Unlocking full-domain flood inundation prediction from sparse point supervision in deep-learning models

Deep learning (DL) surrogates enable rapid flood inundation prediction, but they are typically trained on hydrodynamic simulations and therefore inherit the structural errors of those models. Training DL surrogates directly on sparse point observations of water depth could in principle avoid this inheritance of error, but the mismatch between sparse point observations and the spatially continuous supervision required by full-domain models remains unresolved. We propose an observation-driven framework that couples a spatiotemporal DL surrogate with a physics-informed spatial expansion method reconstructing full-domain water depth fields from sparse observations. The two components are combined through two candidate modelling pathways - pre-training expansion (Pre-Exp) and post-prediction expansion (Post-Exp) - that differ in when the expansion is applied: before training or after prediction. Evaluating the framework across two cases in urbanized floodplains, we compare the two pathways, test the framework under imperfect observations, and assess how observation density and placement affect predictive accuracy. The results show that with 50 observation points covering <0.1 % of domain cells, the framework predicts full-domain water depth at a cross-validated Critical Success Index (CSI) of 0.863 and 0.769 in the two study areas. The Pre-Exp pathway outperforms Post-Exp by avoiding the amplification of point errors during expansion, and proves robust to noisy, incomplete and asynchronous observations. Under sparse observation conditions, a terrain-based placement strategy improves accuracy over uninformed placement at matched density. These findings establish a methodological foundation for observation-driven flood inundation prediction.