Estimating multiyear sea-ice concentration using passive microwave data and MLP neural networks

Three ice-type classification methods utilizing SSM/I passive microwave data were compared. Each applied a multilayer perceptron (MLP) neural network (NN) with OKEAN (radar and passive microwave) sea-ice learning data, a different learning algorithm based on, respectively, error back propagation and simulated annealing (M1), dynamic learning and polynomial basis functions (M2), and dynamic learning with two-step optimization (M3). M2 and M3 methods used the Kalman filtering technique. Our studies demonstrated, that for sea-ice inversions the modified MLP NN with M1 algorithm was more efficient because M2 and M3 algorithms caused overfitting. Multiyear (MY) sea-ice concentration maps were generated from SSM/I Tbs (19 GHz V, 19 GHz H and 37 GHz V channels) using modified MLP NN with M1 algorithm, OKEAN and ERS learning data. These maps were compared with respective MY sea-ice concentration maps developed using NASA Team algorithm (NTA). Our studies demonstrated the superiority of the NN method compared to the NTA.