Synthesizing long-term satellite imagery consistent with climate data: Application to daily snow cover

Abstract

High-resolution daily snow cover estimation is challenging due to irregular satellite data availability. In this study, we create synthetic 30 m snow cover images for dates with no satellite data. It is based on the relationship between meteorological predictors and available clear sky 30 m Landsat/Sentinel-2 snow cover images. Our approach relies on the fact that while snow cover can vary in a matter of days, its patterns may repeat from year to year if meteorological characteristics are similar. In this context, we apply a K-nearest neighbor algorithm with a similarity metric tailored to estimating snow cover.The approach is tested with different data availability scenarios to generate twenty years of daily synthetic snow cover images for two regions of interest in Switzerland. The results are assessed based on a leave-one-out analysis, comparisons with PlanetScope, MODIS, and Copernicus High Resolution Snow & Ice Fractional Snow Cover On Ground (HRSI-FSCOG) images, comparisons with data collected at ground stations, and a comparison with a simple snow accumulation and melt model.The performance of our mapping approach suggests that it is possible to use recurrent snow patterns and climate reanalysis to generate missing data that are virtually and statistically indistinguishable from real observations. The results also indicate that low-resolution climate data, such as in the ERA5-Land reanalysis dataset, provide sufficient performance in the generated synthetic snow maps, opening possibilities for applications in areas with limited in situ snow or climatic data. Our approach could be used to create historical snow cover images for Pre-Satellite periods, or even for future periods by using climate projections.