SCIENCE@home

A Deep Learning Approach for Wet Snow Monitoring in Mountainous Regions From SAR Image Time Series Based on Sentinel-1 and Sentinel-2 Snow Products

Thu Trang Lê,A. Atto,Emmanuel Trouvé,Fatima Karbou

Published 2025 in IEEE Transactions on Geoscience and Remote Sensing

ABSTRACT

Snow is a vital environmental parameter that holds significance across various disciplines, such as hydrology, meteorology, and natural disaster management. With the increasing accessibility of snow products derived from synthetic aperture radar (SAR) and optical data, like Sentinel-1 wet snow (WS) and Sentinel-2 total snow, users have benefited from improved snow mapping and monitoring. However, snow mapping in the mountainous areas remains challenging due to the difficulty of obtaining reliable ground-truth data on steep mountain terrain. In this study, we introduce a deep semantic segmentation framework, separable atrous convolution U-Net (SACUNet), specifically designed for WS detection from SAR image time series in mountainous environments. To address the lack of ground truth, we constructed a high-confidence training and validation database through a rigorous decision-fusion process combining multitemporal Sentinel-1 WS detections with Sentinel-2 total snow maps. We also propose two complementary metrics, the conditional agreement rate (CAR) and the WS intersection over union (WSIoU), to quantify the robustness and consistency of the fusion procedure, thus ensuring the reliability of training labels in the absence of in situ data. SACUNet integrates advanced techniques like: 1) depthwise separable convolution (DSC), which captures cross-channel dependencies and adapts feature representations and 2) atrous separable convolution (ASC), which further refines and consolidates the learned features, into the U-Net architecture. The proposed framework has been successfully employed to monitor WS in the Mont-Blanc massif, using a time series of 69 Sentinel-1 images acquired from July 5, 2020 to August 29, 2021. SACUNet demonstrates remarkable accuracy in WS detection, with an overall accuracy (OA) of 97%, precision of 94%, recall of 97%, IoU at 92%, and an $F1$ -score reaching 96%. Validation against meteorological records from four alpine stations confirmed that SACUNet effectively tracks seasonal WS dynamics, suppresses false detections during cold periods, and captures realistic high-altitude melt events. Moreover, the model trained in Mont-Blanc generalized successfully to the Vanoise massif, demonstrating its transferability to other alpine regions. Beyond quantitative accuracy, SACUNet enables the spatiotemporal analysis of WS evolution, offering insights into its extent, frequency, and seasonal progression across elevation bands. These findings highlight the framework’s potential as an operational tool for large-scale WS monitoring in mountainous environments.

PUBLICATION RECORD

  • Publication year

    2025

  • Venue

    IEEE Transactions on Geoscience and Remote Sensing

  • Publication date

    Unknown publication date

  • Fields of study

    Computer Science, Engineering, Environmental Science

  • Identifiers
    DOI 10.1109/TGRS.2025.3631318
  • External record

    Open on Semantic Scholar

  • Source metadata

    Semantic Scholar

CITATION MAP

EXTRACTION MAP

CLAIMS

  • No claims are published for this paper.

CONCEPTS

  • No concepts are published for this paper.

REFERENCES

Showing 1-69 of 69 references · Page 1 of 1

CITED BY

  • No citing papers are available for this paper.

Showing 0-0 of 0 citing papers · Page 1 of 1