Optimizing environmental covariates for digital mapping of soil organic carbon in Hunan Province, China Authors: Xiaojing Cai, Falin Liu, Zhengui Cai Published: 2026 ABSTRACT Soil organic carbon (SOC) is widely recognized as a fundamental indicator of soil fertility, ecosystem functioning, and overall soil health. Effective land management requires continuous monitoring of SOC variations through modern technological approaches. In this study, 477 soil samples were meticulously collected and analyzed for SOC content in the laboratory. Terrain attributes and spectral indices were then derived from satellite data. Machine learning models, including support vector machine (SVM), artificial neural network (ANN), and random forest (RF), were employed to predict SOC content. To improve computational efficiency and model accuracy, the variance inflation factor (VIF) and Boruta’s variable selection methods were applied, identifying the most relevant environmental covariates. Results demonstrated that only 5 out of 40 environmental covariates were optimal for SOC modeling. Using these selected covariates, the RF model achieved the highest prediction accuracy (R² = 0.84, RMSE = 0.069%, and PRD = 3.6%). The RF model effectively captured the inherent variability and complexity of soil properties, yielding precise and reliable SOC predictions. The results emphasize the capability of machine learning in predicting SOC levels, aiding in the enhancement of soil management strategies and agricultural planning. Ultimately, this study provides a foundation for integrating advanced predictive techniques to enhance SOC assessment in future study. Introduction Soil is a dynamic and evolving natural matrix, with its properties continuously changing due to various environmental and biological processes1–5. Among these properties, soil organic carbon (SOC) content is particularly significant, as it plays a crucial role in enhancing soil fertility and quality through both physical and chemical improvements2,6–8. SOC is also a key indicator of soil health and productivity, consisting of decomposed plant and animal residues that are processed by soil microorganisms9,10. Despite its importance for maintaining soil health and ecosystem stability, soils with high SOC content are relatively scarce worldwide11. This scarcity highlights the need for effective SOC management strategies, particularly in soils with medium to low SOC levels12,13. In recent years, industrial activities and land use changes have accelerated the decomposition of SOC, increased soil aerobic respiration, enhanced mineralization, and intensified humus oxidation, leading to the release of CO₂ from the soil into the atmosphere14,15. These challenges have driven researchers to prioritize SOC mapping on a global scale. Traditional approaches for assessing SOC content rely on extensive field sampling and laboratory analyses, which are both time-consuming and costly16. Recently, significant research has explored the potential of modern technologies in agricultural applications10,17. The primary objective of integrating these advanced techniques in agriculture is to enhance precision management and ensure food security. Consequently, remote sensing (RS), proximal sensing, and other innovative digital soil mapping (DSM) methods have witnessed substantial advancements18. These technologies facilitate large-scale soil property analysis with improved efficiency and accuracy8,19–24. Numerous studies have established strong correlations between remote sensing (RS) indices, terrain attributes, and SOC content6,8,10,24–26. These parameters, when integrated into machine learning models, can yield highly accurate predictions, typically reflected by performance metrics such as an R² greater than 0.70–0.85 and RMSE values that indicate minimal deviation between observed and predicted SOC values. In this context, “highly accurate” refers to prediction performance that exceeds models relying solely on traditional covariates such as limited field-measured soil properties or coarse-resolution environmental data. One key reason for their effectiveness is that an increase in SOC content leads to a decrease in soil spectral reflectance. However, it is important to note that the influence of terrain attributes and other covariates on SOC content can vary depending on the spatial scale and landscape complexity; variables that are significant at local scales may become less important at broader scales27. According to Daniel, et al. 28, the near-infrared (NIR) band (2.16–2.9 μm) exhibits the highest correlation with SOC content. Khosravi Aqdam, et al. 29 predicted SOC content in northwestern Iran using satellite data through partial least squares regression (PLSR) and random forest (RF) models. Their results validated that the RF model delivered precise predictions of SOC content. Moreover, Guo, et al. 20 emphasized that the precision of SOC estimation enhances with higher spatial resolution of satellite data. Due to the critical role of SOC, numerous models have been developed to predict its distribution30–32. However, the most effective model is the one that achieves high predictive performance while utilizing the fewest variables, ensuring computational efficiency and facilitating a clear understanding of the studied variable21,33. The complexity of SOC modeling stems from the varying influence of different factors, leading to uncertainties in predictions. In agricultural regions, these uncertainties are further complicated by land use, land use changes, and anthropogenic activities, which can significantly alter SOC dynamics. These uncertainties arise due to complex interactions among SOC-related parameters and the challenges in identifying the key influencing factors34. Additionally, discrepancies in model structures contribute further to prediction uncertainties35. Given the importance of SOC content and its associated processes, especially in agricultural lands, optimizing its prediction is essential for achieving sustainable agricultural development and long-term food security. The development of accurate SOC maps plays a critical role in effective land use planning and resource management, particularly in regions where agricultural productivity directly affects the well-being of densely populated areas. By integrating SOC data with advanced predictive models, agricultural practices can be optimized, soil fertility can be enhanced, and environmental protection can be supported. Therefore, this study aims to: (1) optimize SOC content prediction by identifying key environmental covariates that significantly influence its spatial distribution, and (2) assess the performance of various machine learning algorithms to improve SOC prediction, providing a framework tailored specifically for regional agricultural policy and planning. Materials and methods Study area The study site is situated in south-central China (Fig. 1), extending between latitudes 24°30′00″ N and 30°00′00″ N and longitudes 108°30′00″ E to 114°30′00″ E, covering an area of approximately 2.12 × 10⁵ km² (Fig. 1). This region features diverse landforms, including mountains, hills, plains, and river systems. The region experiences an average yearly temperature of 18.3 °C and an annual total precipitation of 1,795 mm (http://data.cma.cn). This region is a crucial agricultural zone, known for its rice, tea, and citrus production, owing to its favorable climate with diverse temperatures and rainfall patterns. However, it is also highly vulnerable to natural disasters, such as floods and droughts. To address these challenges, effective management and adaptation strategies are required for sustainable development and climate resilience. In Hunan Province, various measures are being adopted, including the improvement of irrigation infrastructure, the use of flood-resistant crop varieties, and the application of advanced agricultural technologies. Moreover, efficient disaster management and climate adaptation strategies are essential to enhance the region’s resilience and long-term sustainability. Field studies, soil sampling, and laboratory analysis A total of 477 topsoil samples (0–30 cm) were randomly gathered from the study region between June 8th and July 30th, 2023 (Fig. 1). The soil sampling was conducted with the formal permission of the private landowners, and all procedures were performed in strict accordance with local environmental regulations. Due to access limitations, dense forest areas were not sampled directly; instead, samples were taken from the forest edges. The objective was to ensure comprehensive coverage of the entire study area with the samples. Random sampling was chosen for its advantage of not requiring prior knowledge of the spatial variation in soil properties36. Once the sampling locations were determined, soil samples were collected from a fixed depth of 0–30 cm and placed into zip bags for transportation. After the samples were transported to the laboratory, they were air-dried for 48 h and passed through a 2 mm sieve. In the subsequent phase, the SOC content was measured using the Walkley-Black method37. GPS-Garmin with an accuracy of ± 3 m was used to determine the exact location of each sampling point. Environmental covariates Satellite images from Landsat 8 and the Digital Elevation Model (DEM), obtained from the Shuttle Radar Topography Mission (SRTM, 30 m resolution), related to the study region were acquired using Google Earth Engine (https://code.earthengine.google.com/). Terrain attributes were derived from the DEM using SAGA-GIS software, which involved processing the DEM to extract topographic parameters. Concurrently, Landsat 8 satellite images were utilized to calculate spectral indices. To improve the precision and eliminate potential errors, radiometric and atmospheric corrections were applied to the images using ENVI software prior to analysis. In total, 40 parameters were extracted from the Landsat 8 imagery and DEM (Table 1). All parameters were resampled to a 30-meter pixel size to minimize modeling errors38. Table 1Environmental covariates used for predicting SOC content 1.CovariatesAbbreviationNormalized different vegetation indexNDVIRenormalized difference vegetation indexRDVIWetness brightness differenceWBDINormalized difference salinity indexNDSINormalized difference moisture indexNDMIClay IndexCIBrightness IndexBIGypsum IndexGIGreen-Red Vegetation IndexGVIIron Oxide RatioIORAdjusted Vegetation IndexSAVIModified Soil Adjusted Vegetation IndexMSAVIModified Normalized Difference Water IndexMNDVIEnhanced Vegetation IndexEVINormalized Ratio Vegetation IndexNRVIAnalytical Hill shadingAHAspectAspectCatchment SlopeCSChannel DirectionCDChannel Network Base LevelCNBLChannel Network DistanceCNDConvergence IndexCoIDiffuse InsolationDiffuseGeomorphonsGeomorphonsLandformsLandformsLS FactorLS FactorMass Balance IndexMBIModified Catchment AreaMCAMRRTFMRRTFPlan CurvaturePCRelative Slope PositionRSPSlopeSlopeStandardized HeightSHMultiresolution index of valley bottom flatnessMRVBFNormalized HeightNHMidslope positionMSTerrain surface textureTESlope gradientSGConvexityConvexityTopographic Wetness IndexTWIValley Depth VD Table 1Environmental covariates used for predicting SOC content 1.CovariatesAbbreviationNormalized different vegetation indexNDVIRenormalized difference vegetation indexRDVIWetness brightness differenceWBDINormalized difference salinity indexNDSINormalized difference moisture indexNDMIClay IndexCIBrightness IndexBIGypsum IndexGIGreen-Red Vegetation IndexGVIIron Oxide RatioIORAdjusted Vegetation IndexSAVIModified Soil Adjusted Vegetation IndexMSAVIModified Normalized Difference Water IndexMNDVIEnhanced Vegetation IndexEVINormalized Ratio Vegetation IndexNRVIAnalytical Hill shadingAHAspectAspectCatchment SlopeCSChannel DirectionCDChannel Network Base LevelCNBLChannel Network DistanceCNDConvergence IndexCoIDiffuse InsolationDiffuseGeomorphonsGeomorphonsLandformsLandformsLS FactorLS FactorMass Balance IndexMBIModified Catchment AreaMCAMRRTFMRRTFPlan CurvaturePCRelative Slope PositionRSPSlopeSlopeStandardized HeightSHMultiresolution index of valley bottom flatnessMRVBFNormalized HeightNHMidslope positionMSTerrain surface textureTESlope gradientSGConvexityConvexityTopographic Wetness IndexTWIValley Depth VD Environmental covariates used for predicting SOC content 1. Optimizing modeling through selection of environmental covariates In this study, the collected data were randomly split into two subsets: 70% for training and 30% for testing39. To select the environmental covariates, two approaches were employed: The Variance Inflation Factor (VIF) and Boruta’s variable selection technique. The VIF is used to evaluate the degree of multicollinearity among predictors in regression analysis. Environmental covariates with VIF values greater than 10 (VIF > 10) were excluded, as such values indicate significant multicollinearity40. The Boruta method, a semi-automated, supervised variable selection technique, further identifies the most relevant environmental features and helps prevent model overfitting, ensuring that only the most important variables are retained41. Following variable selection, soil organic carbon (SOC) content was modeled using machine learning algorithms, including Random Forest (RF), Support Vector Machine (SVM), and Artificial Neural Network (ANN) models42. RF model The RF model is an ensemble learning method that constructs a multitude of decision trees during training and outputs either the mode of the classes (for classification) or the mean prediction (for regression) of the individual trees. This approach reduces the risk of overfitting compared to a single decision tree and can handle large datasets with high dimensionality. The RF algorithm works by creating multiple bootstrap samples from the training data, building a decision tree for each sample, and then aggregating the results. It also computes variable importance scores, which allow researchers to identify the most influential predictors in modeling SOC content. This makes RF both a robust and interpretable tool for understanding SOC dynamics and optimizing management strategies29. SVM model The SVM is a supervised machine learning algorithm designed for classification and regression tasks. In SVM regression (also called Support Vector Regression, SVR), the algorithm aims to find a function that approximates the relationship between input variables and SOC content while minimizing prediction errors within a defined margin (epsilon). SVM transforms the input data into a high-dimensional feature space using kernel functions (e.g., linear, polynomial, or radial basis function), which enables it to capture complex nonlinear relationships. The key advantages of SVM include its effectiveness in high-dimensional spaces, its ability to model nonlinear relationships, and its robustness against overfitting, especially when the number of observations is smaller than the number of features43. ANN model ANNs are computational systems modeled after the structure and operations of the human brain. These models are widely used to solve complex problems, including predicting SOC content. An ANN consists of several simple processing units known as neurons, organized into different layers: input, hidden, and output. Neurons in one layer are connected to those in the previous and following layers17. In general, when a machine learning model becomes too complex, especially when utilizing numerous environmental parameters for prediction, it may overfit the training data, leading to poor performance20. Therefore, before applying the machine learning models in this study, two methods (VIF and Boruta’s variable selection) were employed to identify the appropriate input parameters for all models. Model evaluation This study evaluated the effectiveness of machine learning models in predicting SOC content by utilizing the coefficient of determination (R²), root mean square error (RMSE), and relative percent difference (RPD)44(Eqs. 1–3).1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{R}}^{\mathrm{2}}\mathrm{=1-}\frac{\sum\:_{\mathrm{i}\mathrm{=1}}^{\mathrm{n}}{\mathrm{(}{\mathrm{O}}_{\mathrm{i}}\mathrm{-}{\mathrm{P}}_{\mathrm{i}}\mathrm{)}}^{\mathrm{2}}}{\sum\:_{\mathrm{i}}^{\mathrm{n}}{\mathrm{(}{\mathrm{O}}_{\mathrm{i}}\mathrm{-}\stackrel{-}{{\mathrm{O}}_{\mathrm{i}}}\mathrm{)}}^{\mathrm{2}}}$$\end{document}2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{RMSE=}\sqrt{\frac{1}{n}\sum\:_{\mathrm{i=1}}^{\mathrm{n}}{\mathrm{(}{\mathrm{P}}_{\mathrm{i}}\mathrm{-}{\mathrm{O}}_{\mathrm{i}}\mathrm{)}}^{2}}$$\end{document}3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{RPD=}\frac{\mathrm{S}\mathrm{D}}{\mathrm{RMSE}}\times\:100$$\end{document} where, n, Oi, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\stackrel{-}{{\mathrm{O}}_{\mathrm{i}}}$$\end{document}, and Pi denote the total number of sampling locations, the actual measured values, the average of the measured values, and the predicted values, respectively45. Furthermore, SD refers to the standard deviation. According to Eq. 3, the Relative Percent Difference (RPD) is categorized as follows. If RPD < 1, the accuracy is considered poor,If 2 < RPD ≥ ,1 the accuracy is medium,If 3 < RPD ≥ 2, the accuracy is satisfactory,If 4 < RPD ≥ 3, the accuracy is good,If RPD ≥ 4, the accuracy is excellent. If RPD < 1, the accuracy is considered poor, If RPD < 1, the accuracy is considered poor, If 2 < RPD ≥ ,1 the accuracy is medium, If 2 < RPD ≥ ,1 the accuracy is medium, If 3 < RPD ≥ 2, the accuracy is satisfactory, If 3 < RPD ≥ 2, the accuracy is satisfactory, If 4 < RPD ≥ 3, the accuracy is good, If 4 < RPD ≥ 3, the accuracy is good, If RPD ≥ 4, the accuracy is excellent. If RPD ≥ 4, the accuracy is excellent. Estimating uncertainty in predictions A fundamental aspect of DSM is the assessment of uncertainties associated with predictive models. In digital mapping, ensuring accuracy is crucial, but it is equally important to represent the inherent uncertainties within these models46. In this study, uncertainty maps for surface SOC content were generated using the 5th and 90th percentiles as lower and upper boundaries. The best-performing model, selected based on validation metrics, was used to predict SOC content. The final model was derived by averaging these predictions, with the 5th and 90th percentiles assigned to each pixel as uncertainty indicators. This methodology enhances the reliability of SOC content simulations across various landscapes. It equips researchers with both precise SOC distribution maps and a clear understanding of model uncertainties, thereby supporting more informed decision-making in soil management and environmental conservation47. According to Lu and Wang47, the process of estimating uncertainty is itself subject to uncertainty, particularly when based on a limited amount of soil data. Consequently, assessing the dependability of these uncertainty estimations is essential. In this study, the evaluation was conducted using the Prediction Interval Coverage Probability (PICP) metric. To be deemed reliable, an uncertainty estimate should have a 90% prediction interval corresponding to a PICP value of 90%. Additionally, the overall research methodology is depicted in Fig. 2.The spatial data processing and map generation in this study were conducted using SAGA-GIS (version 9.1.1; https://saga-gis.sourceforge.io) and ENVI (version 5.6; https://www.nv5geospatial.com/Software-Technology/ENVI). Additionally, certain maps were developed using R (version 4.3.1; https://www.r-project.org) and ArcGIS Desktop (version 10.8.1; https://www.esri.com). All statistical analyses were performed using IBM SPSS Statistics (version 28.0; https://www.ibm.com/spss). Results Statistical analysis of SOC content The statistical analysis of SOC content showed that it varied between 0.47% and 1.49%, with a mean value of 1.26% in the study area. Based on the classification from Vinhal-Freitas, et al. 48, these values place the SOC content within the moderate range. Furthermore, the calculated statistical parameters, such as variance (0.11), standard deviation (0.33), and coefficient of variation (25.69%), highlight the variability in SOC across the samples. The coefficient of variation, being relatively high, indicates considerable fluctuations in SOC levels within the region, which, according to Vieira and Paz Gonzalez49, places the variability within the moderate category. Optimal selection of environmental covariates for estimating SOC content In this study, we employed a two-step machine learning-based approach to identify the most relevant environmental covariates for accurately estimating SOC content. The first step involved evaluating the multicollinearity of the environmental covariates using the VIF index. Environmental covariates with high VIF values, greater than 10, were identified as collinear and subsequently removed to ensure that the remaining predictors were not redundant and would not distort the model’s performance. In the second step, the Boruta’s variable selection approach was applied to further refine the selection process. This method evaluates both the predictive ability and the relevance of each remaining environmental covariate. By doing so, we ensured that only the most influential variables, those with significant predictive power, were retained for the final model. This process facilitated the creation of an optimal predictor set that would result in a robust and reliable model for SOC estimation. Ultimately, from an initial pool of 40 environmental covariates, derived from Landsat 8 satellite images and DEM, five covariates were selected. These included the Normalized Difference Vegetation Index (NDVI), stand height (standh), Topographic Wetness Index (TWI), Curvature Index (CI), and Temperature Index (TE). These five variables were found to have the most significant relationship with SOC content, making them ideal for modeling SOC values using machine learning models such as RF, SVM, and ANN. Fig. 1The geographic location of Hunan Province, China, and the sampling points across the study area. Sampling point coordinates were collected during field surveys. The map was generated using ArcGIS Desktop (version 10.8.1; https://www.esri.com); background imagery was accessed via Google Earth (https://www.google.com/earth). The geographic location of Hunan Province, China, and the sampling points across the study area. Sampling point coordinates were collected during field surveys. The map was generated using ArcGIS Desktop (version 10.8.1; https://www.esri.com); background imagery was accessed via Google Earth (https://www.google.com/earth). Fig. 2Overall flowchart of this study for environmental parameter selection and SOC content estimation using selected parameters and machine learning methods. Overall flowchart of this study for environmental parameter selection and SOC content estimation using selected parameters and machine learning methods. Fig. 3The assessment metrics for training and test datasets consist of R² (a), RMSE (b), and PRD (c). The assessment metrics for training and test datasets consist of R² (a), RMSE (b), and PRD (c). Fig. 4Spatial distribution of SOC content predicted by machine learning methods, including RF (a), ANN (b), and SVM (c). The prediction models were developed in R (version 4.3.1; https://www.r-project.org), and the final maps were generated using ArcGIS (version 10.8.1; https://www.esri.com). The spatial data used in this figure were obtained from Google Earth Engine (https://code.earthengine.google.com/). Spatial distribution of SOC content predicted by machine learning methods, including RF (a), ANN (b), and SVM (c). The prediction models were developed in R (version 4.3.1; https://www.r-project.org), and the final maps were generated using ArcGIS (version 10.8.1; https://www.esri.com). The spatial data used in this figure were obtained from Google Earth Engine (https://code.earthengine.google.com/). Fig. 5Variable importance to predict the SOC content by the RF (a) and SVM models (b). Variable importance to predict the SOC content by the RF (a) and SVM models (b). Modeling SOC content After identifying the most relevant environmental covariates for predicting the SOC content, three machine learning models—RF, SVM, and ANN—were evaluated and compared for their performance in predicting SOC. The comparison revealed that the RF model significantly outperformed the other two models in terms of prediction accuracy. The R² values for the training data were 0.69 for SVM, 0.71 for ANN, and 0.84 for RF, whereas the R² values for the test data were 0.72, 0.75, and 0.85 for SVM, ANN, and RF, respectively (Fig. 3a). In addition to the R² values, further evaluation was carried out using the RMSE and the PRD indices. These indices demonstrated that the RF model consistently delivered superior results compared to the SVM and ANN models, as shown in Fig. 3b and c. The lower RMSE and PRD values for RF suggest a more precise prediction with less error and greater reliability. Considering the R², RMSE, and PRD results, it is evident that the RF model is the most appropriate choice for SOC prediction in this region. Its higher accuracy and better generalization capability make it the most reliable method for estimating SOC content. Based on the findings of this study, it is concluded that the RF method provides more accurate and consistent results compared to other methods like ANN and SVM. Spatial distribution of SOC content The spatial distribution maps of SOC content, generated using the RF, ANN, and SVM models, are presented in Fig. 4. While all three models exhibit comparable spatial trends, the findings highlight that the highest SOC content is predominantly found in the western, southern, and central parts of the study region. These areas exhibit significantly higher SOC content compared to other regions. Furthermore, the spatial analysis indicates that high SOC content is confined to a limited portion of the study area, whereas most regions have SOC content below 1%. This suggests that although certain areas exhibit relatively high SOC content, the majority of the land falls within moderate to low SOC levels, with most values remaining under the 1% mark. Influence of environmental covariates on SOC content prediction Figure 5 presents the relative importance (%) of various parameters for estimating SOC content using the RF and SVM models. The results show that the VIF values for the selected parameters in the RF model are as follows: standh at 0.72, NDVI at 0.68, TE at 0.50, TWI at 0.25, and CI at 0.18. In comparison, for the SVM model, the relative influence (%) of the parameters is as follows: TWI (0.68), TE (0.52), NDVI (0.50), standh (0.21), and CI (0.10). These results indicate that the most significant parameters for predicting SOC content with the RF model are stand height (standh) and NDVI, whereas for the SVM model, TWI and TE are the most influential parameters. Fig 6 illustrates the structure of the ANN used for SOC content estimation, including the input, hidden, and output layers of the model. Based on the ANN model, the most important environmental covariates for SOC content estimation are the CI and TWI indices. Although the importance of these environmental covariates differs across the three models, it is important to note that all selected covariates have an impact on SOC content. In arid and semi-arid regions, SOC content is largely influenced by natural vegetation cover and the amount of plant residues left in the soil. Moreover, changes in elevation can influence vegetation cover and soil moisture, both of which contribute to increasing SOC content. According to the RF model results, which demonstrated the highest accuracy, standh and NDVI are the most critical factors in estimating SOC content. Fig. 6An ANN model to estimate the SOC content by selected environmental covariates. An ANN model to estimate the SOC content by selected environmental covariates. Uncertainty analysis in predicting SOC content Following the modeling process for estimating SOC content, the results demonstrated that the RF model outperformed other approaches and was identified as the most effective method for SOC content prediction. Accordingly, the uncertainty associated with SOC content was also assessed using this model. Fig. 7 presents the uncertainty maps for SOC content predictions. In this analysis, uncertainty is represented by the lower (5%) and upper (90%) limits within a 90% confidence interval. A comparison of the spatial patterns of these bounds with the mean predictions shown in Fig. 4 reveals a strong correspondence between them. This indicates that areas with higher predicted SOC values generally exhibit greater levels of uncertainty. Another significant observation is the considerable gap between the upper and lower bounds in SOC estimation, highlighting the need for improving the accuracy of spatial predictions in this domain. Therefore, uncertainty analysis plays a crucial role in optimizing soil resource management and planning. Table 1 presents the PICP values for predicting SOC content using the RF model based on test data. The results indicate that the average PICP value is approximately 90%, demonstrating the high accuracy of the prediction intervals generated by the RF model. This suggests that the uncertainty estimates are generally reliable. Fig. 7Spatial distribution of uncertainty in SOC content at 90% (a) and 5% (b) limits (https://www.r-project.org). The prediction models were developed in R (version 4.3.1; https://www.r-project.org), and the final maps were generated using ArcGIS (version 10.8.1; https://www.esri.com). The spatial data used in this figure were obtained from Google Earth Engine (https://code.earthengine.google.com/). Spatial distribution of uncertainty in SOC content at 90% (a) and 5% (b) limits (https://www.r-project.org). The prediction models were developed in R (version 4.3.1; https://www.r-project.org), and the final maps were generated using ArcGIS (version 10.8.1; https://www.esri.com). The spatial data used in this figure were obtained from Google Earth Engine (https://code.earthengine.google.com/). Table 2Percentage values of PICP calculated based on test data (n=123).MethodDataPICP (%)RFTest (n=123)0.93 Table 2Percentage values of PICP calculated based on test data (n=123).MethodDataPICP (%)RFTest (n=123)0.93 Percentage values of PICP calculated based on test data (n=123). Discussion SOC content is a pivotal factor influencing agricultural sustainability and ecosystem stability. It plays a significant role in maintaining soil fertility, enhancing water retention, and contributing to carbon sequestration. Given its importance, the ability to accurately predict SOC content is essential for sustainable land management and climate change mitigation. Various modeling approaches have been developed to estimate SOC content, yet the selection of optimal environmental covariates and the most effective modeling techniques remain ongoing challenges26,50. In this study, we aimed to improve SOC content prediction by leveraging a refined selection of environmental covariates using VIF analysis and Boruta’s variable selection method. By integrating these selected variables, we implemented RF, ANN, and SVM models to assess SOC content. The principal objective was to enhance model efficiency by reducing data redundancy while maximizing predictive accuracy13,51. Our results underscored substantial variation in predictive performance across different models (Fig. 3). This variability arises due to multiple factors, including inherent differences in model architectures, spatial heterogeneity in soil properties, and inconsistencies in environmental covariates. Consequently, our findings confirm that no single model can universally outperform others in all contexts. Instead, model effectiveness is inherently tied to the choice of predictor variables and methodological framework9,19,24,30,52. This highlights the necessity of a systematic variable selection strategy to enhance model generalizability and robustness. Traditional SOC prediction studies often relied on empirical and pedological guidelines for selecting environmental covariates, leading to inconsistencies and reduced model reliability. The absence of a universally accepted standard for variable selection has historically hampered predictive accuracy and increased model uncertainty. Our study addresses this gap by implementing a rigorous variable selection protocol that offers two major advantages. First, it identifies the most significant predictor variables, optimizing model input. Second, it establishes a standardized approach for variable selection, enhancing model interpretability and reproducibility. However, despite these advancements, certain complexities remain unresolved due to the dynamic interactions between environmental covariates, underscoring the inherent limitations of data-driven approaches1. Keel, et al. 53 emphasized the critical role of variable selection in predictive modeling, acknowledging that some influential factors may remain uncontrolled, thereby necessitating careful consideration in model development22. In this study, we identified five key environmental covariates—NDVI, standh, CI, TE, and TWI—as the most influential variables in SOC prediction (Fig. 5). These variables not only exhibit clear correlations with SOC content but have also been consistently validated in previous studies9,20,29,31,50,54,55. Traditionally, many studies included all available environmental covariates in machine learning models, allowing the algorithm to determine the most relevant features. However, this practice often results in excessive model complexity, reduced interpretability, and increased uncertainty22. The results of this study showed that applying a variable pre-selection approach can improve prediction performance. Nevertheless, such outcomes are context-dependent and may vary with factors such as the extent and complexity of the study area and sampling density. The optimized set of variables used in this study enhanced prediction accuracy while maintaining computational efficiency. The necessity for this study stems from the critical role of SOC in ensuring soil fertility and agricultural sustainability, particularly in highly populated agricultural regions such as Hunan Province, China. Given the province’s dependence on agriculture, precise SOC prediction is essential for optimizing land use practices and mitigating soil degradation. The findings of this study provide a strong foundation for integrating machine learning-driven SOC prediction into precision agriculture frameworks. By demonstrating the efficacy of a structured variable selection process, our research offers a practical pathway for improving soil health assessments, formulating data-driven agricultural policies, and fostering sustainable environmental stewardship. Future studies should focus on integrating remote sensing technologies and hybrid modeling approaches to further refine SOC predictions, ensuring their applicability across diverse ecological landscapes. Conclusion The main objective of this study was to **identify and determine the most influential environmental covariates for accurately predicting soil organic carbon (SOC) content in Hunan Province, China. To achieve this, variable selection methods, including VIF and Boruta, were employed to pinpoint the most relevant covariates, thereby optimizing the predictive performance of machine learning algorithms. A comprehensive set of environmental covariates potentially affecting SOC content was evaluated, and the most significant variables were incorporated into predictive models, including RF, ANN, and SVM. The results demonstrated that integrating these selected covariates not only enhanced prediction accuracy but also reduced computational complexity, thereby improving the operational efficiency and robustness of the models. This approach provides a reliable and powerful framework for SOC estimation across large geographic scales and offers a solid foundation for the development and refinement of predictive models in future studies.