Classification of multispectral imagery using dynamic learning neural network

The paper presents the results of classification of SPOT high resolution visible (RHV) multispectral imagery using neural networks. The test site, located near Taoyuan county of the northern Taiwan, is an agriculture area containing small ponds, bare and barren soils, vegetation, built-up land, near shore sea, and man-made buildings. The classifier is a dynamic learning neural network (DL) using the Kalman filter technique as an adaptation rule. The network architecture involves multi-layer perceptrons, i.e., feed-forward nets with one or more layers of nodes between the input and output nodes. The methodology of selection of training data sets is addressed. Then, accordingly, selected data sets from a 512×512 pixels three-band image are used to train the neural nets to categorize different types of the land-cover. Both simulated and real images are used to test the classification performance. Results indicate that the DL substantially reduces the training time as compared to the commonly used back-propagation (BP) trained neural network whose slow training process is shown to impede it in certain practical applications. As for classification accuracy, the presented results are shown to be excellent. It is concluded that the use of a dynamic learning network gives very promising classification results in terms of training time and classification accuracy. In particular, the proposed network significantly improves the practicality of the land-cover classification