Ocean Surface Wind Field Retrieval From SSM/I Data Using Neural Networks

Based on the SSM/I brightness temperature data, three kinds of neural network models of different structures, which are Single-parameter Artificial Neural Network (SANN), Multi-parameter Artificial Neural Network (MANN) and a new Compound Multi-parameter Artificial Neural Network (CMANN), are built to retrieve the global ocean surface wind speed at real time. The result shows that, by adding the brightness temperature of the 85GHz vertical and horizontal polarization channels to the neural networks, the retrieval precision of the ocean surface wind speed is improved. The rms error of CMANN decreases with the increasing of the wind speed and therefore there is a better performance at the high wind speed (>15m/s) than at the low wind speed. The structure of the NN, such as layers, nodes and activation functions, are taken into account to enhance the performance of the wind speed retrievals. In addition, three new NN methods are developed, which are the classified NN model, the circular NN model and the NN ensemble model, to reduce the rms error of the wind speed. The analogous NN methods are used to retrieval the wind direction. The data used to train and test the NN models is from the matchup of the data of the TAO, the NDBC and the SSM/I brightness temperature. The wind speed range of the CMANN algorithm is 0-25m/s. Compared with the buoy wind, the rms errors are 1.61m/s and 1.46m/s under clear plus cloudy conditions and clear conditions, respectively. The rms error between the retrieved relative wind direction and the buoy measurements is about 40°.