A Comparison of Different Artificial Intelligence and Machine Learning Methods for Gully Erosion Susceptibility Mapping in the Upper Narmada Basin

Gully erosion (GE) is one of the most important mechanisms of soil loss worldwide. In this study, various machine learningMachine learning techniques, such as classification and regression trees (CART), random forest (RF), and artificial neural networks (ANN), have been used to ascertain gully erosion susceptibilitySusceptibility (GES) in the Upper Narmada Basin (UNB). The mapping and analysis were achieved using R programming and ArcGISArcGIS 10.8 software. Initially, a gully inventory map (GIM) of 1501 gully locations was prepared from Sentinel-2 and Google Earth imagesGoogle Earth Image and extensive field surveys. Out of the 1501 gullies in the study area, 1051 gully locations (about 70%) were used for training, and 450 gully locations (about 30%) were used for validating the modelsModels. For GES modeling, 12 gully conditioning factors (GCFs) were used, and the relationships between these GCFs and gully erosion were evaluated. The GES maps were prepared using the CART, RF, and ANN modelsModels and divided into three susceptibilitySusceptibility-based classes: low, moderately, and highly susceptible GE classes. A large part of the study area was highly susceptible to GE. Subsequent validation tests proved the high efficacy of these modelsModels in ascertaining the GES. The RF modelModels performed best compared to the others in this respect, with an AUC-ROC value of 0.78. Therefore, this modelModels can be used in the UNB and other such areas to evaluate the GES zones and thereby aid in framing suitable measures to mitigate soil loss through gully erosion.