A Neural Network Technique for Improving the Accuracy of Scatterometer Winds in Rainy Conditions

Abstract

We exhibit a technique for improving wind accuracy in Ku-band ocean wind scatterometers in the presence of rain. The technique is autonomous in that it only makes use of measurements made by the scatterometer itself, so that no colocation of an external data set (e.g., rain radiometers) is required to perform the correction. The only inputs to the technique are the normalized radar cross-section measurements for each wind vector cell, the cross-track distance of the cell as a proxy for measurement geometry, and the nominal retrieved wind vector for the cell without rain correction. This last input is used to avoid modifying winds not contaminated by rain. The technique was applied to QuikSCAT data for the month of January 2008, resulting in a marked improvement to rainy data. For data that were determined to be rain contaminated by the Jet Propulsion Laboratory rain flag, the rms speed error with respect to National Data Buoy Center buoy winds improved from 8.9 to 3.5 m/s for colocations within 25 km. The rms speed error in rain also improved when compared with the European Centre Medium-Range Weather Forecast winds from 7 to 3 m/s. Data that were not flagged as rain contaminated were not significantly changed, despite the fact that the technique does not make use of the rain flag. The technique was able to distinguish between rain-contaminated wind cells and rain-free wind cells and to substantially improve the wind speed accuracy of the former using QuikSCAT data alone without recourse to any external information about the extent of the rain.