LASDM: A Linear Attention based Hybrid CNN-Transformer for Snow Downscaling

Read full text Summarise with VIRAT
Authors: Mubashshir Ali, Farid Ait-Chaalal, Siddharth Gururani, Alison Dobbin
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

High-resolution Snow Water Equivalent (SWE) data are essential for hydrological modeling, water resource management, and climate impact assessments in snow-dominated regions. Regional models offer improved resolution but are computationally expensive, motivating the use of deep learning-based emulators. In this work, we introduce the Linear Attention Snow Downscaling Model (LASDM), a hybrid architecture that combines convolutional neural networks (CNNs) for local feature extraction with linear attention blocks for modeling long-range dependencies. LASDM is designed to be highly parameter-efficient using fewer than one million trainable parameters. We evaluate LASDM for downscaling and bias correction against convolutional U‑Net, a Swin Transformer–based model, and a statistical baseline based on empirical quantile mapping over the Great Lakes region of North America. Results show that LASDM consistently outperforms these benchmarks across both typical and extreme SWE regimes, while maintaining substantially lower computational complexity. The case studies of two winter storms further demonstrates model performance. Remaining challenges include correcting topographic and snow melt-related biases. These findings highlight the potential of lightweight hybrid CNN-transformer architectures for efficient downscaling and bias correction tasks.

Fields of Study & Subjects

Cryospheric studies and observations Hydrological Forecasting Using AI Meteorological Phenomena and Simulations

Publication details

Publication date
25 March 2026
Language
EN
Document type
Preprint