Observational Uncertainty in Deep Learning Climate Downscaling and Projections

Deep learning (DL) models are increasingly used for climate downscaling, yet the impact of observational uncertainty on their performance remains poorly understood. Here we assess this effect by training ensembles of DL downscaling models using four independent gridded observational data sets over continental Spain that differ in source, methodology, and spatial coverage. Each ensemble includes multiple training realizations of the same model to capture training variability. The models are then applied to downscale global climate model temperature projections. Results show that both intra‐model (across realizations) and inter‐dataset (across observational products) variability substantially influence the downscaled climate change signal. Differences among data sets dominate this uncertainty, especially for extremes (annual 99th percentile) at regional scales, leading to warming discrepancies of up to 2.5°C. However, these discrepancies are smaller for high‐quality regional data sets built from dense observation networks and rigorous quality control, whereas global observational data sets, known to misrepresent local climate features (Bedia et al., 2013, https://doi.org/10.1016/j.gloplacha.2013.04.005 ), exert a larger impact on downscaled projections. These findings highlight that the use of well‐constructed regional data sets can substantially reduce uncertainty and improve the reliability of deep learning–based downscaling. Explicitly incorporating observational uncertainty into training and evaluation frameworks remains essential to ensure robust and credible regional climate projections.