Computationally Efficient Retrieval of Snow Surface Properties From Spaceborne Imaging Spectroscopy Measurements Through Dimensionality Reduction Using K-Means Clustering

Snow albedo is a crucial component to the energy balance of our seasonal snowpack on Earth, reflecting incoming solar irradiance and altering the Earth system. Seasonal snow surface properties undergo constant change (e.g., melt-freeze cycle, faceting, sublimation, and windblown compaction) and have high spatial variability, especially in mountain regions, making it difficult to scale ground measurements. Snow albedo, fractional snow-covered area, and snow-specific surface area can be modeled using top of atmosphere radiance measurements from spaceborne imaging spectroscopy. We model these snow properties by testing inversions geared specifically for complex topography, as well as incorporating the impacts of the mixed snow pixel. Additionally, we avoid computing every pixel in an image by creating tens-of-thousands of k-means clusters based on the rounded values of the cosine of the local illumination angle to the nearest ten-thousandth digit. This computation is further sped up by leveraging message passing interface to scale with more nodes. We present this work as an open-source algorithm, which we refer to as Global Optical Snow properties via High-speed Algorithm With K-means clustering (GOSHAWK). We validate our algorithm with PRecursore IperSpettrale della Missione Applicativa L1 radiance imagery across eight sites in the Northern Hemisphere from 2021 to 2023 and compare outputs with field spectroscopy, specific surface area measurements, airborne LiDAR surveys, and four-component net radiometers. More work in algorithm development and calibration–validation work is needed in steep terrain and dense canopy to improve snow property retrieval prior to the Surface Biology and Geology and Copernicus Hyperspectral Imaging Mission for the Environment missions.