融合光谱及形态学信息的对象级空间特征提取方法

Objected-Based Structural Feature Extraction Method Using Spectral and Morphological Information

针对传统的高分辨率遥感影像分割方法仅利用光谱特征或者形态学特征的弊端,提出了一种融合光谱信息和形态学信息的多尺度分割算法。该算法首先利用差分多尺度形态学序列特征与影像光谱特征构造光谱-形态学特征集,然后利用Hausdorff距离计算相邻像素的边权值并构造图模型,利用最小生成树Kruskal算法完成影像的初始分割,最后结合分形网络进化的区域异质性准则完成区域合并。在该分割结果的基础上,提出了面向对象的灰度共生矩阵特征和面向对象的像元形状指数特征。实验结果显示,所提出的分割方法在效果和效率上均优于eCognition 8.0和Meanshift算法,并且对象级灰度共生矩阵特征和对象级像元形状指数特征明显优于传统的像素级特征。

In order to overcome the drawbacks of using either spectral or morphological features for traditional image segmentation methods,a multi-scale image segmentation method using both the spectral and morphological information is proposed.First of all,Differential Morphological Profiles are combined with spectral features to form spectral-morphological characteristics.Then,Hausdorff distance is implemented to calculate the weight of edges based on graph theory and minimum spanning tree algorithm Kruskal is applied to complete the initial segmentation of color images.Finally,the obtained segmentation result is refined by a region merging procedure with the regional heterogeneous criteria proposed in fractal network evolution.Furthermore,object-based Gray Level Co-occurrence Matrix and object-based Pixel Shape Index are proposed on the basis of segmentation results.Experimental results show that the proposed segmentation method is more effective and more efficient than eCognition software 8.0 and Meanshift algorithm.In addition,object-based Gray Level Co-occurrence Matrix and object-based Pixel Shape Index are apparently better than traditional pixel-based methods.