Optimization of the region-growing image segmentation for object-oriented land cover classification

Abstract

An object-oriented method has been recognized as superior to a pixel-based method in land cover classification using high resolution satellite images (e.g. IKONOS, QuickBird). Accurate object-oriented land cover classification is premised on precise image segmentation. Many region-growing methods are suggested and applied to improve the image segmentation. However, most of them do not have an objective criterion to decide the segmentation scale. This study inquires an objective and convenient method to optimize the region-growing image segmentation. The observation toward process of region-growing has showed the impracticality of using single common threshold over the whole scene. We suggest a method to extract the optimal regions by detecting the earliest stable stage of region-growing in every pixel. The evaluation (comparing the result with conventional region-growing method) has pointed out the effectiveness of the method for image segmentation as follow. This method can 1) restrain over-segmentation, 2) extract image objects with low coefficient of variation, 3) eliminate the subjectivity of the operator's work and 4) decrease the work load by its automatic algorithm.