Abstract
The ESA CryoClim project previously developed a binary snow cover product for long-term systematic climate monitoring. The algorithm is based on fusion of AVHRR GAC and SMMR + SSM/I + SSMIS data. The product set consists of a time series (1982-2015) of daily, snow cover maps of 5 km spatial resolution with full global coverage independent of clouds and polar-night darkness. The ESA Climate Change Initiative (CCI) Snow ECV project has now developed an advancement to a fractional snow cover (FSC) product.
The original binary algorithm is based on a hidden Markov model (HMM) simulating snow states based on the satellite observations. The model is described by the different states and the possible transitions between these states. The states are given by probability density functions and the transitions by transition probabilities. A Viterbi algorithm is used to find the most likely snow cover sequence throughout the hydrological year at each grid cell. The HMM solution represents not only a multi-sensor model but also a multi-temporal model.
The advancement from binary to fractional snow cover has followed two main paths: First, we introduced more HMM states to be able to classify the snow cover into 10% FSC intervals. However, introducing 100 states to obtain 1% FSC intervals would not give a stable model. For obtaining higher precision, we have interpolated between HMM states using a secondary Viterbi sequence. The two probabilities are used as weights to estimate the FSC.
The algorithm has been implemented to run on a supercomputer as three components. The optical and passive microwave radiometer (PMR) data are processed in two processing chains giving the probability of snow from each type of data, respectively. The probabilities are applied in the HMM multi-sensor multi-temporal model generating the fractional snow cover map. The map also includes a per-grid-cell estimate of the FSC uncertainty. A new 38-year time series (1982-2019) of daily, global products was generated recently and will be made freely available to the climate community.
Two different validation approaches have been applied. The first was based on using in-situ data from stations (points) and snow courses, following the same approach as in past CryoClim activities. The other approach closely followed the validation technique developed by the Snow CCI project of using classified high-resolution data based on Landsat imagery as a reference. Based on the in-situ data approach, very similar results to those obtain previously for the binary product, were found. Yearly overall accuracy was found to be mostly between 90 and 94%, with some exceptions. Using Landsat data, the overall accuracy was found to be high with an RMSE in the order of 16% and a bias lower than 2.4%. For the FSC uncertainty estimates, a general overestimation of 2-3% was found.