Synthesis Of Multi-Scale Segmentation Results Based On Land Cover Categories

In remote sensing imagery, ground objects belonging to the same land cover category always have similar optimal segmentation scales. The paper proposed a method using the land cover categories as a prior knowledge to guide the synthesis of multi-scale image segmentation results. This method took into account the variety of scale characteristics of different ground objects as well as the similarity of scale of objects belonging to the same land cover category. Firstly, the image was coarsely divided into multiple regions, and each of them belonged to a land cover category. Then for each category, we selected the optimal segmentation scale by the supervised accuracy assessment of segmentation results. Finally, the optimal scales of segmentation results were synthesized to get the final segmentation result. To validate this method, the Quickbird image was segmented and classified. Experimental results showed that this method could generate accurate segmentation results for the latter classification.