The feasibility gap: a systematic review of AI flood forecasting strategies in data-scarce regions

Floods claim thousands of lives annually across Africa, yet the region remains underserved by operational flood forecasting systems. Artificial intelligence has shown promise over physics-based models for flood forecasting. However, developing these models requires climate observations. This systematic review synthesizes strategies proposed for developing AI models in data-scarce settings. We classify them into data-centric and model-centric approaches. Beyond this taxonomy, we note that the reviewed studies do not explicitly distinguish between flood driver scarcity (i.e., unavailable hydrometeorological inputs) and flood event scarcity (i.e., missing records of past flood events). While flood driver scarcity can be mitigated through various data-substitution strategies, flood event scarcity cannot, as flood occurrence can only be confirmed through direct, local observations. Furthermore, we show that many of the proposed strategies are not scalable in Africa because they rely on local in-situ data, which are largely unavailable across much of the continent. Specifically, 97% of the reviewed studies depended on in-situ observations for training, fine-tuning, or validation. Consequently, many approaches intended for data-scarce settings remain operationally unusable in regions where climate observations are actually scarce. We conclude that future research must pivot from low-in-situ-data adaptation towards fully ungauged strategies.