Retrieval of snow physical parameters by neural networks and optimal estimation: Case study for ground-based spectral radiometer system

A new retrieval algorithm for estimation of snow grain size and impurity concentration from spectral radiation data is developed for remote sensing applications. A radiative transfer (RT) model for the coupled atmosphere-snow system is used as a forward model. This model simulates spectral radiant quantities for visible and near-infrared channels. The forward RT calculation is, however, the most time-consuming part of the forward-inverse modeling. Therefore, we replaced it with a neural network (NN) function for fast computation of radiances and Jacobians. The retrieval scheme is based on an optimal estimation method with a priori constraints. The NN function was also employed to obtain an accurate first guess in the retrieval scheme. Validation with simulation data shows that a combination of NN techniques and optimal estimation method can provide more accurate retrievals than by using only NN techniques. In addition, validation with in-situ measurements conducted by using ground-based spectral radiometer system shows that comparison between retrieved snow parameters with insitu measurements is acceptable with satisfactory accuracy. The algorithm provides simultaneous, accurate and fast retrieval of the snow properties. The algorithm presented here is useful for airborne/satellite remote sensing.