Advanced machine learning for urban flood hazard assessment

Rapid urbanization in Bangkok leads to severe urban flooding that damaging infrastructure. This impact disrupting economic stability and hindering sustainable growth. Existing flood hazard assessment approaches fail to accurately depict the dynamic relationships between climate variations, infrastructure elements, and human activities. This study introduces an integrated machine learning framework for flood prediction and hazard classification which combines meteorological, hydrological, and socioeconomic data from district areas over 22 years from 2002 to 2023. Three primary prediction tasks were evaluated which included classification of flood occurrences as well as regression analysis of flood levels and classification of flood hazard levels. The research evaluated numerous machine learning algorithms such as Random Forest, LightGBM, XGBoost, Support Vector Machine, Multi-Layer Perceptron, Long Short-Term Memory along with Temporal Fusion Transformer. The novel Graph Convolutional Network–Temporal Fusion Transformer (GCN-TFT) model is proposed to capture spatial dependencies between districts alongside long-range temporal patterns in flood behavior. The GCN-TFT model demonstrated superior performance compared to all baseline models across multiple evaluation metrics and achieved a 0.995 F1 score for hazard level classification. The model produced high-resolution hazard maps that exposed consistent spatial distribution patterns in flood-prone areas along with significant temporal regime changes. We revealed the most influential predictors through feature importance analysis, enhancing model interpretability by highlighting which variables most significantly contributed to flood hazard prediction. The developed data-driven framework achieves high accuracy in urban flood hazard mapping and provides transferable and interpretable solutions that aid disaster planning and infrastructure development in cities vulnerable to climate change.