A historical time series of glacial debris cover change in the Himalayas using multi-source satellite data and deep learning

Accurate estimation of current and historical glacier area provides crucial information for glacier monitoring and for projecting future glacier change. Despite significant advances in remote sensing, automated glacier delineation techniques still exhibit limited accuracy in debris-covered areas, often requiring time-consuming manual corrections, which are impractical for widespread application. While some regions have rich historical records of glacial development due to ease of access and scientific prioritization, many regions are represented by only a few outlines.

The Himalayas contain many large and heavily debris-covered glaciers, whose retreat and increase in debris coverage could drastically affect the primary source of freshwater for many communities. This vast and important region, however, has very few harmonized and temporally consistent datasets available for scientific use. Existing inventories are generally derived from imagery acquired in different years and exhibit substantial differences in their debris-covered area outlines.

Thus, to improve on these inconsistencies and inaccuracies, we are creating a historical time series spanning the entirety of the Himalayas between 1984 and 2025 with two-year intervals. To overcome the difficulties inherent to debris-covered glacier inclusion, several data sources are used, including optical, infrared, thermal, spectral indices, elevation change rate, surface velocity, and topographic parameters. These data sources are trained on a deep learning network comprised of a pre-trained U-Net with a long short-term memory (LSTM) module, which enhances debris segmentation by introducing historical data to the training process. In addition, the potential benefit of including radar coherence data is evaluated to assess whether glacier outlines produced in recent decades can be further refined.