Flood susceptibility mapping using remote sensing and ensemble machine learning in Upper Jhelum Sub-catchment, India

Flood being the most devastating natural hazard has caused loss to human lives and damage to infrastructure globally. This study makes an attempt to assess flood susceptibility in Upper Jhelum Sub-catchment, India. Effectiveness of multilayer perceptron (MLP), sequential minimal optimization (SMOreg), M5P and bagging ensembles (B-MLP, B-M5P, B-SMOreg) models were utilized for achieving accurate results of flood susceptibility. The results revealed that B-MLP model was found to be a more effective model based on assessors and validation (ROC-AUC: 0.936; correlation coefficient: 0.925; precision: 0.906; F1-score: 0.887; accuracy: 0.863; root mean square error: 0.043; relative absolute error: 2.078 and mean absolute error: 0.013). Thus, this paper integrates bagging (B-MLP) with geospatial analysis in the watersheds of the Sub-catchment for the prediction of flood susceptibility mapping. The results revealed high flood susceptibility was found in the watershed 1E1D2 (Rembaira) followed by 1E1D3 (Vishav) and 1E1D4 (Lidder). The watershed 1E1D3 (Vishav) experienced largest area under moderate flood susceptibility followed by 1E1D4 (Lidder) and 1E1D2 (Rembaira). The findings may help in enhancing flood prediction and devising effective management strategies. B-MLP model can effectively be utilized for flood susceptibility mapping in other geographical regions at spatial scales.