Machine Learning for inpainting QuikSCAT winds in Hawaii's Lee Region

Ship and aircraft observations (e.g., Smith and Grubisic 1993) have revealed a wake behind Hawaii’s Big Island. Two mountains, Mauna Kea and Mauna Loa, induce counter-rotating lee vortices and westerly reverse flow that opposes the prevailing easterly trade winds in the summer season. Unfortunately, the westerly reverse flow is absent in QuikSCAT and ASCAT satellite wind climatologies. We hypothesize that JPL/RSS QuikSCAT wind direction errors in the Big Island wake are due to reliance on coarse resolution nudge (NCEP) winds. A dynamic, summer-time mesoscale Weather Research and Forecasting (WRF) model has been applied to the Hawaii region and the reversal region exists in this model (Zhang et al. 2012). From this WRF data, we have developed a masked region, in which the satellite winds are untrustworthy, and seek to develop a model that can be used to inpaint the untrusted region, given the winds in the surrounding region. Using the WRF winds as training data, we apply and compare multiple supervised methods to reconstruct the lee wind region from the surrounding region 1) Linear Maximum and Cross Correlation Analysis (MCA/CCA), 2) Feed-Forward and Convolutional Neural Networks. We then compare length of training times to determine the optimal method, given a varied amount of dynamic model generated (WRF) data. This work aims to determine the benefit of non-linear relationships inherent to deep learning methods, and if there are added gains of skill over traditional methods for spatial reconstruction. We also seek to examine the convolutional neural networks to recognize similar detected patterns inherent to the empirical orthogonal functions of MCA/CCA.