Projecting future snow changes at kilometer scale for adaptation using machine learning and a CMIP6 multi-model ensemble

Abstract

Assessing future snow cover changes is challenging because the high spatial resolution required is typically unavailable from climate models. This study, therefore, proposes an alternative approach to estimating snow changes by developing a super-spatial-resolution downscaling model of snow depth (SD) for Japan using a convolutional neural network (CNN)-based method, and by downscaling an ensemble of models from the Coupled Model Intercomparison Project Phase 6 (CMIP6) dataset. After assessing the coherence of the observed reference SD dataset with independent observations, we leveraged it to train the CNN downscaling model; following its evaluation, we applied the trained model to CMIP6 climate simulations. The downscaled mean ensemble reproduced the spatial distribution and seasonality of the reference observations. We found an average decrease in the snow-covered area by about 20 % in winter and 25 % in early spring, an altitude-dependent of the SD changes, and a delayed snow cover appearance by the middle of the 21st Century under a high emission scenario. Overall, the downscaling model captures physically plausible relationships, enables high-resolution assessments of future SD based on a multi-model ensemble, produces results consistent with regional climate models, and provides valuable insights into how future snow changes will affect winter tourism and water resources, highlighting its potential benefits for a wide range of adaptation studies.