Geographically weighted regression random forest for modeling soil particles

Clay particles play a vital role in determining soil quality, particularly in the fields of agriculture and conservation. However, the complex and non-linear spatial distribution of clay particles is difficult to capture using conventional modeling methods. This study develops a hybrid model, Geographically Weighted Regression Random Forest (GWRRF), which combines the ability of Geographically Weighted Regression (GWR) to capture spatial heterogeneity with the strength of Random Forest (RF) in handling non-linear relationships. The data used in this study were derived from soil texture and local morphologic analysis across 50 observation points in the Kalikonto watershed. The results indicate that the GWRRF model achieved a higher explanatory power (R² = 0.735) compared to the conventional GWR model (R² = 0.475), demonstrating its better capability in capturing complex spatial variability. However, the RMSE value of the GWRRF model (4.314) was slightly higher than that of the GWR model (3.485), reflecting a trade-off between model flexibility and prediction accuracy. Overall, the integration of GWR and Random Forest in the GWRRF framework provides a more adaptive and context-aware approach for analyzing spatial heterogeneity in clay particle distribution, offering valuable insights for data-driven and sustainable land management practices.