Simulation of snow stratigraphy using full-waveform inversion applied to data from an upward-looking radar system

Snow stratigraphy is a key contributing factor for assessing avalanche danger, but so far only destructive methods \ncan provide this kind of information. Furthermore, continuous monitoring of the temporal evolution of the \nsnowpack is not possible with destructive methods. Radar technology provides information on the snowpack nondestructively \nand allows deriving internal snow properties from its signal response. In our previous work, we \ndemonstrated that it is feasible to quantitatively derive snowpack properties relevant for avalanche formation and \nmonitor their evolution in time using an upward-looking ground penetrating radar system (upGPR) that was buried \nin a wooden box underneath the snow. Reliable results could only be obtained for the time when the snow cover \nwas dry. In addition, to determine some properties, we still needed additional information such as independently \nmeasured snow height or modeled snow density. Hence, the system was not yet able to provide information from \navalanche starting zones, since this type of information is generally not available in avalanche-prone terrain. To \nfully exploit the information content of upGPR data, and thus to at least partially compensate for the lack of information, \nwe applied full-waveform inversion (FWI) techniques. We refined the model of the snowpack by repeated \nforward modeling the waveforms and updating the model parameters to match it with recorded data. The forward \nmodel took into account both the effect of the snow density on the velocity of the electromagnetic wave, as well \nas the influence of snow wetness on the attenuation. This allowed the density and the liquid water content for each \nlayer in the snowpack to be determined. As we conducted a measurement every 3 hours (every 30 minutes as soon \nas the snowpack became wet), we could also simulate the temporal evolution of the density and the liquid water \nprofiles. The method worked without assumptions or external measurements, even when the snow cover was wet.