Generalist-specialist transfer learning for cross-city spatial generalization in pluvial flood prediction

Pluvial flood prediction across cities remains challenging, as data-driven models often generalize limited beyond their training cities. Here we develop a generalist-specialist transfer learning approach for cross-city pluvial flood prediction using datasets from cities with diverse characteristics. A generalist model is trained on multi-city simulations, and specialist models are then adapted to target cities using a single local design rainfall event, either by full fine tuning or with the encoder frozen. Across 10 cities in Southwest England, the specialist models substantially outperform the generalist, with the overall Recall and Precision increased from 0.650 to 0.807 and 0.681 to 0.833, and Root Mean Square Error (RMSE) reduced from 0.094 m to 0.047 m. Feature analysis using centered kernel alignment (CKA) and representational similarity analysis (RSA) shows that cross-city transfer largely reuses shared representations across cities, with only minor local adaptation. These results demonstrate that strong spatial generalization can be achieved by leveraging transferable representations, revealing a data-efficient and scalable pathway for adapting flood prediction models to new cities using only one local design flood event.