Comparison of neural network algorithms for remote sensing applications

Neural networks (NN) are rapidly gaining acceptance within many disciplines including remote sensing. This is due primarily to the use of more efficient training algorithms and better understanding of their capabilities and limitations. NNs have been found to be robust and well suited for the wide variety of data found in remote sensing. They are used in the classification of land-use as well as in the estimation of target properties. The type of neural network to use, the method of training, and the capabilities and limitations of the various neural networks are still the subject of some debate. This paper examines the performance of four neural network algorithms found in the literature: (1) the traditional multi-layer perceptron trained with the well known back-propagation algorithm of Werbos (1974) and Rumelhart and McClelland (1988) (BP-MLP), (2) the dynamic-learning NN (DLNN) of Tzeng et al. (1994) trained using a linear Kalman-based technique, (3) the fast learning NN (FL-MLP) of Manry et al. (1994), and (4) the polynomial-based functional link NN (FLNN) of Pao (1989).