The object-oriented segmentation is a critical process in the classification and recognition of high-resolution remote sensing images. Multi-threshold segmentation methods have been widely used in multi-target recognition and information extraction of high-resolution remote sensing images because they are simple, easy-to-implement, and has ideal segmentation effect. However, the determination of thresholds for existing multi-threshold segmentation algorithms is still a problem, which limits to get the best effect of segmentation. To address this issue we propose a self-adapted multi-threshold segmentation method, based on region merging, toward segmenting remote sensing images. This method involves four steps: image preprocessing based on morphological filtering, improved watershed transformation to initiate primitive segments, optimal region merging, and self-adapted multi-threshold segmentation. The performance of the proposed algorithm is evaluated in QuickBird images and compared to the existing region merging method. The results reveal the proposed segmentation method outperforms the existing method, as indicated by its lower discrepancy measure.