Unveiling urban process-dependent hydrological responses via rainfall “fingerprints”

Characterizing rainfall variability is essential for hydrometeorological analysis and building flood-resilient cities. However, progress is hindered by an incomplete grasp of rainfall heterogeneity, ambiguous links to hydrological impacts, and a lack of tools for systematically and operationally identifying event-scale rainfall processes. Here, we analyze 23 years of hourly rainfall observations from 61 gauges in Shanghai and introduce an integrated machine learning methodology combining unsupervised and supervised learning to: (1) extract reproducible spatiotemporal rainfall “fingerprints,” hydrologically meaningful event-scale rainfall process archetypes, (2) validate their hydrological responses, and (3) build an automated fingerprint classifier. A pronounced west–east contrast in fingerprint diversity emerges across the city, with more homogenized fingerprints in highly urbanized sectors. Fingerprints also display strong spatiotemporal heterogeneity in rainfall process evolution, forming station-specific archetypes that are internally coherent yet indistinguishable by rainfall intensity or accumulation alone. Hydrological analysis reveals an asymmetric relationship between rainfall fingerprints and hydrological response: peak water level is primarily associated with cumulative rainfall and duration, whereas the rising rate is more sensitive to peak intensity. Identical fingerprints can produce different responses, while distinct fingerprints may yield similar responses under varying boundary conditions. The proposed classifier achieves >[jls-end-space/]80% accuracy in identifying the rainfall fingerprints of unseen events at over 50 stations, supporting the potential for automated rainfall process type recognition. The methodology established here not only deepens understanding of urban rainfall but also provides a scientific and technical foundation for site-adaptive, process-informed urban flood governance.