Tuning spatial parameters of Geographical Random Forest: the case of agricultural drought

Machine learning algorithms are widely used methods in geographical research. However, these algorithms are not properly exploiting the underlying spatial relationships present in the geographical data. One of the approaches, which addresses this problem, is based on an ensemble of local models, which are constructed from samples in close proximity to the location of prediction. This concept was applied to the Random Forest (RF) algorithm, creating a Geographical Random Forest (GRF). This study aims to further develop GRF by tuning the spatial parameters for each location in case of agricultural drought. In addition to tuning, the explanatory property of RF within the framework GRF is explored. Four machine learning models were constructed; regular RF, regular RF with spatial covariates, GRF, and GRF with the tuning of spatial parameters. Models were evaluated using Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE). Although the decrease in RMSE in this very case is relatively small, the method may provide higher improvement with different datasets.