A physics-informed deep learning urban canopy model for Mediterranean cities: Development, validation, and application to Thessaloniki, Greece

Urban canopy models (UCMs) are essential for representing urban surface–atmosphere exchanges, but their computational cost and detailed parameter requirements limit operational use. This study develops a deep neural network urban canopy model (DNN-UCM) trained on quality-controlled flux observations from 16 Urban-PLUMBER sites, with Heraklion, Greece withheld for independent Mediterranean validation. The model predicts upwelling shortwave and longwave radiation, sensible heat flux, and latent heat flux from meteorological forcing and six urban morphological parameters using a CNN-Attention-LSTM architecture. A physics-informed variant (DNN-UCM-PI) was further developed using curriculum training with energy-balance, albedo, surface-temperature, Bowen-ratio, and nocturnal-flux constraints. Cross-validation showed strong predictive skill, with DNN-UCM-PI improving R 2 for all four fluxes relative to the unconstrained model. At Heraklion, DNN-UCM-PI achieved R 2 = 0.997 for SWup, 0.991 for LWup, and 0.930 for Qh, while Qle remained less predictable (R 2 = 0.377), reflecting unresolved hydrological controls. Comparison with a literature-based, lightly adjusted offline TEB configuration showed better DNN-UCM-PI performance for radiation and Qh, although fully calibrated physics-based UCMs would likely narrow this gap. Applied to Thessaloniki, the model reproduced characteristic Mediterranean energy partitioning, with a Bowen ratio of 4.9, indicating dry urban surfaces and limited evaporative cooling. Local observations showed an urban heat island of +0.8 K relative to ERA5, causing 37% underestimation of Qh when reanalysis forcing was used. Physics-informed constraints improved extrapolation to dense urban canyons, upgrading Qle behavior at H/W = 2.42 from unreliable to plausible. The DNN-UCM-PI provides a computationally efficient framework for Mediterranean urban climate applications.