The application of geospatially-integrated machine learning models for flood prediction: A review

With the increase in flooding events globally, Artificial Intelligence (AI) approaches, particularly geospatially integrated machine learning (ML), have emerged as a viable tool for flood modeling. This review systematically analyzed 100 peer-reviewed articles published between 1990 and 2024 to examine trends in ML applications across different flood types, their performance, opportunities, and limitations, and to identify gaps between research activity and flood occurrence, highlighting the need for capacity building in some regions. Results show that ML models are most frequently applied to pluvial floods (44%), followed by fluvial (40%), compound (9%), and coastal (7%) floods, while applications for pluvial, coastal, and compound flooding continue to increase. ANN-based models were the most commonly applied approaches, particularly in fluvial and compound flood studies, followed by SVM and RF models. Reported model performance varied substantially across studies depending on flood type, data availability, and validation strategy, with several localized studies reporting very high predictive accuracy under controlled validation conditions. The geographical distribution of 2,395 flood occurrences from the EM-DAT database revealed disparities between research output and flood occurrence. Publications align relatively well with Asia (50% of publications; 1,450 flood events), Europe (22%; 336 events), and North America (17%; 207 events), whereas Africa (5%; 97 events) and South America (5%; 244 events) remain underrepresented despite significant flood exposure. These findings highlight the need for improved data access, capacity building, and transferable ML approaches to support flood prediction research in underrepresented but flood-prone regions.