Retrieval of ocean winds from satellite scatterometer by a neural network

This paper presents the reconstruction of a wind field from three-beam scatterometer measurements under the framework of a neural network. A neural network is adopted to implement the inversion of a geophysical model function (GMF) that relates the scatterometer measurements of normalized radar cross section to surface wind speed and direction. To illustrate the functionality and applicability of the neural network, a set of wind fields generated by means of the Monte Carlo simulation are used. At each sample point of the wind field, the speed and direction are simulated. Then, a GMF CMOD4 is used to synthesize the normalized radar cross section at three pointing antennas according to the ERS-1 configuration. In such a case, the neural network is constructed to model the inverse transfer function. For inputs, a pixel-based and area-based scheme are considered. The network training is accomplished by mapping input-output pairs that are randomly selected from the database of simulated wind fields. The effectiveness of the neural network as an inverse transfer function is validated. Four data sets of ERS-1 scatterometer data over the western Pacific were selected for case study. Intercomparison with other methods concludes that the use of neural network has its indispensable advantages and better retrieval accuracy can be obtained.