Comparative assessment using boosted regression trees, binary logistic regression, frequency ratio and numerical risk factor for gully erosion susceptibility modelling

Abstract The initiation and development of gullies as worldwide features in landscape have resulted in land degradation, soil erosion, desertification, flooding and groundwater level decrease, which in turn, cause severe destruction to infrastructure. Gully erosion susceptibility mapping is the first and most important step in managing these effects and achieving sustainable development. This paper attempts to generate a reliable map using four state-of-the-art models to investigate the Bayazeh Watershed in Iran. These models consists of boosted regression trees (BRT), binary logistic regression (BLR), numerical risk factor (NRF) and frequency ratio (FR), which are based on a geographic information system (GIS). The gully erosion inventory map accounts for 362 gully locations, which were randomly divided into two groups (70% for training and 30% for validation). Sixteen topographical, geological, hydrological and environmental gully-related conditioning factors were selected for modelling. The threshold-independent area under receiver operating characteristic (AUROC) and seed cell area index (SCAI) approaches were used for validation. According to the results of BLR and BRT, the conditioning parameters namely, NDVI and lithology, played a key role in gully occurrence. Validation results showed that the BRT model with AUROC = 0.834 (83.4%) had higher prediction accuracy than other models, followed by FR 0.823 (82.3%), NRF 0.746 (74.6%) and BLR 0.659 (65.9%). SCAI results indicated that the BRT, FR and BLR models had acceptable classification accuracy. The findings, in terms of model and predictor choice, can be used by decision-makers for hazard management and implementation of protective measures in gully erosion-prone areas.