Spatial Prediction and Mapping of Gully Erosion Susceptibility Using Machine Learning Techniques in a Degraded Semi-Arid Region of Kenya

This study aimed at (i) developing, evaluating and comparing the performance of support vector machines (SVM), boosted regression trees (BRT), random forest (RF) logistic (LR) models in mapping gully erosion susceptibility, (ii) determining important conditioning factors (GECFs) a Kenyan semi-arid landscape. A total 431 geo-referenced points were gathered through field survey visual interpretation high-resolution satellite imagery on Google Earth, while 24 raster-based GECFs retrieved from existing geodatabases for spatial modeling prediction. The resultant exhibited excellent performance, although machine learners outperformed benchmark LR technique. Specifically, RF BRT returned highest area under receiver operating characteristic curve (AUC = 0.89 each) overall accuracy (OA 80.2%; 79.7%, respectively), followed by SVM 0.86; 0.85 & OA 79.1%; 79.6%, respectively). In addition, importance varied among models. best-performing model ranked distance to stream, drainage density valley depth as three most region. output susceptibility maps can efficient allocation resources sustainable land management area.