Gully headcut susceptibility modeling using functional trees, naïve Bayes tree, and random forest models

Abstract

Abstract Gully headcuts are due to erosion generated by concentrated overland flow, non-uniform infiltration, the presence of impermeable sub-surface soil layers, and a hydraulic gradient. It has been seen in an extensive range of continuous and categorical conditioning factors in several countries. Researchers fail to determine the factor that is the most important cause of GHs occurrence. This study determines the morphological and physio-chemical soil features, the locations with significant GHs, and the soil and geomorphic parameters in the loess area of Golestan Province, Iran. The spatial distribution of GHs was determined by field surveys and analysis of 1\u202fm-resolution aerial photography collected with an UAV. Physical and chemical features of soil samples were determined through lab analysis. The relationships between features and GHs were ranked by statistical significance using chi-square tests. Functional trees - (FT), naive Bayes tree (NBTree), and random forest (RF) models were employed to generate GHs sensitivity maps. One-hundred and twenty-seven GHs were identified and randomly divided into a training dataset comprised of 70% (89) of the GHs, and a validation dataset encompassing 30% (38) of the GHs. Twenty-two GHs conditioning factors were input into the models. The results of the chi-square analysis confirmed land use (90.8%) had the greatest influence on GHs occurrence. The results of training indicated that the AUC value for the FT, NBTree, and RF models were 83.4%, 94.8%, and 96.5%, respectively. The confidence interval (CI) for these methods was 0.771–0.885, 0.905–0.975, and 0.926–0.987, respectively. The highest standard error (SE) was related to the FT model (0.0314), followed by NBTree (0.0151), and RF (0.0119). The validation results indicated an AUC of 0.888, 0.947, and 0.981 for these methods, respectively. Therefore, RF was found to be the most effective model for predicting and mapping GHs in the future.