Developing high-resolution soil organic carbon maps across Ethiopia through a machine learning approach

Accurately mapping the spatial distribution of Soil Organic Carbon (SOC) is crucial for understanding ecological processes, including nutrient and carbon cycling, food security, and agricultural productivity. Despite their importance, a national- and regional-scale SOC map is lacking in Ethiopia. Thus, this study addresses this gap by developing high-resolution national maps of SOC using a machine learning (ML) based digital soil mapping approach. Three ML models, Random Forest (RF), Gradient Boosting (GB), and Extreme Gradient Boosting (XGB), were applied to predict the SOC content of 0 to 20 cm depth at a 100-m spatial resolution. Predictor variables, including topographic and climatic data, Sentinel-2 spectral bands, and derived vegetation and moisture indices, were used to upscale a digital SOC map. The RF model produced the highest predictive accuracy among the three models, with performance metrics including a coefficient of determination (R2) of 0.57, root mean square error (RMSE) of 0.65, mean absolute error (MAE) of 0.49, bias of 0.01, and Nash–Sutcliffe Efficiency (NSE) of 0.57, outperforming the GB and XGB models. The results of variable importance revealed that precipitation, green normalized difference vegetation index (GNDVI), temperature, shortwave infrared (SWIR-1), moisture stress index (MSI), and enhanced vegetation index (EVI) were the best predictors for the predictive models. Spatial analysis revealed higher SOC concentrations in the southeastern grasslands, central highlands, and southwestern mountainous regions, particularly within forested and grassland land covers. This study provides insight into Ethiopia's national digital soil map products, offering a baseline for future monitoring of land use and climate change impacts.