Neural network approach to the classification of urban images
Over the past few years considerable research effort has been devoted to the study of pattern recognition methods applied to the classification of remotely sensed images. Neural network methods have been widely explored, and been shown to be generally superior to conventional statistical methods. However, the classification of objects shown on greylevel high resolution images in urban areas presents significant difficulties. This thesis presents the results of work aimed at reducing some of these difficulties. High resolution greylevel aerial images are used as the raw material, and methods of processing using neural networks are presented. If a per-pixel approach were used there would be only one input neuron, the pixel greylevel, which would not provide a sufficient basis for successful object identification. The use of spatial neighbourhoods providing an m x m input vector centred on each pixel is investigated; this method takes into account the texture of the pixel's neighbourhood. The pixel's neighbourhood could be considered to contain more that textural information. Second order methods using mean greylevel, Laplacian and variance values derived from the pixel neighbourhood are developed to provide the neural network with a three neuron input vector. This method provides the neural network with additional information, improving the strength of the relationship between the input and output neurons, and therefore reducing the training time and improving the classification accuracy. A third method using a hierarchical set of two or more neural networks is proposed as a method of identifying the high level objects in the images. The methods were applied to representative data sets and the results were compared with manually classified images to quantify the results. Classification accuracy varied from 69% with a window of raw pixel values and 84% with a three neuron input vector of second order values.