Neural network retrieval of winds from combined surface and satellite observations

A variety of methods have been applied to deriving atmospheric winds from satellite observations. Cloud and moisture tracking techniques use the movement of features of cloud and moisture images to derive horizontal winds. Geopotential heights derived from soundings can be used to compute gradient winds, geostrophic winds, or quasi-geostrophic winds. Radar scatterometry can yield sea surface wind. Limb scanning Doppler interferometry has also shown promise for upper air winds. Still to be tried is satellite based Doppler lidar measurement of tropospheric winds. Each of these methods, except possibly the yet untried Doppler lidar-in-space, has significant limitations in accuracy or coverage.

Our work is aimed at developing methods for determination of soundings of atmospheric parameters using a highly mobile system incorporating surface sensors, wind radar, microwave radiometry, and satellite data receivers, with minimal use of rawinsondes. The primary deficiency of this system at present is the inability to produce accurate winds above the range of the wind radar. The present work is part of a series of investigations in an attempt to overcome that deficiency.

We have been experimenting with neural network methods for retrieving winds from satellite sounder data. While experiments to date have yielded errors comparable to those achieved by other sounder based methods, a primary motivation of the neural network experiments was the idea of using them for fusion of data from diverse sources. Our current experiments use surface-based radar winds in combination with satellite winds in an attempt to improve the accuracy of the retrievals. In these experiments, real or simulated radar wind measurements and satellite radiances have been used as inputs and rawinsonde measurements are used as "truth" comparisons.