The impact of scale and variable selection for predicting topsoil organic carbon using machine learning and geostatistics based on legacy soil data

Soil organic carbon (SOC) is a significant soil property that informs agricultural management practices, resource planning, understanding land degradation, natural resource planning, and climate mitigation. In digital soil mapping (DSM), the spatial scale and the choice of covariates strongly influence model accuracy. These factors influence the degree of spatial autocorrelation as well as the capacity of the models to capture complex interactions. This study explores the impact of scale and choice covariates for predicting SOC in the topsoil at national and regional scales in South Africa. Random forest (RF) and regression kriging (RK) methods were employed along with the soil legacy data and environmental covariates, including long-term climate, organisms, soil variables, and static topographic variables. In addition, Boruta feature selection was used to identify a subset of the most important variables. Results showed that model performance varied with scale. SOC prediction at the national scale yielded slightly higher accuracy than that at the regional scale. At the regional scale, RF yielded an R² of 0.53, while RK yielded an R² of 0.50. Similarly, RF produced slightly lower RMSE and MAE compared to RK at the regional scale. In contrast, RK yielded slightly more accurate SOC predictions at the national scale than RF. RK yielded an R2 of 0.56 compared to 0.54 for RF at the national scale. The RMSE and MAE for RK were also lower compared to those of RF. Boruta identified 22 of 35 variables as relevant at the regional scale and 34 at the national scale. Feature selection slightly improved the predictions of SOC regional scale using RF, while that of RK decreased slightly. However, at the national scale, the performance of models remained unchanged. Regarding covariate importance, shortwave radiation, evapotranspiration, precipitation, maximum temperature, and runoff were identified as the most influential at the regional level. On the other hand, climate, soil variables and vegetation indices (NDVI and soil brightness index) were identified as the most influential at the national level. These findings highlight the critical role of scale and covariate selection for the accurate prediction of SOC in the topsoil using DSM.