Monitoring glacier-fed river width dynamics in High Mountain Asia from Sentinel-2 time series using a deformable UNet and skeleton evolution framework

• DUNet with deformable convolutions captures complex glacier-fed river dynamics. • A shape-preserving skeleton evolution algorithm conducts reliable width estimation. • Glacier-fed rivers exhibit approximately two-fold seasonal width variations. Glacier-fed rivers are widely distributed in High Mountain Asia (HMA) areas and play a vital role in regional water resources and flood hazards. However, their spatiotemporal dynamics are very challenging to monitor due to their multi-scale, highly sinuous, and highly variable braided morphologies. Here we propose a new framework to monitor glacier-fed river width dynamics in HMA from Sentinel-2 time series by integrating a deformable UNet (DUNet) deep learning model and a discrete, shape-preserving skeleton evolution algorithm. The DUNet semantic segmentation model adaptively achieves learnable offsets and adjusts the scale and morphology of receptive fields; thus, it is suitable for capturing complex river variations based on river masks extracted independently from each image. The skeleton evolution algorithm is capable of preserving continuous and accurate river centerlines; thus, it can reliably estimate river width, particularly for highly braided channels. We map multi-temporal glacier-fed rivers in representative HMA regions and compare our mapping results with those obtained from three conventional deep learning models (UNet, DeepLabV3+, and Dynamic World). The results show that our proposed method can accurately monitor multi-temporal glacier-fed rivers in HMA, achieving a Kappa coefficient of 0.896 ± 0.038, which is higher and more stable compared with that of the other three models. Independent test regions across HMA further validate the spatial generalization ability of our method. In addition, our method yields more reliable estimates of river width, with coefficient of determination (R 2 ) values ranging from 0.953 to 0.988 and mean absolute errors ranging from 19.87 to 45.14 m across the study regions. Moreover, we find significant seasonal dynamics in river width, where the river width varies by approximately a factor of 2 between its relatively wide and narrow states. In summary, this study demonstrates that the specifically optimized deep learning models hold strong promise for capturing spatiotemporal variations of complex landscape features such as terrestrial rivers.