Fusion of high spatial and high temporal snow surface properties from satellite observations for estimating snow water equivalent

Abstract:

Snow surface properties including snow cover fraction, snow grain size, snow albedo, and the impact of light absorbing particles on snow albedo measured by satellite can be used to validate or calibrate models of snow water equivalent or be assimilated into these models directly. The tradeoff between spatial and temporal resolution of available satellite data for estimating these snow surface properties remains a challenge. Observations from satellites are available daily from MODIS at 500 m and VIIRS at 1 km, but snow surface properties vary at a much finer scale. Satellites such as Landsat 5, 7, and 8 and Sentinel 2a and 2b provide higher spatial resolution information but even when combined do not observe the Earth’s surface daily. Snow cover and snow albedo can change quickly especially during the accumulation and melt seasons so daily or even sub-daily observations are needed to capture variability. We show that estimates of snow cover from these satellites using spectral mixture analysis applied to multispectral satellites at any spatial resolution are unbiased relative to each other. We combine snow surface properties from these different satellite sources using a two-stage random forest algorithm and a suite of predictor variables. The training model uses only 1-5% of the high spatial resolution data and is assessed using the other 95-99% of the data. For snow cover, our technique shows an overall accuracy of 97% while mean difference and RMSE are less than 1%. Building on our work with snow cover, we apply this technique to other snow surface properties such as grain size used for snow albedo. We analyze a time series of data in the Sierra Nevada during the MODIS period of record and test with modeled snow water equivalent from reconstruction at both 500 m and 30 m spatial resolution to understand the utility of the fused snow surface properties.