Information content of binary snow/non-snow data for melt prediction

Snow spatial evolution and distribution in mountainous terrains is difficult to capture since these remote areas are often inaccessible. Nevertheless, they constitute important features for water resources management. To face this issue, the concept of data assimilation has recently been introduced in hydrology to reconstruct snow evolution with a snow model, given observed snow states. So far, snow cover area observations from satellites imagery or field measurements have been employed in this approach. A procedure has been developed in order to extract the snow evolution at the catchment scale from terrestrial photography. The pictures are first geo-referenced and snow is then detected by the mean of a clustering algorithm to finally establish binary snow cover maps. The benefit of using these binary maps, over the usual observations, in a data assimilation framework has been assessed by comparing two commonly used methods: ensemble Kalman filter (EnKF) and particle filter (PF). Binary snow maps over the melting period in Val Ferret catchment (Switzerland) has been computed for this purpose and introduced in a snow model. Even though both methods brought improvement in estimating snow evolution, EnKF showed limitations in binary data manipulation. In comparison, PF proved to be more efficient and allowed an important reduction of associated errors in the snow estimation. The sensitivity of the method still has to be evaluated to be able to propose a reliable approach handling binary observations. But the use of terrestrial photography could bring significant improvement in snowpack reconstruction as it provides accessible high definition products.