Gully erosion susceptibility assessment using machine learning methods and geostatistical multivariate approach

The spatial mapping of areas prone to gully erosion is essential for planning control strategies related to soil conservation. This study aimed to evaluate machine learning methods for the modeling of gully erosion in the Baba-Arab region in Iran. To accomplish this, a support vector machine (SVM), maximum entropy (ME), artificial neural network (ANN), and their ensemble models were used, and susceptibility maps were produced. The 18 effective factors in gully erosion including distance to roads, elevation, slope degree, slope aspect, topographic wetness index (TWI), distance to rivers, land use, normalized difference vegetation index (NDVI), stream density, plan curvature, clay, silt, pH, organic matter (OM), Na, mean weight diameter (MWD), electrical conductivity (EC), and average annual rainfall were selected for the modeling process. Spatial variations in soil and climatic factors were studied using geostatistical methods. Factor interpolation was performed using inverse distance-weighted and kriging methods. The fractional variance-to-threshold ratio showed that pH, silt, and clay content had strong spatial structures, while EC, Na, OM, MWD, sand, and mean annual rainfall showed moderate spatial structures. The validation results of the models using the ROC curve considering 30% of the gully points showed that the models used had similar accuracy as SVM-ME with the highest performance. According to the prioritization of effective factors based on SVM, elevation and distance to roads were the most important factors in the occurrence of gully erosion in the study area. The results of spatial modeling of gully erosion can be of considerable help to land managers in identifying gullies promptly and in preventing the spread of gullies in the study area.