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

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.