Snow characterization at a global scale with passive microwave satellite observations

[1] The sensitivity of passive microwave satellite observations to snow characteristics is evaluated, between 19 and 85 GHz, for a winter season, for the Northern Hemisphere. The surface emissivities derived from the Special Sensor Microwave/Imager measurements are systematically compared with in situ snow measurements at 2784 stations, in North America and Eurasia. In addition, coincident satellite responses from active microwave sensors (ERS scatterometer) and visible observations (AVHRR) are also analyzed. Vegetation interferes with the signal that is received by the satellites. Snow emissivities also react to scattering by the snow grain growth that is related to the snow metamorphism during the winter. This phenomenon increases with frequency and is already very sensitive at 37 GHz. Passive microwaves at high frequency (85 GHz) are very sensitive to the presence of snow on the ground, even for very low snow depth. None of the tested satellite measurements is well correlated to the snow depth at a global scale, making snow depth retrieval from these observations very difficult on a global basis. The sensitivity of the satellite observations to snow characteristics depends on local conditions. To partly alleviate these difficulties, a neural network inversion scheme based on local statistics is developed to combine satellite observations, in situ measurements, and land surface model outputs. The combination of different wavelengths partly limits the ambiguities related to the individual sensitivity of each satellite observation to the various sources of variabilities. The final retrieval algorithm is compatible with an assimilation strategy that would better constrain the behavior of surface models. Finally, a clustering algorithm is applied to the suite of satellite observations and clearly shows a strong sensitivity to the snow characteristics and metamorphism during the winter. Characterization of the snowpack using satellite observation classification can yield qualitative information for snow model parameterization.