Computer vision in flash flood forecasting: A narrative review of applications, integration pathways, and future directions

Flash floods are among the most destructive hydrometeorological hazards, requiring forecasting approaches that balance accuracy, timeliness and scalability. Computer vision (CV) has become a key enabler, providing near-real-time information from unmanned aerial vehicles (UAVs), satellites and ground-based imagery. This structured narrative review synthesises advances across six domains relevant to forecasting workflows: flood extent mapping, debris and water-level detection, land use and land cover (LULC) classification, change detection, impact assessment, and image compression. Integration pathways between CV outputs and hydrological or hydraulic models are examined, revealing that such coupling remains limited in current literature. It also highlights the future areas of research in this domain. The systematic assessment shows that convolutional neural network (CNN)-based segmentation remains the most practical approach for extracting real-time information from image data, while transformers and lightweight models show promise for real-time use. Persistent challenges include the scarcity of UAV benchmarks, reproducibility gaps and weak operational integration.