An application of autocorrelation to the identification of tree types from satellite images

In order to classify different types of tree growth in Eire, data from three bands, single-date SPOT satellite images were analysed. Initially the analysis was only statistical using a variety of methods to differentiate between, for example, pine and oak. It was found that simple statistical approaches to pixel values within the three bands did not give sufficiently different values to be able to classify with the degree of confidence required. An alternative approach was developed which used initially spatial and then statistical techniques. It uses a relatively simple autocorrelation technique to describe local texture features as a small vector. The process is not as computationally intensive as some other texture processes. The technique creates, from a rectangular texture region (N×M), a vector size N+M-6 comprising autocorrelation statistics from the block. This vector is computed for a training area and then used as a search statistic in the rest of the image. An N×M rectangular window is stepped through the image, creating a new vector for each step. The product moment correlation between the sample window vector and the training area vector is plotted for each pixel position. The process is, as far as we can discover, a new one for creating a localised texture measurement and hence a new band which can then be incorporated into the statistical model, or used alone to classify an image. We have shown that this new band does add further significant classifying power to the process