Scene-to-Scene classification variability in spatiotemporal glacier surface facies mapping in Svalbard

Spatiotemporal analysis of glaciological phenomena provides a robust assessment of a glacier’s health. Glacier surface facies (GSF) are direct indicators of the state of snow and ice within a glacier. However, long-term analyses of GSF via optical satellite products presents challenges stemming from data availability, weather conditions, and variations in mapping methodologies. In this study we present the first decadal analysis of GSF in Svalbard using optical satellite data. Utilizing unsupervised classification on images of the Vestre Broggerbreen glacier from 2013 to 2023, we identify and quantify the variations between facies observed on the images and their classified spatial distributions. The identified facies comprise of snow, firn, glacier ice, and dirty ice, with a fifth thematic class of shadowed snow. In certain imagery, snow and firn are labelled as ‘snow 1’ and ‘snow 2’ due to the derived reflectance and appearance of firn differing from established patterns in the literature, while remaining spectrally distinct from snow. Our analysis suggests that shadowed snow induces the most misclassification in overall assessment of GSF. Shadowed snow and dirty ice produce convoluted spectral reflectance that in combination with the overall darkening of the glacier severely misrepresented firn, underreported dirty ice, and produced inaccurate maps. Spatially, dirty ice was classified with a lower distribution in 2023 (0.49 km2) than in 2013 (0.58 km2). Moreover, firn/snow 2 was classified as a larger area in 2023 (1.01 km2). These results indicate a pressing need to identify long-term trends affected by scene-to-scene distortions. Our future experiments involve multi-decadal supervised analyses of GSF in the high Arctic involving image-specific discussions, thereby providing crucial information for refinement of supraglacial monitoring with multispectral data.