Spatial distribution and modeling of soil organic carbon using machine learning and a geostatistical approach
6 affiliations
- 1Department of Natural Resource Management, Debre Tabor University
- 2Department of Forest Science, Luiz de Queiroz College of Agriculture, University of São Paulo
- 3Department of Plant Sciences, Debre Tabor University
- 4Department of Natural Resource Management, Injibara University
- 5Department of Soil Science, ESALQ, University of São Paulo
- 6Department of Agricultural and Environmental Engineering, School of Engineering, Fluminense Federal University
Abstract
Reliable soil organic carbon (SOC) estimation is crucial for ensuring agricultural production and environmental amenity. Under data-scarcity conditions, geospatial modeling is a suitable tool for mapping and managing SOC. This study evaluated the spatial variability of SOC and compared the predictive performance of five modelling approaches: Linear Regression (LR), Random Forest (RF), Support Vector Machine (SVM), Artificial Neural Network (ANN), and Extreme Gradient Boosting (XGBoost) using environmental covariates, including temperature, soil texture, drainage, soil depth, erosion, altitude, soil type, and land use, in the Fogera Plain, northwestern Ethiopia. Descriptive statistics showed that grassland contained the highest mean SOC (1.43%), followed by shrubland (1.40%) and cultivated land (1.35%), indicating significant land-use effects on carbon storage. Correlation analysis identified soil erosion (r = − 0.756, p < 0.01) as the dominant negative driver of SOC, whereas soil texture exhibited a strong positive relationship (r = 0.501, p < 0.01). Variable importance analysis consistently ranked erosion and soil texture as the most influential predictors across all machine-learning models. Among the evaluated algorithms, SVM produced the highest predictive accuracy (R2 = 0.936), outperforming LR (R2 = 0.925), RF (R2 = 0.851), XGBoost (R2 = 0.842), and ANN (R2 = 0.774). Semivariogram analysis indicated weak spatial dependence of SOC (nugget/sill = 79.07–91.02%), with the exponential model providing the best geostatistical fit (R2 = 0.88). Spatial prediction revealed higher SOC concentrations in the western and southwestern parts of the landscape and lower values in the northern and southeastern areas. These findings demonstrate that integrating machine-learning techniques, particularly SVM, with environmental covariates provides an effective framework for accurate SOC prediction and uncertainty assessment. The resulting high-resolution SOC maps provide valuable information for site-specific soil management, carbon accounting, erosion mitigation, and sustainable land-use planning in tropical agricultural landscapes. Future research should integrate high-resolution remote sensing data, long-term field monitoring, additional environmental covariates, and ensemble deep-learning approaches to improve the transferability, temporal prediction, and uncertainty quantification of SOC models under changing climate and land-use conditions.
Related
- Modeling of fracture intensity by integrating geostatistical and neural network methods: The asmari reservoir, Rag-e-Sefid Oilfield, SW Iran
- Molecular-Level Insights into the Spatial Distribution, Transformation, and Potential Sources of Dissolved Organic Nitrogen in the Beidagang Reservoir
- Spatial variability of soil fertility parameters in Kaski District, Nepal: A geostatistical approach
- Gravity topography modeling of the Denman Glacier region using a geostatistical approach
- Spatial distribution, origins, and driving mechanisms of groundwater nitrate contamination in a rural-agricultural-mining catchment: Insight from PMF, machine learning and multi-isotopic techniques
BibTeX
@article{tiruneh2026spatial,
title = {{Spatial distribution and modeling of soil organic carbon using machine learning and a geostatistical approach}},
author = {Gizachew Ayalew Tiruneh and Ciro Abbud Righi and Tiringo Yilak Alemayehu and Genetu Fekadu and Simeneh Demissie and Marcelo H. Procópio Pelegrino and Carlos Rodrigues Pereira},
journal = {Discover Sustainability},
year = {2026},
month = jul,
doi = {10.1007/s43621-026-04201-w},
url = {https://doi.org/10.1007/s43621-026-04201-w},
note = {CC BY},
}
RIS
TY - JOUR
TI - Spatial distribution and modeling of soil organic carbon using machine learning and a geostatistical approach
AU - Gizachew Ayalew Tiruneh
AU - Ciro Abbud Righi
AU - Tiringo Yilak Alemayehu
AU - Genetu Fekadu
AU - Simeneh Demissie
AU - Marcelo H. Procópio Pelegrino
AU - Carlos Rodrigues Pereira
PY - 2026
DA - 2026/07/01
JO - Discover Sustainability
DO - 10.1007/s43621-026-04201-w
UR - https://doi.org/10.1007/s43621-026-04201-w
AB - Reliable soil organic carbon (SOC) estimation is crucial for ensuring agricultural production and environmental amenity. Under data-scarcity conditions, geospatial modeling is a suitable tool for mapping and managing SOC. This study evaluated the spatial variability of SOC and compared the predictive performance of five modelling approaches: Linear Regression (LR), Random Forest (RF), Support Vector Machine (SVM), Artificial Neural Network (ANN), and Extreme Gradient Boosting (XGBoost) using environmental covariates, including temperature, soil texture, drainage, soil depth, erosion, altitude, soil type, and land use, in the Fogera Plain, northwestern Ethiopia. Descriptive statistics showed that grassland contained the highest mean SOC (1.43%), followed by shrubland (1.40%) and cultivated land (1.35%), indicating significant land-use effects on carbon storage. Correlation analysis identified soil erosion (r = − 0.756, p < 0.01) as the dominant negative driver of SOC, whereas soil texture exhibited a strong positive relationship (r = 0.501, p < 0.01). Variable importance analysis consistently ranked erosion and soil texture as the most influential predictors across all machine-learning models. Among the evaluated algorithms, SVM produced the highest predictive accuracy (R2 = 0.936), outperforming LR (R2 = 0.925), RF (R2 = 0.851), XGBoost (R2 = 0.842), and ANN (R2 = 0.774). Semivariogram analysis indicated weak spatial dependence of SOC (nugget/sill = 79.07–91.02%), with the exponential model providing the best geostatistical fit (R2 = 0.88). Spatial prediction revealed higher SOC concentrations in the western and southwestern parts of the landscape and lower values in the northern and southeastern areas. These findings demonstrate that integrating machine-learning techniques, particularly SVM, with environmental covariates provides an effective framework for accurate SOC prediction and uncertainty assessment. The resulting high-resolution SOC maps provide valuable information for site-specific soil management, carbon accounting, erosion mitigation, and sustainable land-use planning in tropical agricultural landscapes. Future research should integrate high-resolution remote sensing data, long-term field monitoring, additional environmental covariates, and ensemble deep-learning approaches to improve the transferability, temporal prediction, and uncertainty quantification of SOC models under changing climate and land-use conditions.
SN - 2662-9984
ER -