Spatial autocorrelation in machine learning for modelling soil organic carbon

Spatial autocorrelation, the relationship between nearby samples of a spatial random variable, is often overlooked in machine learning models, leading to biased results. This study compares various methods to account for spatial autocorrelation when predicting soil organic carbon (SOC) using random forest models. This kind of systematic comparison has not been done previously. Five models incorporating spatial structure were compared against baseline models with no added spatial components. Cross-validation showed slight improvements in accuracy for models considering spatial autocorrelation, while Shapley Additive Explanations confirmed the importance of spatial variables. However, no decrease in spatial autocorrelation of residuals was observed. Random Forest Spatial Interpolation emerged as the top performer in capturing spatial structure and improving model accuracy. Raster-based models exhibited enhanced prediction detail. The findings emphasize the value of incorporating spatial autocorrelation for better prediction of SOC with machine learning. Considerations such as the spatial distribution of predictions and computational complexity should help guide the selection of suitable approaches for specific spatial modelling tasks. • Five raster-based RF models incorporated unique spatial autocorrelation strategies. • Spatial ML strategies reduced residual spatial autocorrelation, improved performance. • RFSI captured spatial structure and improved model the most, followed by BD and XY models. • Raster-based models provided more detailed predictions than vector-based ones.