Physics-guided temporal fusion for long-term daily urban land surface temperature: Accuracy, uncertainty, and interpretability in Geneva, Switzerland

Urban climate analyses often rely on satellite-derived land surface temperature (LST), which provides spatially extensive observations of surface thermal conditions but remains fragmented by cloud contamination and complex urban heterogeneity. We introduce SEB-TwinTFT, a physics-guided hybrid modelling framework that reconstructs daily gap-free urban LST with quantified uncertainty while simultaneously estimating effective surface energy balance (SEB) parameters. Unlike conventional deep-learning approaches that emulate LST directly, SEB-TwinTFT produces spatially and temporally varying fields of effective albedo ( α ), aerodynamic resistance ( r a ), evaporative efficiency ( β ), and ground heat-flux fraction ( γ ), which describe how the model partitions the surface energy balance. Embedding a differentiable SEB formulation within the network reduces purely black-box behavior and provides interpretable diagnostics of the processes influencing modeled urban surface temperature variability. Architecturally, SEB-TwinTFT couples a temporal-fusion transformer head with a differentiable SEB module through a learnable gate that adaptively blends data-driven and physics-guided predictions. Applied to summertime conditions in Geneva for 2000–2024, the framework reproduces observed spatial and temporal LST variability, yields SEB parameter patterns broadly consistent with established surface–atmosphere exchange processes, and captures coherent spatial structures such as a systematic lakeward displacement of nocturnal temperature maxima that is consistent with lake–land breeze influences. Cross-decadal validation with a strict 2011–2020 hold-out demonstrates stable performance across typical summers and upper-tail LST extremes. Clear-sky sampling bias in thermal infrared observations is addressed through inverse-probability weighting of Missing-Not-at-Random observations, and heteroscedastic training provides calibrated prediction intervals. SEB-TwinTFT therefore offers an uncertainty-aware and physically interpretable framework for reconstructing urban LST and analyzing surface energy-balance variability in complex lake–urban environments.