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

The Science of The Total Environment, Journal Year: 2025, Volume and Issue: 964, P. 178606 - 178606

Published: Jan. 24, 2025

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

Language: Английский

A Snow Depth Downscaling Algorithm Based on Deep Learning Fusion of Enhanced Passive Microwave and Cloud-Free Optical Remote Sensing Data in China

Remote Sensing, Journal Year: 2024, Volume and Issue: 16(24), P. 4756 - 4756

Published: Dec. 20, 2024

High spatial resolution snow depth (SD) is crucial for hydrological, ecological, and disaster research. However, passive microwave SD product (10/25 km) increasingly insufficient to meet contemporary requirements due its coarse resolution, particularly in heterogeneous alpine areas. In this study, we develop a superior downscaling algorithm based on the FT-Transformer (Feature Tokenizer + Transformer) model, termed FTSD. This fuses latest calibrated enhanced brightness temperature (CETB) (3.125/6.25 with daily cloud-free optical data (500 m), including cover fraction (SCF) days (SCD). Developed evaluated using 42,692 ground measurements across China from 2000 2020, FTSD demonstrated notable improvements accuracy of retrieval. Specifically, RMSE temporal spatiotemporal independent validation 7.64 cm 9.74 cm, respectively, indicating reliable generalizability stability. Compared long-term series (25 km, = 10.77 cm), m, 7.67 cm) provides accuracy, especially improved by 48% deep (> 40 cm). Moreover, higher effectively captures SD’s heterogeneity mountainous regions China. When compared algorithms utilizing raw TB traditional random forest CETB model optimize 10.08% 4.84%, which demonstrates superiority regarding sources regression methods. Collectively, these results demonstrate that innovative exhibits performance has potential provide robust foundation meteorological environmental

Language: Английский