value of the training set and the test set after each partitioning; select the partitioning with the smallest D value as the final partitioning scheme.

[0104] (5) Prediction model construction and parameter optimization. Use the random forest model and the extreme gradient boosting model for modeling. The specific steps are as follows:

[0105] Random forest model parameter optimization. Key parameters: n_estimators (number of decision trees), max_depth (maximum depth of the tree), max_features (maximum number of features considered when splitting nodes); parameter search range: n_estimators ∈ [100, 1000], max_depth ∈ [5, 20], max_features ∈ [0.3, 0.8]; optimization method: use the soil organic carbon content of the test set as the dependent variable and the optimized feature set as the independent variable, and use ten-fold cross-validation to evaluate the model performance, and select the parameter combination with the smallest mean absolute error (MAE).

[0106] Extreme gradient boosting model parameter optimization. Key parameters: n_estimators (number of base learners), max_depth (maximum depth of the tree), learning_rate (learning rate); parameter search range: n_estimators ∈ [100, 1000], max_depth ∈ [3, 10], learning_rate ∈ [0.01, 0.3]; optimization method: use the soil organic carbon content of the test set as the dependent variable and the optimized feature set as the independent variable, and use ten-fold cross-validation to evaluate the model performance, and select the parameter combination with the smallest mean absolute error (MAE).

[0107] Model training and integration. Adopt the Bootstrap method for model integration: set the number of iterations num_iterations = 100, and draw 100 sub-sample sets from the training set with replacement; train the random forest and extreme gradient boosting base models for each sub-sample set respectively.

[0108] (5) Prediction of organic carbon. The optimized feature set corresponding to the 1-km resolution spatial points in the Tarim River Basin was input into the base models, and the predicted values and statistics (mean, variance) of all base models were stored for subsequent prediction accuracy and uncertainty assessment.

[0109] (6) Model application and effect evaluation.

[0110] The prediction accuracy of the model was evaluated by calculating the coefficient of determination (R 2 ), mean absolute error (MAE), and root mean square error (RMSE) between the mean of the predicted organic carbon values and the measured organic carbon values of the test set at the corresponding spatial points:

[0111]

[0112] where y i is the measured organic carbon value of the i-th sample in the test set; is the predicted organic carbon value of the i-th sample in the test set; is the mean of the measured organic carbon in the test set; n is the number of samples in the test set;

[0113] Accuracy verification results: Random forest model: R2 = 0.77, MAE = 2.28 g·kg -1 , RMSE = 3.55 g·kg -1 ; Extreme gradient boosting model R2 = 0.89, MAE = 1.91 g·kg -1 , RMSE = 2.39 g·kg -1 . The results show that the extreme gradient boosting model has better prediction effect ( Figure 3 ).

[0114] Uncertainty visualization: The distribution map of soil organic carbon content at 1-km resolution ( Figure 4 a), upper limit of 90% confidence interval prediction ( Figure 4 b), lower limit of prediction ( Figure 4 c) and interval value distribution map ( Figure 4 d) were drawn to visualize the error boundary.

[0115] Uncertainty quantification: The prediction uncertainty of the model was evaluated by calculating the total prediction variance (VARtotal) and mean prediction interval (MPI):

[0116]

[0117] VAR total = VAR bootstrap + MSE test

[0118]

[0119] PIR i = UPL i - LPL i

[0120]

[0121] where y i is the measured organic carbon value of the i-th sample in the test set; is the predicted organic carbon value of the i-th sample in the test set; n is the number of samples in the test set; MSE test is the mean squared deviation between the predicted and measured values in the test set; is the variance of the predicted results of 100 iterations for the i-th sample based on the Bootstrap sampling method and training 100 base models; is the total predicted variance of the i-th sample; is the mean of the predicted results of the i-th sample; 1.645 is the z-value for the 90% confidence interval; UPL i is the upper limit value of the 90% confidence interval of the i-th sample; LPL i is the lower limit value of the 90% confidence interval of the i-th sample; PIR i is the range value of the 90% confidence interval of the i-th sample. The narrower the interval range, the more stable the prediction result; m is the number of predicted samples in the target area; MPI is the average prediction interval value of the prediction results, reflecting the average uncertainty range of the model's prediction results in the target area. The smaller the value, the more stable the model's prediction results..

[0122] Model uncertainty comparison: In this embodiment, the MPI of the random forest model and the extreme gradient boosting model are 13.88 g·kg -1 and 10.50 g·kg -1 respectively. The smaller MPI value of the extreme gradient boosting model indicates that its prediction result has lower uncertainty and higher reliability.

[0123] During application, input the optimized feature set of the soil samples in the target area into the trained random forest model and extreme gradient boosting model, and perform the following operations:

[0124] The predicted organic carbon values and predicted variances output by each model;

[0125] Calculate the 90% confidence interval based on the predicted variance;

[0126] Compare the prediction results of the two models, and preferentially use the output value of the model with higher prediction accuracy as the final prediction result.

[0127] As described above, it is only a further embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the scope disclosed by the present invention, making equivalent substitutions or changes according to the technical solution and concept of the present invention, shall fall within the protection scope of the present invention.

Claims

1. A high-precision prediction method for soil organic carbon in arid regions based on machine learning and multi-source variables, characterized in that: It includes the following steps: Step 1: Collect surface soil samples in arid areas, measure the organic carbon content using a high-frequency combustion-infrared carbon-sulfur analyzer to construct an actual measurement dataset, and the spatial distribution of the samples needs to cover all land use types and soil types and be relatively uniform spatially; Step 2: Based on the geographical coordinates of the soil samples, obtain seven types of environmental variables including climate, vegetation, terrain, soil properties, soil type, land use type, and remote sensing, and construct an initial feature set; Step 3: Use the Boruta algorithm, input the actual measured soil organic carbon data and the initial feature set into the random forest model, and retain the key features with the feature importance marked as Confirmed to form an optimized feature set; Step 4: Use the Kolmogorov-Smirnov test method based on the assumption of the same distribution, generate candidate partitioning schemes by randomly sampling the actual measured organic carbon data 100 times according to a ratio of 9:1, and select the scheme with the smallest K-S statistic to partition the training set and the test set; Step 5: Select the random forest and extreme gradient boosting algorithms, input the training set and the optimized feature set, determine the optimal parameter combination through grid search and ten-fold cross-validation, and train 100 base models using Bootstrap sampling; Step 6: Input the optimized feature set of the target area into the 100 base models and output the prediction results of the soil organic carbon in the target area; Step 7: Calculate the decision coefficient, mean absolute error, and root mean square error based on the test set for accuracy evaluation, and calculate the total prediction variance and average prediction interval for uncertainty evaluation.

2. The high-precision prediction method for soil organic carbon in arid areas based on machine learning and multi-source variables according to claim 1, wherein: In the above Step 1, collect surface soil samples of 0-20 cm, and obtain the samples using the five-point mixing method.

3. The high-precision prediction method for soil organic carbon in arid areas based on machine learning and multi-source variables according to claim 2, characterized in that: In the above Step 2, the climate variables include annual average rainfall, annual average temperature, and potential evapotranspiration; The vegetation variables include vegetation normalized difference index, net primary productivity, and enhanced vegetation index; the terrain variables include elevation, slope, aspect, and terrain wetness index; The soil property variables include clay content, gravel content, and silt content; The remote sensing variables include Bands 1-7 of Landsat8 OLI and the calculated soil salinity index and normalized humidity index.

4. The high-precision prediction method for soil organic carbon in arid areas based on machine learning and multi-source variables according to claim 3, wherein: In the above Step 3, perform Boruta feature screening in the R4.1.2 environment, and set the number of random forest iterations to 200 times.

5. The high-precision prediction method of soil organic carbon in arid areas based on machine learning and multi-source variables according to claim 4, characterized in that: When optimizing the parameters of the random forest model in the above Step 5, the parameter search range is n_estimators∈[100,1000], max_depth∈[5,20], max_features∈[0.3,0.8]; When optimizing the parameters of the extreme gradient boosting model in the above Step 5, the parameter search range is n_estimators∈[100,1000], max_depth∈[3,10], learning_rate∈[0.01,0.3].

6. The high-precision prediction method for soil organic carbon in arid areas based on machine learning and multi-source variables according to claim 5, characterized in that: In the above Step 6, use 100 base models to obtain the prediction results of the soil organic carbon in the target area.

7. The high-precision prediction method of soil organic carbon in arid areas based on machine learning and multi-source variables according to claim 6, characterized in that: In the above Step 7, the accuracy evaluation formula is: where y i is the measured organic carbon value of the i-th sample in the test set; is the predicted organic carbon value of the i-th sample in the test set; is the average value of measured organic carbon in the test set; n is the number of samples in the test set; R 2 is the decision coefficient, which measures the explanatory power of the model for the variability of the target variable. The closer R 2 is to 1, the better the model fitting effect; MAE measures the average absolute deviation between the model prediction value and the actual measured value. The closer MAE is to 0, the higher the model prediction accuracy; RMSE measures the average deviation between the predicted values and the measured values of the model. The closer the RMSE is to 0, the higher the prediction accuracy of the model; The formulas for the total prediction variance and the average prediction interval for uncertainty evaluation are as follows: VAR total = VAR bootstrap + MSE test PIR i = UPL i - LPL i where y i is the measured organic carbon value of the i-th sample in the test set; is the predicted organic carbon value of the i-th sample in the test set; n is the number of samples in the test set; MSE test is a measure of the average squared deviation between the predicted values and the measured values in the test set; Based on the Bootstrap sampling method, 100 base models are trained, and for the i-th sample, the variance of the prediction results of 100 iterations is calculated; is the predicted total variance of the i-th sample; is the average of the prediction results of the i-th sample; 1.645 is the z-value for a 90% confidence interval; UPL i is the upper limit value of the 90% confidence interval for the i-th sample; LPL i is the lower limit value of the 90% confidence interval for the i-th sample; PIR i is the 90% confidence interval range value of the i-th sample. The narrower the interval range, the more stable the prediction result; m is the number of predicted samples in the target area; MPI is the average prediction interval value of the prediction result, which reflects the average uncertainty range of the model's prediction result in the target area. The smaller the value, the more stable the model's prediction result.

8. The high-precision prediction method for soil organic carbon in arid areas based on machine learning and multi-source variables according to claim 7, characterized in that: It also includes a high-precision prediction system for soil organic carbon in arid areas for the high-precision prediction method of soil organic carbon in arid areas. The system includes the following: The sample collection and dataset construction module is used to collect surface soil samples in arid areas and measure the organic carbon content to construct a measured dataset; The environmental variable acquisition and feature set construction module is used to acquire seven types of environmental variables to construct an initial feature set; The feature screening module uses the Boruta algorithm to screen key features and generate an optimized feature set; The dataset division module uses the Kolmogorov-Smirnov test method to divide the training set and the test set; The model construction and optimization module selects the random forest and extreme gradient boosting algorithms, and optimizes the parameters and trains the base model through grid search and ten-fold cross-validation; The model prediction module inputs the optimized feature set of the target area into the base model to obtain the prediction result of soil organic carbon; The model evaluation and application module evaluates the accuracy and uncertainty of the model prediction result, and applies the model to predict and map soil organic carbon.

9. The high-precision prediction method for soil organic carbon in arid areas based on machine learning and multi-source variables according to claim 8, characterized in that: The sample collection and dataset construction module collects 0-20 cm surface soil samples and obtains samples using the five-point mixing method; The environmental variables obtained by the environmental variable acquisition and feature set construction module include seven types of variables: climate, vegetation, terrain, soil properties, soil type, land use type, and remote sensing.

10. The high-precision prediction method for soil organic carbon in arid areas based on machine learning and multi-source variables according to claim 9, characterized in that: The parameter search range of the random forest model and the extreme gradient boosting model in the model construction and optimization module.

Images