Advanced deep learning techniques for automated extraction of non-debris-covered areas of glaciers in High-Mountain Asia using time-series remote sensing data

Deep learning approaches have gained prominence for automatic glacier boundary extraction due to their localized nature of convolutional operations, potentially leading to incomplete or fragmented glacier pixel representations. Moreover, the accuracy of extracting glacier boundaries from a single remote sensing image (RSIs) is often influenced by seasonal snow, clouds, shadows, and frozen lakes. To overcome these challenges, we introduce a novel model for extracting the non-debris-covered areas of glaciers (NDCAG) from RSIs, termed GlacierSTR-UNet. This model enhances information flow and overall performance by embedding the Swin Transformer (ST) as an encoder into a U-shaped architecture and reduces training time and improves gradient handling by incorporating the ResNet block in the decoder. We deploy the GlacierSTR-UNet model on the Google Earth Engine (GEE) platform to efficiently generate multiple NDCAG results from RSIs taken at different periods. A pixel-by-pixel synthesis algorithm is then applied to aggregate the multiple NDCAG extraction results, producing the final NDCAG. Accuracy assessments indicate that GlacierSTR-UNet achieves an overall accuracy of 0.8817, and the relative deviation between automatically extracted and manually interpreted NDCAG remains within 2 %. Finally, we obtain the NDCAG datasets for the periods of 2015/2016 and 2022/2023 in High-Mountain Asia, revealing a reduction of 4,185.12 ± 7,870.96 km² in NDCAG from 2015/2016 to 2022/2023. These findings demonstrate the effectiveness of our approach in efficiently and accurately extracting NDCAG, highlighting its potential for monitoring glacier changes and supporting glacier inventory efforts.