基于分水岭变换及分层区域合并的城市高分辨率影像分割

Segmentation of High Resolution Imagery over Urban Area Using Watershed Transformation and Stratified Region Merging

针对城市区域高分辨率图像的特点,以及传统的基于分水岭变换的图像分割方法中存在的过分割问题,提出一种分割区域的分层合并方法来改进分割结果。首先采用多通道分水岭分割得到初始分割结果,然后通过定量分析城市不同地物内部光谱变化性的特点,对影像进行分层,并对不同的层分别进行合并,得到最终的分割结果。采用一景北京地区的QuickBird影像,从目视评价、定量指标计算以及应用等3个方面,对提出的方法进行验证和评价,并与现有的分割方法比较。结果表明,与现有方法相比,基于分层区域合并的方法可得到更准确的分割结果,适合城市高分辨率图像的分割。

Given that watershed transform based image segmentation methods usually produce obvious over-segmentation, and the high resolution imagery of urban areas shows some peculiar characteristics, the authors propose a hierarchical region merging method to optimize the segmentation results initially produced by watershed transformation. A multi-channel watershed transformation was adopted to generate an initial segmentation result. Through quantitative analysis of internal spectral variability of different ground objects in urban areas, the image was stratified to several layers. Region merging was separately conducted in each layer. A final segmentation result was obtained by aggregating segmentation results from these layers. The proposed method was evaluated by comparing with the existing watershed segmentation method in terms of visual inspection, quantitative measures and applications in urban land cover classification, using a QuickBird image of Beijing area. The experimental results indicate that the proposed method outperforms the existing method and it is suitable for segmentation of high resolution imagery over urban areas.