A systematic review of flood prediction (2018–2025): Flood categories, input features, and Machine Learning, Deep Learning and hybrid approaches

Abstract

Floods continue to pose a significant global threat, resulting in substantial human, infrastructural, and economic losses each year. Although floods cannot be entirely prevented, advancements in modern technologies, particularly data-driven approaches, offer significant potential to mitigate their impacts through accurate prediction. This paper conducts a Systematic Literature Review (SLR) of flood prediction research published between 2018 and 2025. The review focuses on four primary aspects: (1) the classification and geographical distribution of different flood types; (2) the key hydrological, meteorological, and geographical parameters utilized in flood prediction; (3) the range of innovative technologies applied, including traditional statistical approach, Machine Learning (ML), Deep Learning (DL), and hybrid models; and (4) the challenges associated with training ML models using historical and geographical datasets. The findings reveal a clear trend toward the adoption of advanced ML and ensemble-based methods due to their improved prediction accuracy and adaptability across diverse geographical contexts. The review also emphasizes the critical role of feature selection in enhancing model performance and highlights the growing importance of integrating real-time data streams for timely flood forecasting. Despite the progress, significant challenges persist, particularly the scarcity of comprehensive historical datasets in many regions, which affects model generalizability and robustness. This paper outlines potential future research directions, including leveraging transfer learning, data augmentation, and integrating heterogeneous data sources to develop more reliable and context-aware flood prediction systems.