Long Time-Series Glacier Outlines in the Three-Rivers Headwater Region From 1986 to 2021 Based on Deep Learning

Abstract

The deep-learning-based approach has drawn significant attention in glacier extraction due to its advantages over traditional techniques. In this study, to verify the feasibility and effectiveness of LandsNet architecture for glacier extraction, we applied a modified LandsNet (M-LandsNet) to extract the glacier outlines in the Three-Rivers Headwater Region. The band ratio method, U-Net, U-Net++, GlacierNet, SaU-Net, U-Net+cSE, and LandsNet, and two scenes were used for comparison. Analysis of the two scenes indicated that the M-LandsNet had the best performance and generalization ability among the eight methods. Weather conditions had the greatest negative impact on the eight methods, followed by geographic environment and geographic location. We further extracted the glacier outlines in the Three-Rivers Headwater Region in 1986−2021 in a total of 12 periods using the M-LandsNet and through manual adjustments. The glacier area in the Three-Rivers Headwater Region has decreased by 416.40 ± 102.71 km2 (16.53 ± 4.08%) in 1986−2021. The reduction rate (16.13 ± 5.63 km2 a−1) in 2003−2021 was almost twice that (7.42 ± 5.97 km2 a−1) in 1986−2003. The reduction rate of the glacier area varied among different periods and areas. Comparison with previous results indicated that the obtained glacier outline dataset in this study is reliable, and can effectively reflect the glacier area and spatio-temporal glacier changes in the Three-Rivers Headwater Region. A long time-series dataset of glacier outlines in the Three-Rivers Headwater Region in 1986−2021 is available at https://doi.org/10.5281/zenodo.5512064. This study can provide data support for the estimation of regional water resources storage.

Keywords

• Deep learning
• long time-series glacier outlines
• M-LandsNet
• three-river headwater region
• U-Net