Abstract
Satellite-based snow cover area (SCA) time-series provide relevant spatial information about snow presence. However, SCA derived from satellite sources still suffers from several classification errors. Notable issues include the underestimation of snow presence in forested areas, where snow beneath the canopy is not visible to satellites (i.e., snow on the ground), and missed snow detection in shadowed regions, which can cover large portions of mountainous areas, particularly during winter. Additional challenges arise from the trade-off between spatial and temporal resolution. Achieving both daily coverage and high spatial resolution requires merging data from multiple satellite sources through downscaling and gap-filling techniques, which can introduce further errors, such as the missed identification of snow patches at the end of the season when using lower-resolution satellites for those specific dates. When SCA data are integrated into snow models through data assimilation schemes, these errors and inconsistencies can propagate, undermining the accuracy of model outputs. For example, errors during melting periods, particularly at the end of the snow season, that result in intermittent snow presence caused by snow patches misclassification, are especially problematic. These challenges complicate the integration of SCA into simple modeling frameworks, such as backward Snow Water Equivalent (SWE) reconstruction, where inconsistencies in the correct determination of the snow disappearance date can lead to significant errors. A previous attempt to harmonize SCA time-series relied on in-situ snow depth data to define accumulation and ablation periods, assuming these periods are uniform across the region (Premier et al., 2023). The method corrected inconsistencies represented by snow cover increases without corresponding accumulation or decreases during periods without melting. While effective for small areas, this approach is limited by the availability of in-situ data and the validity of its assumptions over larger regions. In this work, we extend the previous harmonization technique by using observed meteorological grids variable in time and space. In detail, for this test we used precipitation and temperature from the EMO-1 dataset, used as forcing in the European Flood Awareness System (EFAS). Despite potential inaccuracies in these meteorological fields, we assume they are reliable in terms of timing and spatial extent. This method ensures the coherence of SCA time-series with snow state dynamics (accumulation, ablation) by imposing logical checks on pixel class changes. For example, during accumulation, a pixel can transition from snow-free to snow only if a snowfall occurs on bare ground or remain as snow if precipitation falls on existing snow. During ablation, a pixel can transition from snow to snow-free only if snow cover disappears or remain as snow if partial melting occurs. Misclassified transitions are corrected by analyzing an appropriate time window and applying a majority rule to determine the most frequent label. The harmonization process is performed iteratively, considering previous steps to avoid introducing new inconsistencies during corrections. Logical masks streamline the computation, enhancing efficiency and scalability. By providing coherent, harmonized SCA time-series, this approach improves the integration of satellite-derived snow cover data into snow models, eliminating the need for further pre-processing.
References
Premier, V., Marin, C., Bertoldi, G., Barella, R., Notarnicola, C., & Bruzzone, L. (2023). Exploring the use of multi-source high-resolution satellite data for snow water equivalent reconstruction over mountainous catchments. The Cryosphere, 17(6), 2387-2407.