Ocean surface wind retrievals from special sensor microwave imager data with neural networks

Abstract

Several fully connected, feed forward neural networks trained on a set of special sensor microwave imager examples matched with buoy winds have yielded retrieval accuracies considerably better than those achieved by the current operational method. Equations and coefficients for using two of these networks, each with four input brightness temperatures and a hidden layer containing two neurodes, are given for implementation in wind retrieval codes. The first demonstrated an rms retrieval error of 1.41 m/s at a reference height of 19.5 m using an independent data set representing clear sky conditions. The second network yielded rms retrieval accuracies of 2.39 m/s under adverse weather conditions. This represents a factor of more than 2 improvement over the alternate algorithms that were examined for nonclear conditions.