A comparative evaluation of statistical and machine learning approaches for debris flow susceptibility zonation mapping in the Indian Himalayas

Spatial prediction of debris flows in terms of susceptibility mapping is the first and foremost requirement for disaster mitigation. In the present study, a comparative evaluation of machine learning and statistical approaches for debris flow susceptibility zonation (DFSZ) mapping has been attempted using 10 causative thematic layers (slope, aspect, elevation, plan curvature, profile curvature, topographic wetness index, stream power index, geology, proximity to streams, normalized difference vegetation index) and a debris flow inventory containing 85 debris flow locations. The employed machine learning (ML) approaches include random forest (RF), naïve Bayes (NB), and extreme gradient boosting (XGBoost) models whereas statistical models include the weight of evidence (WoE) and index of entropy (IoE). The results indicated that in all 5 DFSZ maps, about 21.20–47.98% of the area is very highly and highly susceptible to debris flows. It is observed that the major debris flows as well as high susceptible zones are distributed along the river Alakananda and its tributaries and at the vicinity of the NH-58. Among the statistical models, the DFSZ map prepared using the weight of evidence (WoE) model provides higher accuracy in terms of the success rate and the prediction rate compared to that prepared using the index of entropy model (IoE). Among the machine learning-based models, both the extreme gradient boosting (XGBoost) and random forest (RF) models give better accuracy and are more efficient than the Naïve Bayes (NB) model. It is also observed that the ML models perform better than the statistical models for a part of Chamoli district, Uttarakhand state (India). The novelty of the present study lies in the spatial prediction of one of the most destructive forms of mass movement (debris flow) in the Indian Himalayas using statistical and ML models and establishing the superiority of the ML Random Forest & XGBoost model over other ML and statistical models for the present case. This study will help make decisions on the suitability of developmental activities and human settlement in the area under consideration. The present study is one among the few studies focused on the DFSZ mapping in Indian Himalayas and can be replicated in other debris flow prone regions worldwide.