A methodological comparison of head-cut based gully erosion susceptibility models: Combined use of statistical and artificial intelligence

A GIS-based hybrid approach for gully erosion susceptibility mapping (GESM) in the Biarjamand watershed in Iran is presented. A database comprised of 15 geo-environmental factors (GEFs) was compiled and used to predict the spatial distribution of 358 gully locations; 70% (251) of which were extracted for training and 30% (107) for validation. A Dempster-Shafer (DS) statistical model was employed to map susceptibility. Next, the results of four kernels (binary logistic, reg logistic, binary logitraw, and reg linear) of a boosted regression tree (BRT) model were combined to increase the efficiency and accuracy of the mapping. Area under receiver operating characteristics (AUROC), true skill statistic (TSS) and efficiency (E) metrics were used to rank the five validated models. The results show that elevation and distance to road play crucial roles in gullying. Integrating BRT and DS enhanced prediction accuracy. Among the four BRT kernels, binary logistic performed best (AUROC of 0.886, TSS of 0.854 and E equal to 0.880). The worst results were produced by the individual DS model (AUROC = 0.849, TSS = 0.774 and E = 0.834). The hybrid binary logistic-BRT and DS map categorized 14.50% of the study area as having very-low susceptibility, 16.99% low susceptibility, 22.77% moderate susceptibility, 24.12% high susceptibility, and 21.59% very-high susceptibility.