Mapping snow cover at fine resolution in complex and forested terrain

Abstract

Complex topography in mountain basins allows deep snow accumulation in depressions throughout the winter season. During the melt period, snow on exposed terrain disappears faster, leaving behind a patchy mosaic of lingering snow cover. These persistent snow patches help sustain late-season streamflow during the drier summer months. In this work, we explored methods to observe heterogeneous snow cover in a complex subbasin in the Sierra Nevada, CA, using Planet Labs, Inc. 3 m resolution commercial optical satellites and an existing random forest model (Yang et al., 2023). Here, we tested two different model training configurations and a spatio-temporal post-processing approach to improve snow mapping throughout the season for three separate years, paying particular attention to snow in forested areas. In this work, we used lidar only for evaluation so that the methods used can be applied anywhere. The new model training approach and the spatio-temporal post-processing approach introduced in this work both performed well, showcased better results than prior methods, and improved F1 scores in the forest by 0.11 and 0.09, respectively. We then compared basin-wide snow-covered area from the 3 m resolution snow maps with the Moderate Resolution Imaging Spectroradiometer (MODIS) Snow-Covered Area and Grain Size (STC-MODSCAG) product (Rittger et al., 2020), which is coarser resolution but more readily available. In general, they compared well, but on average our snow maps reported snow disappeared over 3 weeks later than STC-MODSCAG, and STC-MODSCAG missed ~303,000 m2 of snow cover that the PlanetScope snow maps identified. Overall, our results demonstrate the capability of high-resolution imagery for detecting snow patches relevant for ecology and the potential for improved snow cover mapping in forested basins using the model training methods or the spatial post-processing methods introduced in this work.