Evidence from XAI for how extreme precipitation relates to urbanisation

Rapid urbanisation is coinciding with a rising trend in the intensity of extreme precipitation. Beyond this large-scale trend, urbanisation further alters local climate by modifying land cover, energy fluxes and airflow. However, it remains unclear how much urbanisation alters heavy precipitation, and how robust different machine learning (ML) models are in uncovering its influence. This uncertainty limits our ability to design targeted adaptation measures (e.g. managing impervious surfaces, cooling hotspots). To address these gaps, we analyse extreme daily precipitation across more than 5,000 stations in Europe using gauge observations, high-resolution meteorological reanalyses and 1km land-use data. Stations are classified into four urbanisation levels (rural, suburban, urban, highly urban) based on impervious surface fraction of surrounding area, and predictors are grouped into geographic, surface, thermal and dynamic categories. We train multiple ML models (ElasticNet, RF, LightGBM, and ANN) under a unified framework and applied explainable AI techniques (SHAP and ALE) to diagnose how these models use physical information across urbanisation levels. Tree-based ensembles achieve the highest skill (R² = 0.45, RMSE=9.28 mm), while all models systematically underestimate the most intense events (>100 mm/d). Our analysis of the ML models finds that thermodynamic variables (dewpoint temperature and heat flux) are the primary controls on extreme precipitation across all urbanisation levels, as evidenced by their larger SHAP ranges (1.36–1.66) compared with the other categories. In contrast, dynamic predictors (U/V component of wind, pressure, and vertical velocity) exert a weaker but relatively consistent influence (SHAP range: 0.74–0.88). In non-urban models, surface processes play a limited role in explaining extreme precipitation. However, in the highly urban model, increasing impervious surface fraction contributes positively to predicted rainfall intensity (net ALE change of about 8 percentage points across the data range). Hence, as urbanisation intensifies, we find impervious surfaces are becoming an increasingly significant explanatory factor in ML models of heavy rainfall.