Deep Learning-Based Snow Depth Retrieval Using Sentinel-1 Repeat-Pass InSAR

TL;DR

This paper introduces a deep learning model that accurately estimates snow depth from Sentinel-1 InSAR data, demonstrating improved transferability across regions and years compared to physics-based methods.

Contribution

A novel learning-based approach that directly maps InSAR observables to snow depth, outperforming traditional physics-based retrievals in accuracy and transferability.

Findings

Pearson correlation of 0.81 with lidar snow depth

Outperforms physics-based Sentinel-1 SWE retrievals

Shows strong temporal and spatial transferability

Abstract

Snow depth plays a central role in seasonal snowpack characterization and the terrestrial water cycle, yet remains challenging to estimate at high spatial resolution. Recent studies have shown that repeat-pass interferometric synthetic aperture radar (InSAR) measurements combined with physics-based models can enable effective snow water equivalent (SWE) retrieval. However, the performance of these methods depends strongly on measurement accuracy and modeling assumptions. Building on the success of InSAR-based approaches, we develop a robust learning-based model that directly learns the relationship between measured InSAR observables and snow depth. The model is trained on a single SnowEx Idaho site and evaluated across independent years and geographically distinct regions. Results demonstrate strong temporal and spatial transferability. In temporal transfer experiments, the proposed approach achieves a Pearson correlation of 0.81 with lidar snow depth, compared to a correlation of approximately 0.47 reported for physics-based Sentinel-1 SWE retrievals over the same site.