Texture Analysis of Remote Sensing Imagery with Clustering and Bayesian Inference

Texture is one of the most significant characteristics for retrieving visually similar patterns in remote sensing images. Traditional approaches for texture analysis are based on symbolic descriptions and statistical methods. This study proposes a new method to extract and classify texture patterns from multispectral Landsat TM satellite images using optimized clustering and probabilistic inference. After the images are preprocessed with Principal Component Analysis and decomposed into regions of interest, Gabor wavelets are computed for each region in the first component image to obtain texture feature vectors. An adapted k-means clustering algorithm with optimized number of clusters and initial starting centers generates training and testing data for Bayes Point Machine classifiers. The classifiers may run in the online mode for binary classification and the batch mode for multi-class classification. The experimental results show the effectiveness of the proposed classification method and its potentials in other image texture pattern recognition applications.