Flooding remains one of the most damaging natural hazards worldwide, and the growing availability of hydrological sensor networks and satellite data has pushed researchers toward data-driven alternatives to conventional physics-based hydrological models. This paper synthesizes findings from six recent bodies of work — three literature reviews and three influential primary studies — covering the evolution of machine learning (ML) applications in flood and flash-flood forecasting, from early architecture benchmarking through real-world operational deployment to static susceptibility mapping. The synthesis traces the field's progression from early artificial neural network (ANN) and support vector machine (SVM) implementations toward the current dominance of Long Short-Term Memory (LSTM) networks and hybrid architectures that combine convolutional, recurrent, and physics-based components. Across the reviewed literature, meteorological and hydrological variables — principally rainfall and water level — account for the substantial majority of model inputs, while error-based metrics such as RMSE and Nash-Sutcliffe Efficiency (NSE) remain the standard for performance evaluation. The synthesis also identifies recurring structural weaknesses in the field: a heavy geographic concentration of research in East and Southeast Asia, limited dataset sharing that constrains reproducibility, underused approaches such as graph neural networks and transfer learning, and insufficient attention to uncertainty quantification. The findings suggest that while data-driven flood forecasting has matured considerably in predictive capability, its transition into dependable operational early-warning systems still depends on resolving these gaps in transparency, generalizability, and interpretability.

Keywords— Flood forecasting; machine learning; deep learning; LSTM; hydrological modeling; hybrid models; flash floods.