A machine learning-based approach for flash flood susceptibility mapping considering rainfall extremes in the northeast region of Bangladesh

• Flash flood susceptibility maps for 8 northeastern districts of Bangladesh. • 200 flood and 200 non-flood points selected on field, historical, and satellite data. • Susceptibility maps considering rainfall extremes for various return periods. • Four machine learning models applied for the maps: ANN , RNN, RFGB, and CatBoost. • ANN model performs the best in terms of both accuracy and the worst-case scenario. Flash floods are catastrophic global events, especially in northeast Bangladesh, and assessing flash flood susceptibility is crucial for preparedness and mitigation. Traditional geographic system-based flash flood susceptibility mapping struggles to capture flash floods’ non-linear and complex nature. However, machine learning models have recently emerged as an efficient alternative to address these limitations. This study evaluates four machine learning models—Artificial Neural Network (ANN), Recurrent Neural Network (RNN), Random Forest – Gradient Boosting (RFGB) Hybrid, and Categorical Boosting (CatBoost)—for flash flood susceptibility. It categorizes areas into five susceptibility levels: very low, low, moderate, high, and very high. Covering 24,424.25 km 2 across eight districts, the study uses 400 points (200 flood and 200 non-flood) for training and validation, based on field investigation, historical flood information, Sentinel-1 SAR GRD: C-band Synthetic Aperture Radar data using Google Earth Engine, and insights from the local people. The models’ predictive performances are evaluated by incorporating topographical attributes and rainfall indices and using accuracy, precision, recall, F1 score, and Receiver Operating Characteristic (ROC) Area Under the Curve (AUC) score, with 70 % of the data for training and 30 % for testing. The ANN model performs best with rainfall indices, achieving high accuracy (maximum AUC = 0.802). The RFGB hybrid model shows excellent training accuracy (AUC ≥ 0.971) but suffers from overfitting during validation (AUC ≤ 0.674), requiring careful hyperparameter adjustment. The CatBoost model effectively uses both rainfall indices and terrain features, achieving AUC = 0.701 in training and AUC = 0.667 in validation. The ANN model conservatively includes the largest area (2198.3 km 2 ) under ’very high’ susceptibility. This study’s flash flood susceptibility maps, which include rainfall extremes, are more robust than those without, helping local administrative authorities and national flood practitioners prepare for flash floods.