Assessing the performance of MCDM, statistical and machine learning ensemble models for gully sensitivity mapping in a semi-arid context

Abstract

Gully erosion is a complex socio-environmental issue that has a negative influence on natural resources and has significant economic costs. This study examined the performance of two ensemble models based on multicriteria decision making (MCDM) analysis, analytic hierarchy process (AHP), weight of evidence (WoE) and random forest (RF) for spatiotemporal monitoring of gully erosion sensitivity (GES) from 1988 to 2019 as well as a projection for 2040 in a semi-arid area. The findings revealed that the vulnerable areas significantly raise between 1988 and 2040 (> 27% of the study area since 2019), in perfect alignment with a rapid deterioration of the vegetation cover (−16%), a general decrease in rainfall (−25% since 2019), and an increase in land surface temperature (LST) average (30°–37° approximatively). Finally, the area under curve (AUC) value revealed a high prediction performance for both developed models (AUC = 0.888 for WoE-RF and 0.886 for MCDM-WoE-AHP).