Item: FROM SCOTTISH SNOW PATCHES TO GLOBAL SNOW COVER: A DEEP LEARNING APPROACH TO SNOW COVER FRACTION RETRIEVAL
-
-
Title: FROM SCOTTISH SNOW PATCHES TO GLOBAL SNOW COVER: A DEEP LEARNING APPROACH TO SNOW COVER FRACTION RETRIEVAL
Proceedings: Proceedings, International Snow Science Workshop, Whistler, BC, Canada, 2026
Authors:
- Leam Howe [ School of GeoSciences, University of Edinburgh, Edinburgh, UK ]
- Richard Essery [ School of GeoSciences, University of Edinburgh, Edinburgh, UK ]
- Elliot J. Crowley [ School of Engineering, University of Edinburgh, Edinburgh, UK ]
Date: 2026-09-28
Abstract: Optical snow mapping products are often developed and validated on regions with deep, clean, continuous snowpacks under clear skies. Maritime mountain ranges with marginal snowpacks suffer from particularly harsh conditions for optical snow observations: fragmented cover, spectrally degraded snow, and near-permanent heavy cloud influence. In the case of cloud, quality is usually assured with a conservative cloud mask, but this discards useful observations. We quantify the cost of that trade-off and offer an alternative. Using 25 cm aerial surveys of the Scottish Highlands, we built an open snow cover fraction (SCF) reference dataset for Sentinel-2 comprising 2455 labelled 1 km chips from 37 scenes, deliberately retaining imagery through cloud gaps and heavy atmospheric interference rather than masking it away. On this dataset we trained SPUN, a U-Net with a ResNet50 encoder and a regression head that predicts continuous SCF per 10 m pixel. Evaluated on a spatially independent test set at the 20 m operational grid, SPUN and the operational Let-It-Snow (LIS) product are separated modestly under clear-sky conditions (snow-only MAE 19.10 % versus 23.26 %; F1 0.606 versus 0.493). Under all-sky conditions the gap becomes distinct: SPUN improves slightly (MAE 18.81 %, F1 0.747), while LIS recall falls to 0.013 and it recovers 1.6 % of the reference snow-covered area. Aggregated to the 1 km chip scale, SPUN tracks snow-covered area with R² = 0.934 against -0.238 for LIS. SPUN is trained only on late-season Scottish snow patches but, nonetheless, transferred to an independent 40-scene global dataset, reaching a macro-averaged F1 of 0.931 and a snow-only F1 of 0.892. We present this as a proof-of-concept rather than a finished operational tool, and argue that the limiting factor in optical snow retrieval is training data rather than model architecture.
Object ID: ISSW2026_P1.59.pdf
DOI: https://doi.org/10.15788/1790098860
Language of Article: English
Presenter(s): Leam Howe
Keywords: snow cover mapping, fractional snow cover, remote sensing, Sentinel-2, deep learning, maritime snowpack, Scotland
Page Number(s): 2278 - 2285
-