Multi-Temporal Glacier Mapping and Change Detection Using Sentinel-2 Imagery and a DEM-Aware Hierarchical Random Forest Framework in Spiti Valley, Himachal Pradesh

Abstract

Himalayan glaciers are undergoing substantial environmental change and their meltwater supports downstream water systems. Monitoring their spatial extent over time is therefore important. Glacier delineation becomes particularly difficult when ice is partly covered by debris or when terrain and atmospheric conditions create spectral confusion. This study presents a DEM-aware hierarchical Random Forest approach that uses Sentinel-2 Level-2A imagery together with Digital Elevation Model (DEM) information to map glacier surfaces and quantify their temporal change in Spiti Valley, Western Himalaya. The proposed framework brings together six Sentinel-2 bands, three spectral indices (NDSI, NDVI and NDBI), and terrain descriptors derived from the DEM, namely elevation, slope, aspect and curvature. Cloud- and shadow-affected observations are screened with the Sentinel-2 Scene Classification Layer. Training observations are additionally weighted using elevation and slope so that samples representing more reliable glacier conditions have greater influence during model fitting. Classification is organized as a two-level procedure: the first level separates glacier from non-glacier pixels, and the second level separates clean ice from debris-covered ice. The classifier is fitted with the 2020 observations and then transferred to the 2021–2026 images without year-specific retraining. The complete framework produced 89% overall accuracy; the two hierarchical stages reached 91% and 96%, respectively. Mapped glacier area declined from 12.38 km² in 2020 to 7.30 km² in 2026, a reduction of 5.08 km² or 41.05%. These findings indicate that the proposed combination of spectral, terrain and hierarchical information can support repeatable glacier monitoring using openly available satellite and DEM datasets.