Flood susceptibility zonation using advanced ensemble machine learning models within Himalayan foreland basin

Abstract

Floods are considered as one of nature's most destructive fluvio-hydrological extremes because of the massive damage to agricultural land, roads and buildings and human fatalities. Rapid development of unplanned infrastructural conveniences and unplanned anthropogenic activities, the frequency and intensity of floods have been accelerated in recent years. Therefore, flood susceptibility analysis is considered as an important flood management approach. Identification of flood susceptibility areas has been performed by applying advanced machine learning (ML) algorithms (random forest (RF), support vector machine (SVM) and extreme gradient boosting (XGBoost)) at the lower part of Raidak river basin. The flood susceptibility maps have been generated based on 14 different flood conditioning factors. Models are evaluated in a conventional way using ROC (receiver operating characteristics) curve. The AUC value of ROC is above 0.80 for all models and XGBoost depicts the highest efficacy (AUC ​= ​0.92). Friedman test and Wilcoxon Signed rank test have been used to measure the statistical variances among the applied models. Models proficiently show that the upper part of Raidak river basin is a less flood probable region whereas the eastern and some middle parts have high flood probability. Around 27% area (285.39 sq.km) within the river basin is highly flood prone (based on XGBoost model) due to the fast changing dynamic landscape and large scale human intervention. The important outcomes of this research will definitely assist the local administrators to take proper sustainable management plans for the reduction of future damages.

Keywords

• Flood susceptibility
• Machine learning (ML) algorithms
• Human intervention
• Wilcoxon signed rank test