Effects of system errors on combined MM/IR neural network inversion of surface snow properties

The University of Nebraska has recently developed a neural network inversion algorithm for the estimation of surface snow properties, viz., surface roughness, wetness, and average grain size. The algorithm uses concurrent measurements of the near-infrared reflectance and millimeter-wave backscatter of the snow surface. The performance of the inversion algorithm was found to be satisfactory under noise-free conditions. However, under operational conditions, noise is invariably present in the data, and the addition of noise causes errors in estimation. The performance of the inversion algorithm was investigated under noise-added conditions. A parameter that was varied was the signal-to-noise ratio. Inversions of free-water content and the grain size were relatively robust, while the surface roughness estimation was very sensitive to added noise. The results of our study can be useful in setting bounds for system performance for accurate snow surface parameter inversion.