Spatial prediction of soil organic carbon: Combining machine learning with residual kriging in an agricultural lowland area (Lombardy region, Italy)

•All the models were able to produce SOC maps with mean values consistent with the actual mean SOC value.•Adding residual kriging to ML models yields marginal improved prediction accuracy of SOC.•VDCN yield information about age and evolution stage of soils.•Agricultural activities play an important role in the interpretation of the results.•Incorporating VDCN and CNBL is vital in future DSM for SOC in lowlands, reflecting their significant impact. Soil organic carbon (SOC) plays a crucial role in the global carbon cycle and in maintaining soil functions in the context of land use and climate change. Understanding the spatial distribution of SOC is essential for the management of agricultural land to optimize soil health and carbon storage. In this study, we investigated the spatial distribution of SOC in an agricultural lowland area of the Lombardy region, Italy, using machine learning (ML) techniques combined with residual kriging. ML models, including the artificial neural network (ANN), extreme learning machine (ELM), and random forest (RF), were trained on 120 SOC observations and eight environmental variables to predict SOC values across the study area. The performance of this ML approach was assessed using a ten-fold nested cross-validation process. The ELM and RF models showed better predictive performances based on the concordance correlation coefficient and root mean square error (RMSE), with RF slightly outperforming ELM based on the RMSE. The residuals of each iteration from the ML models were interpolated by ordinary kriging (OK) and added to the ML-based trend model in a hybrid regression-kriging approach. This approach which accounted for the spatial autocorrelation of the prediction residuals, resulting in a marginally improved prediction accuracy in the ML models. In addition, we found that vertical distance to the channel network and channel network base level are important predictor variables that should be considered in future digital soil models for SOC in lowland areas, given their importance in this study. Furthermore, this study highlights that predicted SOC values were low, particularly in Luvisols, which can be explained by the long history of agricultural land use depleting SOC due to agricultural management and loss of organic plant residues. The prediction maps depicted spatial variation and patterns of SOC in the study area. Our findings may help to refine soil management practices and contribute to improving soil health and c