Training loss matters: exploring new fit-for-purpose loss combinations for statistical downscaling
Year: 2025
Type: preprint
Abstract: Deep learning-based statistical downscaling is increasingly adopted in weather and climate science. However, the choice of training loss function, which is central to guiding model optimisation, remains largely unexplored, with most studies relying on standard losses, such as the mean squared error and the negative log-likelihood Bernoulli-Gamma. In this work, we benchmark a wide range of loss functions for statistical downscaling, including standard losses and others novel to this field, such as wavelet and gradient-based formulations. We also investigate the benefits of combining complementary losses. The experiments cover two representative downscaling tasks with distinct statistical and physical behaviours: 10m zonal wind speed and precipitation. We found that the best levels of the considered performance aspects are reached combining different losses. Most of these combinations have not been used before. Besides, we observed that no loss combination dominates across all tasks or evaluation aspects; instead, the choice and combination of losses strongly influence accuracy, performance on extreme events, and visual rendering of the model. These findings highlight the importance of designing fit-for-purpose training losses, and also provide practical guidance for choosing effective ones. These results apply not only to deep learning-based statistical downscaling but can also be considered in related applications sharing similar challenges, such as weather forecasting.
Institutions: Centre National de la Recherche Scientifique, Centre Européen de Recherche et de Formation Avancée en Calcul Scientifique, Institut de Recherche pour le Développement, Climat, Environnement, Couplages et Incertitudes, Université de Toulouse, Centre National de Recherches Météorologiques, Météo-France
Language: English
Subfield: Atmospheric Science
Field: Earth and Planetary Sciences
Domain: Physical Sciences
SDG: Climate action
Open Access status: gold
Funder: Total
OpenAlex ID: W4414870727
Cites:
Cited by:
Related to: