Comparison of back propagation and binary diamond neural networks in the classification of a Landsat TM image

Multi-layer perceptron neural networks using Back Propagation training methods (often referred to as MLP or Back Propagation neural networks) are currently the neural algorithm of choice for classifying remote sensing images. This work examines problems that exist with the Back Propagation algorithm and raises questions about the suitability of this algorithm for nonexpert users. The utility of the Back Propagation neural network is compared with that of the simpler, recently introduced Binary Diamond neural network algorithm. The classification abilities of the two algorithms are tested in attempting to classify a Landsat TM image from the Kanagawa region of Japan. The results of this classification exercise show that the Binary Diamond algorithm is simpler to use and that it can be trained more rapidly. The Back Propagation algorithm produced better results in this example classification, but the differences found may not justify its increased complexity for some users.