Flood susceptibility mapping using supervised machine learning models: insights into predictors’ significance and models’ performance

Abstract

Eastern Hindu Kush (EHK) is one of the most flood-prone regions due to its diverse topographic features, complex climatic conditions, and fragile socioeconomic situations. Yet there are limited studies on robust assessment and prediction of flood susceptibility in this region. This study aims to predict flood susceptibility hotspots, identify significant flood predictors, and evaluate the models’ performance in the transboundary Kabul River Basin (KRB) of the EHK region. The study employs a set of six Supervised Machine Learning (SML) models, namely Logistic Regression (LR), Artificial Neural Network (ANN), eXtreme Gradient Boost (XGBoost), Random Forest (RF), K-Nearest Neighbors (KNN), and Naïve Bayes (NB), along with sixteen topographical, hydrological, vegetational, and environmental predictors and a flood inventory of 570 flooded and non-flooded locations each. Among the selected SML models, XGBoost demonstrated the highest performance, followed by RF, outperforming the rest of the SML models. The outcomes of the LR, ANN, XGBoost, RF, KNN, and NB models indicate that southern, southeastern, and stream-adjacent regions are susceptible to flooding. These findings provide a robust prediction of flood susceptibility in the transboundary KRB. This can help the relevant authorities strengthen the existing early warning systems, implement mitigation strategies, and foster community resilience.