Remote sensing and GIS-based machine learning models for spatial gully erosion prediction: A case study of Rdat watershed in Sebou basin, Morocco

Abstract

The Rdat watershed is part of the Sebou basin, one of the most hydrological units in Morocco, in which agricultural activity is developed and constitutes an important mode of land use pattern. The expansion of urbanization, population growth, climate change, drought and water resource scarcity are making the land more prone to erosion, where gully erosion is the dominant driver of soil loss and agricultural land degradation. Hence, three machine learning (ML) algorithms were used to predict the gully erosion susceptibility (GES) in the Rdat watershed. Afterwards, gully erosion locations were collected and 16 conditioning factors of gully erosion were selected including topographic, hydrologic, environmental and geologic features. The results of a prediction models were compared and validated using accuracy (AC), precision, and area under receiver operating characteristics curve (AUC). The precision, AC and AUC value, respectively, for the Random Forest (RF) model were 88.5%, 85.6% and 88.4%, whereas for Boosted Regression Trees (BRT) model were 86.6%, 83.2% and 87.9%, while for Support Vector Machine (SVM) model were 78.5%, 82.6% and 85.1%. This finding indicates that the RF model is the most efficient in mapping the GES. Indeed, about 33% of the Rdat watershed is subject to gully erosion at high to very high level, indicating that gullies are more susceptible to develop in this area. Elevation, land use/cover, and drainage density are also indicated to be the most effective factors in this area for increasing gully erosion. Thus, most gullies are located in downstream with low elevation, bare and agricultural lands. The predicted gully erosion map can be an effective support to help decision makers in implementing appropriate soil and water management measures.