Effectiveness of Random Forest Model for Flash Flood Susceptibility in the Himalayan Region

Abstract

Machine learning (ML) approach is being increasingly recognized as vital tool for disaster risk reduction owing to its ability to address both the scale and the impact of a disaster. Risk of flash floods is currently a major problem across the Himalayan region. In present study we evaluated effectiveness of Random Forest (RF) model for flash flood susceptibility, based on real-world data in the Indian Central Himalaya Region (ICHR), where recurrent flash floods are being experienced every year. A geospatial dataset including 200 flash flood locations and eight conditioning factors namely elevation, slope, aspect, profile curvature, distance from river, annual rainfall, land use land cover (LULC) and lithology was used for performance evaluation of Random Forest model for flash flood susceptibility assessment. The effectiveness of the model is evaluated using area under the ROC curve (0.922), accuracy (0.925), precision (0.903), recall (0.921) and F-score (0.911) metrics. The results show that random forest could be an effective tool for flash flood susceptibility assessment in the Indian Central Himalayan Region. Furthermore, by considering optimum conditioning factors based on topographical, geological and hydrological conditions, the model can be used by managers and planners for flash flood management and sustainable conservation of the human society in the other Himalayan regions.