Using Learning by Discovery to Segment Remotely Sensed Images

In this paper, we describe our research in computer-aided image analysis. We have incorporated machine learning methodologies with traditional image processing to perform unsupervised image segmentation. First, we apply image processing techniques to extract from an image a set of training cases, which are histogram peaks described by their intensity ranges and spatial and textural attributes. Second, we use learning by discovery methodologies to cluster these cases. The first methodology we use is based on COBWEB/3, a conceptual clustering approach whose objective is to cluster the cases incrementally as the concept hierarchy is refined. The second methodology is based on an Aggregated Population Equalization (APE) strategy. This approach attempts to maintain similar strengths for all populations in its environment. The clustering result of either approach tells us the number of visually significant classes in the image (and what these classes are) and thus enables us to perform unsupervised segmentation, i.e., the labeling of all image pixels. Based on the results of the visual evaluation of the segmented images, we have built an unsupervised segmentation software tool called ASIS and have applied it to a range of remotely sensed images such as sea ice and vegetation index. In this paper, we present our machine learning approach to unsupervised image segmentation and discuss our experiments and their results.