Analysis of large multi-dimensional data with a backpropagation neural network

The feasibility of using the neural network technique of backpropagation for analyzing the large multidimensional data acquired by satellites is discussed. Techniques are developed for selecting a training data set and accelerating the convergence during the learning phase. Attention is directed to the computational efficiency; both the computer code and the algorithm are carefully examined. A technique for proper selection and preprocessing of the data is developed to provide a good training set. An empirically based equation for selection of a stable learning rate is also presented. The adaptive backpropagation learning technique developed is 5-10 times faster than standard backpropagation