Characterizing recurrent flood hazards in the Himalayan foothill region through data-driven modelling

Abstract

Floods are frequently occurring hazards in the Himalayan foothills that primarily affect the physical environment and socio-economic profile of the local people. To this end, we used some data-driven models and techniques such as multi-criteria decision analysis, bivariate statistical method, machine learning, and sub-watershed prioritization by morphometric analysis (SWPMA) in the Kaljani River basin for the flood susceptibility mapping (FSM). The study used 200 non-flood points and 200 historical flood points for developing the testing (30%) and training (70%) datasets. Multicollinearity diagnostic test (VIF) and information gain ratio (IGR) were used to measure the consistency of the selected twelve influencing factors. Rainfall, drainage density, clay content, distance from the river and elevation were the most important drivers of the selected influencing parameters. We found that the high and very high FSM areas of the basin were around 32.24% for random forest (RF), 47.46% for frequency ratio (FR), 46.41% for analytical hierarchic process (AHP), and 40.5% for SWPMA. Most of these areas lie in Alipurduar town surrounding areas and the lower part of the basin. The ROC (receiver operating characteristic), MAE (mean absolute error), RMSE (root mean square error), Kappa index and Pearson’s correlation coefficients were employed for the performance efficiency of four data-driven models. Achieved ROC values of the RF model, FR model, AHP model and SWPMA model were at 92.4%, 86.6%, 83.1% and 79.4% respectively. Compared to all four methods the machine learning approach, the RF method performed most accurately. This study can help researchers for choosing the right approaches to the FSM. The present FSM may be instrumental before the various stakeholders while designing the flood evacuation plan, preparation of future flood shelters, and planning for future development for the local community.