Time-series Fusion of Optical and Sar Data for Snow Cover Area Mapping

We have developed a new approach based on modelling and assimilation to combine SAR and optical data for snow cover area mapping. SAR data would typically be acquired a few times a week, while optical data is acquired daily but is limited by cloud cover. The algorithm we use analyses the current time series to estimate the current Fractional Snow Cover (FSC) per pixel. A set of snow states is defined. Each snow state has a corresponding reflectance model for optical data and a backscatter model for SAR data. The snow states defined are ‘dry snow, full snow cover’, ‘wet snow, full snow cover’, ‘fractional snow cover’ and ‘snow-free ground’. A Hidden Markov Model (HMM) has been established to compute the likelihood of a transition from one state to another, given the current observations. The backscatter and reflectance observations are processed by an algorithm comparing them to their respective models given by the current state. Based on this, the most likely current FSC is calculated for each pixel being analysed. Each pixel is processed independently and might therefore be in different stages (which is typical for mountainous terrain). The approach has been tested for a mountain plateau in South Norway combining Terra MODIS and ENVISAT ASAR from four snowmelt seasons (2003-2006). The results indicate that it is possible to obtain consistent results of high accuracy from the combination of the two sensors. Further work includes testing and tailoring of the approach to areas with steeper terrain. INTRODUCTION The snow cover has a substantial impact on the interaction processes between the atmosphere and the surface, thus the knowledge of snow variables is important in climatology, weather forecasting, and hydrology. In mountainous areas and in the northern Europe, snowfall is a substantial part of the overall precipitation. In order to perform sustainable management of water, in particular for hydropower production and flood protection, information on the snow cover is mandatory. The first experiments trying to combine SAR and optical data for snow cover area mapping took place about 15 years ago. So far, no published approach has worked very well due to the very different characteristics of the two sensor types. While the SAR signal is dominated by the dielectric properties of the medium measured and its geometrical properties at the scale of the wavelength, the optical sensor is sensitive to reflection, absorption and scattering properties of the snow grains in the top level of the snowpack. Hence, the sensors are measuring entirely different physical phenomena. The latest generation of optical and SAR sensors has opened for multi-sensor time-series mapping of snow cover. A few algorithms for this have been published. Raggam, Almer and Strobel (1994) demonstrated how snow cover retrieved from multi-parameter airborne SAR and SPOT HRV can be combined. Koskinen et al. (1999) analysed a time series of NOAA AVHRR and ERS-2 SAR images. However, they did no actual combination of the two other than studying how the snow cover developed as observed by the two sensors. Tait et al. (2000) developed a true combination of data from two sensors to produce a snow map. NOAA AVHRR data and SSM/I data were analysed together with climate station data and a digital terrain model in a decision tree in order to produce continental-scale snow maps for North America. The lack of access to frequent acquisitions of both SAR and optical data changed with the launches of Radarsat and ENVISAT (with ASAR) that in wide swath modes are able to deliver freEARSeL Land Ice and Snow Special Interest Group Workshop, 11-13 February 2008, Berne 2 quent coverage for a given geographical area. This allows multi-sensor fusion with optical sensors like AVHRR and MODIS on a frequent basis. Examples of such fusion can be found in Solberg et al. (2004a and b). The optimal situation is when optical and SAR sensors are on the same platform, which ensures acquisitions under exactly same conditions. The only satellite platform delivering such data currently is ENVISAT. ASAR and MERIS can be acquired simultaneously. However, cloud detection over snow-covered surfaces is not possible with MERIS, and the alternative sensor AATSR has a much smaller swath width. So this gives no practical/operational solution to the problem. Anyway, an example of using AATSR for detecting clouds in a MERIS sub-scene can be found in Tampelini et al. (2003). An example of the use of ASAR and MERIS in combination can be found in Solberg et al. (2004b). METHODS A serious challenge of multi-sensor fusion algorithms is that the optical and SAR sensors measure different physical phenomena. The effects from photon scattering, transmission and absorption near the snow surface at the snow-grain-size level dominate the optical snow spectrum. The radar signal is dominated by effects due to dielectric properties of the snow medium as well as snow surface roughness (for wet snow) or a combination of the snow pack structure and the ground below. In addition there are contributions from the bare ground surface for fractional snow cover conditions. When blending the snow cover fraction (SCF) retrieved from these two types of sensors into a fractional snow cover product, heterogeneities will easily appear as shown in the example in Figure 1, which is based on the algorithm in Solberg et al. 2004a. This is a problem in several applications. Variability in the retrieved parameter that is not related to the true SCF may create wrong interpretations when the snow cover is used as an indicator for climate change or as a variable in a hydrological model.