Neural Network Multi-Parameter Algorithms to Retrieve Atmospheric and Oceanic Parameters from Satellite Data

A new approach in satellite retrievals, multi-parameter empirical retrievals using neural networks (NNs), is introduced. Conventional methods for using satellite data involve solving an inverse (or retrieval) problem and deriving a transfer function which relates a single geophysical parameter of interest (e.g., surface wind speed over the ocean), to a vector of satellite measurement (e.g., SSM/I brightness temperatures). Usually satellite measurements contain information on several related geophysical parameters. Therefore, in principle, under favorable conditions, these parameters can be retrieved simultaneously from the vector of satellite measurements. Single-parameter retrieval algorithms: (1) do not utilize the information about other geophysical parameters which is available in satellite measurements (2) produce retrievals for a single parameter (e.g., wind speed) which do not correspond to any particular atmospheric state or ocean surface state. These retrievals correspond to some unknown "mean" atmospheric and surface states which can not be specified without additional information. Thus, single-parameter retrieval algorithms effectively average over an ensemble of atmospheric and surface states for all of the related geophysical parameters except for the one which is retrieved. This averaging gives rise to additional "artificial" errors in this single retrieved parameter which do not arise in the multi-parameter approach. It is shown that single-parameter retrievals, compared with multi-parameter retrievals, contain significant additional "artificial" systematic and random errors. These errors may be avoided using multi-parameter retrieval algorithms. Neural networks are well suited for developing such multi-parameter retrieval algorithms. The NN approach for developing empirical multi-parameter algorithms is presented. A new NN empirical algorithm (OMBNN3) which simultaneously retrieves four geophysical parameters: surface wind speed, columnar water vapor, columnar liquid water, and sea surface temperature from five SSM/I (Special Sensor Microwave Imager) brightness temperatures (T19V, T19H, T22V, T37V, and T37H) is presented and compared with several single-parameter algorithms to illustrate advantages of the multi-parameter approach. Although geophysical parameters considered have previously been retrieved separately by other techniques, it is the simultaneous retrieval procedure (using NNs) that allows the information from one parameter to contribute to a better estimate of the other parameters. The parameters retrieved by the NN, when analyzed together, can provide information about synoptic weather patterns over the oceans that is more comprehensive and internally consistent than that from a single parameter algorithm. Data produced by multi-parameter OMBNN3 algorithm are currently operationally used for assimilation into numerical weather prediction models and to improve the interpretation of weather conditions at the ocean surface and can be seen at: http://polar.wwb.noaa.gov/winds/ every day. The OMBNN3 algorithm was incorporated as the operational algorithm at National Centers for Environmental Prediction in April of 1998.