Pixel clustering is a basic process in image segmentation for classifying various regions in an image. However, this clustering becomes complex for multispectral and hyperspectral images due to the high dimensionality of these images. For such cases, dimensionality reduction is usually used to remove component images unsuitable for pixel clustering, in order to reduce computaional cost and improves accuracy. We propose a simple dimensionality reduction method in which a subset of component images is selected from multiple images using the joint singular value decomposition. Results of experiments for LandsatTM multispectral images demonstrate the effectiveness of the proposed method. Segmentation using a subimage chosen by the proposed enables us to avoid mixture of inappropriate component images and improve the performance of the segmentation.
