Abstract:
Susceptibility modeling of flash floods plays a pivotal role in identifying distinct areas prone to flash flood inundation, which is very important for disaster preparedness and mitigation strategies. The traditional approach of geographic system-based mapping has limited capability to capture the non-linear pattern of flash flood characteristics, where, in recent times, machine learning models have proved to be an efficient alternative. This thesis presents a comprehensive assessment of four specific machine learning models, namely the Artificial Neural Network (ANN), Recurrent Neural Network (RNN), Random Forest – Gradient Boosting (RFGB) Hybrid Algorithm, and Categorical Boosting (CatBoost), for modeling the flash floods’ susceptibility. Moreover, it identifies and divides susceptible areas under flash floods into five categories: very low, low, moderate, high, and very high. The study area includes an area of 24,424.25 km2, including eight districts: Brahmanbaria, Habiganj, Kishoreganj, Maulvibazar, Mymensingh, Netrakona, Sunamganj, and Sylhet. Four hundred points (200 flood and 200 non-flood points) are considered here for training and validation purposes 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. Among these points, 200 points are flood points, and 200 are non-flood points, where each district gets 25 flood points and 25 non-flood points. By harnessing diverse datasets incorporating topographical attributes (elevation, slope, aspect, curvature, topographic roughness index, sediment transport index, stream power index, land use/land cover, topographic wetness index) and rainfall indices, the study rigorously evaluates and contrasts the models' predictive performances. The complex picture of their skills is shown by a range of commonly used statistical measures, including accuracy, precision, recall, F1 score, and Receiver Operating Characteristic(ROC) Area Under the Curve(AUC) score. Among the 400 flood and non-flood points, 70% of the data is used as a training dataset, and 30% of the data is used as a testing dataset. The ANN model performs the best characterized by rainfall indices, exhibiting recommendable accuracy values (maximum AUC = 0.802 for maximum consecutive 5-day rainfall, 100-year return period). The RFGB hybrid model, on the other hand, exhibits a complicated interaction of advantages and disadvantages. It exhibits outstanding accuracy during training (AUC ≥ 0.971 for all cases), but during validation, it suffers from overfitting issues (AUC ≤ 0.674 for all cases) that need careful hyperparameter tweaking and regularization. The CatBoost model shows a clear sensitivity to both rainfall indices and terrain features, expertly using both characteristics to improve classification accuracy (highest AUC = 0.701 in the training dataset and AUC = 0.667 in the validation dataset). This model further excels by demonstrating outstanding generalization capabilities with little overfitting, giving it a distinct advantage. Moreover, the ANN model presents the most conservative scenario, including the most area (2198.3 km2, for the maximum consecutive 5-day precipitation, 100-year return period case) under the ‘very high’ susceptibility category. In the light of a comparative analysis contextualizing the relationship between a rainfall event and a flash flood event, the example of the case of 2017 shows that the rainfall event causing the 2017 flash flood event resembles the rainfall of 50 years to 100 years return period considering monthly maximum 1-day rainfall. This comparativestudy has practical implications in many cases. It provides vital information for building reliable and trustworthy flood susceptibility models, considering the various and context-specific aspects of the models, making a substantial contribution to successful flash flood management techniques.