Guo et al. (2026) Monitoring glacier-fed river width dynamics in High Mountain Asia from Sentinel-2 time series using a deformable UNet and skeleton evolution framework

Identification

• Journal: International Journal of Applied Earth Observation and Geoinformation
• Year: 2026
• Date: 2026-03-26
• Authors: Xiaoyu Guo, Kai Yang, W. Zhang, Xiaodong Yi, Yinghui Jiang
• DOI: 10.1016/j.jag.2026.105244

Research Groups

• School of Geography and Ocean Science, Nanjing University, China
• Jiangsu Provincial Key Laboratory for Advanced Remote Sensing and Geographic Information Technology, China

Short Summary

This study developed a novel framework integrating a deformable UNet (DUNet) deep learning model and a discrete, shape-preserving skeleton evolution algorithm to accurately monitor glacier-fed river width dynamics in High Mountain Asia using Sentinel-2 time series. The proposed method demonstrated superior performance over conventional deep learning models and existing global datasets, revealing significant seasonal variations in river width.

Objective

• To develop and validate a new framework, integrating a deformable UNet (DUNet) deep learning model and a discrete, shape-preserving skeleton evolution algorithm, for monitoring spatiotemporal dynamics of glacier-fed river widths in High Mountain Asia using Sentinel-2 time series.

Study Configuration

• Spatial Scale: High Mountain Asia (HMA), focusing on three representative regions within the Indus River basin (Indus-Shigar River, Shyok-Hushe River, Shyok-Nubra River) and eight independent test regions across the Tarim River, Amu Darya, Indus River, and Ganges–Brahmaputra basins. The study targets glacier-fed rivers of Strahler order 3 and above.
• Temporal Scale: Multi-temporal Sentinel-2 imagery from May to October 2024 for primary analysis of seasonal dynamics. Comparative analysis with GLOW-S (2017-2022) and SWORD (1984-2015) datasets for the mean annual river discharge period of 2022.

Methodology and Data

• Models used:
  • Proposed: Deformable UNet (DUNet) for semantic segmentation of river masks, and a discrete, shape-preserving skeleton evolution algorithm for generating continuous river centerlines and estimating widths.
  • Comparative (for river mask extraction): UNet, DeepLabV3+, Dynamic World.
• Data sources:
  • Satellite Imagery: Multi-temporal Sentinel-2A/B Level-1C optical images (visible, near-infrared, and shortwave-infrared bands at 10 m and 20 m resolution).
  • Digital Elevation Model (DEM): Copernicus GLO-30 Digital Elevation Model (30 m resolution).
  • Ancillary Datasets: Randolph Glacier Inventory version 7.0 (RGI), HydroBASINS Level 4 to 6, HydroRivers.
  • Comparative River Width Datasets: Global Long-term River Width from Sentinel-2 (GLOW-S), Surface Water and Ocean Topography Mission River Database (SWORD).
  • Derived Data: Modified Normalized Difference Water Index (MNDWI), Height Above the Nearest Drainage (HAND) index, manually delineated ground truth river masks.

Main Results

• The proposed DUNet method accurately mapped multi-temporal glacier-fed rivers in HMA, achieving a Kappa coefficient of 0.896 ± 0.038, F1-score of 0.900 ± 0.038, Precision of 0.925 ± 0.031, and Recall of 0.876 ± 0.073, outperforming UNet, DeepLabV3+, and Dynamic World.
• DUNet demonstrated strong spatial generalization ability, successfully extracting rivers across diverse HMA regions, including challenging areas with turbid water, valley shadows, or high topographic relief.
• The method yielded reliable estimates of river width, with coefficient of determination (R2) values ranging from 0.953 to 0.988 and mean absolute errors (MAE) ranging from 19.87 m to 45.14 m across the study regions.
• Glacier-fed rivers exhibited significant seasonal dynamics, with river width varying by approximately a factor of 2 between its relatively wide and narrow states (Interquartile Ranges of 162.22 m and 133.58 m, Coefficients of Quartile Variation of 0.28 and 0.29 in SHR and SNR, respectively). Rivers typically expanded from May to August and contracted from August to October.
• Comparative global river width products (SWORD and GLOW-S) systematically overestimated glacier-fed river widths in the study regions, with MAE values for SWORD up to 508.87 m and for GLOW-S up to 219.78 m, significantly higher than DUNet's MAE of 12.91 m to 30.37 m.

Contributions

• Introduction of a novel framework combining a deformable UNet (DUNet) and a discrete, shape-preserving skeleton evolution algorithm, specifically optimized for monitoring complex glacier-fed river width dynamics in High Mountain Asia.
• Demonstration of superior accuracy and spatiotemporal stability in river mask extraction and width estimation compared to conventional deep learning models and existing global river width products (SWORD, GLOW-S) in challenging, morphologically complex environments.
• Highlighting the importance of adaptive deep learning models with flexible receptive fields for capturing multi-scale, highly sinuous, and braided river features, addressing limitations of fixed-receptive-field models.
• Provision of a more reliable method for tracking intra-annual and inter-annual glacier-fed river width dynamics, which is crucial for regional water resource management, flood mitigation, and fully utilizing the capabilities of satellite missions like SWOT.

Funding

• National Key Research and Development Program of China (2022YFB3903601)
• National Natural Science Foundation of China (42271320)
• Fundamental Research Funds for the Central Universities (2025300383)

Citation

@article{Guo2026Monitoring,
  author = {Guo, Xiaoyu and Yang, Kai and Zhang, W. and Yi, Xiaodong and Jiang, Yinghui},
  title = {Monitoring glacier-fed river width dynamics in High Mountain Asia from Sentinel-2 time series using a deformable UNet and skeleton evolution framework},
  journal = {International Journal of Applied Earth Observation and Geoinformation},
  year = {2026},
  doi = {10.1016/j.jag.2026.105244},
  url = {https://doi.org/10.1016/j.jag.2026.105244}
}