Data supplement: Spatial autocorrelation in machine learning for modelling soil organic carbon

Spatial autocorrelation in machine learning for modelling soil organic carbon: Data supplement

Alexander Kmoch, Clay Taylor Harrison, Jeonghwan Choi, Evelyn Uuemaa

Spatial autocorrelation, the relationship between nearby samples of a spatialrandom variable, is often overlooked in machine learning models, leading tobiased results. This study investigates various methods to account for spa-tial autocorrelation when predicting soil organic carbon (SOC) using randomforest models. Five models incorporating spatial structure were comparedagainst baseline models that did not have any added spatial components.Cross-validation showed slight improvements in accuracy for models consid-ering spatial autocorrelation, while Shapley Additive Explanations confirmedthe importance of spatial variables. However, no decrease in spatial autocor-relation of residuals was observed. Raster-based models exhibited enhancedprediction detail, but high-resolution validation data availability limited thor-ough validation. The findings emphasize the value of incorporating spatialautocorrelation for improved SOC prediction in machine learning models.Considerations such as the distribution of predictions and computationalcomplexity should help guide the selection of suitable approaches for specificspatial modelling tasks.