An edge direction adaptive multi-scale watershed segmentation algorithm

Abstract

针对图像分割存在过分割、欠分割以及分割边界具有不确定性等问题,提出一种边缘走向自适应的多尺度分水岭遥感图像分割算法.此算法根据梯度变化的最大方向来确定单个波段的梯度值,通过各像元邻域内波段间的相关性合成多个波段的梯度,对梯度图像进行形态学重建之后,采用多尺度标记算法进行标记分水岭分割.选取QuickBird,SPOT,Landsat TM 3种不同空间分辨率的遥感影像对此算法进行试验分析;同时,将该算法与eCognition软件中多尺度分割方法、多波段组合的传统分水岭分割方法、形态学分水岭分割方法进行比较.结果表明:该算法分割结果的边界和真实的地物边界非常接近,分割结果精度优于eCognition软件中多尺度分割方法、多波段组合的传统分水岭分割方法和形态学分水岭分割方法. Aiming at the problems of over-segmentation, insufficient segmentation and uncertainties of segmentation boundary in image segmentation, an edge direction adaptive multi-scale watershed segmentation algorithm for remote sensing images was proposed. This method mainly consisted of three steps. First, the gradient of each band was separately computed along twelve directions, and the largest value was taken as the final gradient. Then, gradient of many bands were composed according to the relevance of bands in each pixel. The gradient image was processed using the morphological reconstruction method to reduce the noise effects. At last, the image was segmented by the watershed algorithm using a multi-scale labelling strategy. The performance of proposed method was validated using three remote sensing images with different spatial resolutions, which are QuickBird, SPOT, Landsat TM, and compared with the methods of eCognition software, the traditional multi-bands watershed segmentation method and the morphological watershed segmentation method. The experimental result is extremely close to the actual ground boundary, and demonstrates that the proposed method is superior to the other three methods in terms of the edge matching percentage and the computational efficiency.