Flash flood susceptibility modeling using optimized deep learning method in the Uttarakhand Himalayas

Abstract

Flash floods, which are influenced by hydro-meteorological conditions, are increasingly being triggered in the Uttarakhand Himalayas due to human-induced environmental and climatic changes. The catchment areas of Himalayan rivers in Uttarakhand experience flash floods every year, primarily caused by heavy rainfall and glacial lake outburst floods (GLOFs). Despite the severity of the issue, few studies have used optimized deep learning methods for robust flash flood susceptibility modeling (FFSM) and mapping. Therefore, this study aims to introduce a novel approach for FFSM using optimized deep learning (DL) models and to identify the most influential factors for prediction using SHapley Additive exPlanations (SHAP) and Partial Dependence Plots (PDPs). In this study, two optimized DL models, namely the Deep Neural Network (DNN) and Convolutional Neural Network (CNN), were trained for FFSM. A spatial database was constructed using 320 past flash flood and non-flash flood sample and twelve flash flood influencing factors: Elevation, Slope, Curvature, Normalized Difference Vegetation Index (NDVI), Topographic Ruggedness Index (TRI), Stream Power Index (SPI), Land Use and Land Cover (LULC), Distance from River, Drainage Density, Topographic Wetness Index (TWI), Annual Rainfall, and Geology. The predictive performance of the models was validated and compared using statistical evaluation metrics, including the Receiver Operating Characteristic (ROC) curve, Precision-Recall Curves (PRC), accuracy, precision, recall, and F1 score. The results show that 4 to 6% of the areas were predicted as being in the very high flood susceptibility zone in both models, demonstrating high accuracy with strong areas under the curve (AUC) for both the ROC and PRC. The DNN model achieved an AUC of 0.91 and 0.94 for prediction, with accuracy, precision, recall, and F1 scores of 0.8265, 0.8723, 0.7885, and 0.8283, respectively. The CNN model achieved an AUC of 0.92 and 0.95, with corresponding accuracy, precision, recall, and F1 scores of 0.8776, 0.8704, 0.9038, and 0.8868. SHAP and PDP analyses revealed that Elevation, Slope, Annual Rainfall, Drainage Density, LULC, NDVI, Distance from River, and TWI were the most influential factors for the trained FFSM. This prediction accuracy emphasises the potential of these models as reliable tools for the strategic planning of flood protection measures. This research thus demonstrates that the use of optimized DL models can significantly improve flash flood susceptibility mapping and provide a quantitative and methodologically sound approach to mitigating the negative impacts of flash floods. The results can help stakeholders to make informed decisions to reduce the risks of flash floods and ensure the safety of people and infrastructure in vulnerable areas.