The results of the classification of SPOT high resolution visible multispectral imagery using a neural network are presented. The test site, located near Taoyuan in northern Taiwan, is in an agricultural area containing small ponds, bare and barren soils, vegetation, built-up land, and man-made buildings near the sea shore. The classifier Is a dynamic learning neural network (DL) using the Kalman filter technique as ifs adaptation rule. The network's architecture consists of multilayer perceptrons, i.e., feed-forward nets with one or more layers between the input and output nodes. Selected data sets from 512- by 512-pixel three-band images were used to train the neural nets to classify the different types of land cover. Both simulated and real images were used to test classification performance. Results indicated that the DL substantially reduces the training time, compared to the commonly used back-propagation (BP) neural network whose slow training process prevents it from being used in certain practical applications. As for classification accuracy, the results were excellent. We concluded that the use of a dynamic learning network provides promising classification results in terms of training time and classification rate. In particular, the proposed network significantly improves the practicality of land-cover classification.
