Factors affecting topographic thresholds in gully erosion occurrence and its management using predictive machine learning models

Soil degradation induced by gully erosion represents a worldwide problem in the many arid and semi-arid countries, such as Iran. This study assessed: (1) the importance of variables that control gully erosion using the Boruta algorithm, (2) the relationship among causative variables and gullied locations using the evidential belief function model (EBF), and (3) gully erosion development using the algorithms of boosted regression tree (BRT) and support vector machine (SVM). Based on the results of the Boruta algorithm, slope, land use, lithology, plan curvature, and elevation were the most important factors controlling gully erosion. The results of the EBF model showed the predominance of gully erosion on rangeland and loess-marl deposition. The predominance of gullied locations on the concave positions, with the slope of 5°–20° in the vicinity of drainage lines, illustrates a preferential topographic zone and, therefore, a terrain threshold for gullying. The correlation of gullied locations with rangelands and weak soils in concave positions demonstrates that the interactions among soil characteristics, topography, and land use stimulate a low topographic threshold for gullies development. These relationships are consistent with the threshold concept that a given soil, land use, and climate within a given landscape encourage a given drainage area and a critical soil surface slope that are necessary for gully incision. Furthermore, the BRF-SVM had the highest efficiency and the lowest root mean square error, followed by BRT for predicting gully development, compared with LN-SVM algorithm. The application of two machine learning methods for predicting the gully head cut susceptibility in northern Iran showed that the maps generated by these algorithms could provide an appropriate strategy for geo-conservation and restoration efforts in gullying-prone areas.