Pixel-scale gully erosion susceptibility: Predictive modeling with R using gully inventory consistent with terrain variables

This study develops a reproducible methodology for gully erosion susceptibility assessment using Multivariate Adaptive Regression Splines (MARS) in the Turkey Creek basin, Central Kansas (USA). MARS models were trained on two predictor sets (A and B) extracted from the Digital Elevation Model (DEM) and ten gully grids derived from a gully inventory. Set A included predictors independent of the catchment area (e.g., slope angle, plan curvature), while set B added catchment area-related variables (e.g., stream order, wetness index). Gully grids were created by snapping digitized gully pixels to DEM flow lines by varying snapping distances and catchment area thresholds. Cross-validation across 20 square zones revealed significant performance improvements with snapped gully data and set B predictors, as measured by AUC and Cohen’s kappa . The modeling framework, supported by open-source R code, offers a valuable tool for erosion susceptibility studies in regions where DEM and gully inventory data are available. • Including catchment area-dependent variables as predictors significantly improved model’s ability to identify gully-prone areas. • A new methodology for automatic snapping of gully pixels to DEM-derived flow lines elevated the precision of gully susceptibility mapping. • A provided open-source R code presents a reproducible framework of gully mapping for environmental planning across diverse landscapes.