Deep learning for wind vector determination

Numerical Weather Prediction (NWP) is a process of using numerical simulation to track and predict weather patterns over time. NWP is presented with a number of challenges in that it aims for precision in a situation that is by its nature uncertain and influenced by factors that are very difficult to enumerate. As an alternative, Empirical Weather Prediction (EWP) attempts to use observational data to construct models of weather phenomena. We evaluate a deep learning method applied to EWP to make wind vector determinations from radiometric data using data collected from Hurricane Sandy. Our approach uses unsupervised pre-training of stacked autoencoders to construct multilayer perceptrons. We then discuss the role of our approach as an important step in positioning EWP as a viable alternative to NWP.