Review of Machine Learning and Deep Learning Techniques for Flood Forecasting

Natural flood incidents are increasingly threatening urban and peri-urban populations; however, the effectiveness of operational early warning systems is jeopardized by the variety of disparate data, insufficient applicability across regions, and inconsistent evaluation criteria. This paper, through a systematic review of published studies related to AI-based flood prediction in urban areas, specifically studies based on traditional machine learning (ML) methods, deep learning, hybrid physics-ML modelling, and probabilistic models, summarizes the existing body of work. The overall findings demonstrate that satisfactory prediction systems require a multiple-modal pipeline that encompasses different hydrology datasets, historical incident data and numerical weather forecasts. Successful implementation requires extensive feature engineering, with an emphasis on time-series analyses (i.e., lagged fills and rolling statistics) and composite risk indicators. Predictive modelling incorporates robust ensemble methods like XGBoost, with the application of stratified cross-validation and the optimization of thresholds specifically for prioritizing high recall in alerts that pose a safety risk. While reported results in studies indicate good discriminative measures across multiple test datasets, they also allude to continued weaknesses in the space, including ongoing issues related to data heterogeneity, transferability across different basins, and uncertainty calibration, along with a fundamental challenge in establishing robustness.