基于优化合并的高分辨率遥感影像分割算法

高分辨率遥感影像的分割算法研究对遥感数据处理与应用具有重要意义。本文提出了一种优化合并的分割算法以提高运算效率,该算法包含局部最优合并和全局最优合并2个阶段。第1阶段采用凝聚层次聚类（Hierarchical Agglomerative Clustering,HAC）方法实现局部最优合并,并对其合并规则进行了优化,使优化后的合并规则先注重光谱特征,再考虑待合并区域的几何特征。第2阶段采用区域邻接图（Region Adjacency Graph,RAG）方法实现全局最优合并,其合并规则主要考虑了区域的光谱和边界信息,减少了区域尺度对合并规则函数产生的负面影响,并且该阶段利用了红黑树来实现全局最优合并,以提高对RAG的搜索效率。最后,利用OrbView3高分辨率遥感影像开展了分割实验,结果表明本文算法可以得到令人满意的分割精度。本文的成果为遥感影像分割及其相关研究提供了新思路。

Segmentation Method Using Optimized Merging for High Resolution Remote Sensing Images

Study on the segmentation method for high resolution remote sensing images is very important for the processing and application of remote sensing data. Image segmentation plays an important role in geographic object-based image analysis, and it is also very useful in GIS data management and remote sensing data compression. A new segmentation algorithm using optimized merging criteria is proposed in this paper. The proposed algorithm divides the merging process into two stages, including the local best merging and the global best merging. Hierarchical agglomerative clustering is used to implement the first stage to meet the main objective of increasing the running efficiency. The merging criterion in the first stage focuses on the regional geometric information to create the visually pleasing segments, and in addition, this criterion is constructed on the premise that the regions to be merged should be sufficiently similar in spectra. Thus, when designing the merging criterion of the local best merge, the spectral and geometric information are both taken into consideration. Moreover, Global Moran′s I is used to determine the ending condition for the first stage. After the local best merging, the region adjacency graph (RAG) is constructed to implement the global best merging, in which the spectral and edge information is taken into account. In this stage, the negative impact introduced by the regions′ scale is found throughout the experiments. Thus, the size information of each region is excluded from the merging criterion of the global best merging. In addition, a special binary search tree, which is called the red-black tree, is used in the implementation to rank the edges of RAG, so as to speed up the graph structure updating after a merging taking place. High resolution images acquired from OrbView3 are adopted to conduct the segmentation experiment, the results of which indicate that our algorithm can produce the satisfactory performance. The conclusions made in this paper may provide new insights for the studies on remote sensing image segmentation and the related researches.