Wind reconstruction from ERS-1 scatterometer data using neural network

A dynamic learning neural network is adopted to relate the normalized radar backscattering coefficient /spl sigma//sup /spl deg// to the wind vector. By given the average wind speed and wind direction and their standard deviations, a set of test wind vector fields are simulated. The corresponding /spl sigma//sup /spl deg// values are calculated according to the well known empirical model named CMOD-4. To improve the accuracy of the estimated wind direction, spatial information is considered in the reconstruction process. Thereafter, the neural network is iteratively trained by the input-output pairs generated from the test wind fields. Finally, the wind parameters are reconstructed upon applying /spl sigma//sup /spl deg// values to the well trained neural network. Experimental results indicate that neural network is an effective tool for wind reconstruction.