基于改进超像素和标记分水岭的高分辨率遥感影像分割方法  

High-resolution remote sensing image segmentation based on improved superpixel and marker watershed

作　　者：张锐 尤淑撑 杜磊 禄競 何芸 胡勇 ZHANG Rui;YOU Shucheng;DU Lei;LU Jing;HE Yun;HU Yong(Land Satellite Remote Sensing Application Center, Ministry of Natural Resources, Beijing 100048, China;Chongqing Institute of Surveying and Monitoring for Planning and Natural Resources, Chongqing 400120, China)

机构地区：[1]自然资源部国土卫星遥感应用中心,北京100048 [2]重庆市规划和自然资源调查监测院,重庆400120

出　　处：《国土资源遥感》2021年第1期86-95,共10页Remote Sensing for Land & Resources

基　　金：国家自然科学基金项目“丘陵山区‘冬水田’时序定量遥感识别模型与方法研究”(编号:41901386);重庆市自然科学基金项目“冬水田定量遥感识别模型与方法研究”(编号:cstc2019jcyj-msxmX0548);中国博士后基金(编号:2020M670609)共同资助。

摘　　要：影像分割是高分辨率影像面向对象分析中的关键步骤,对信息提取精度起到至关重要的作用。为提高高分辨率遥感影像面向对象算法分割性能,提出一种改进超像素和标记分水岭的分割方法,包括特征融合、超像素初分割、控制标记符的标记分水岭再分割3个主要步骤。在超像素初分割阶段,利用高分辨率遥感影像纹理特征突出的优势,结合颜色空间、空间位置信息以及相位一致性纹理特征等信息提出一种新的距离测度计算规则,按照符合颗粒形状的圆形邻域进行搜索相似点,对影像进行超像素粗分割,并标记超像素斑块;计算超像素分割后每个斑块的灰度值,超像素分割后的影像重建,利用形态学的扩展技术提取局部极小值控制分割区域的数量,对传统数学形态学分水岭分割算法产生的过分割进行优化改进;对重建的影像进行高斯滤波,然后采用控制标记符分水岭算法对重建后的影像进行再分割,得到多尺度综合分割结果。在实验部分,利用资源三号卫星影像和机载航空影像验证本文提出方法,基于准确率和召回率定量评价分割精度,并将本文方法与其他分割方法的结果进行比较,证明本文提出方法的分割有效性。Image segmentation is a key step in object-oriented analysis of high resolution images and plays an important role in information extraction accuracy.In order to improve the segmentation performance of object-oriented algorithms for high-resolution remote sensing images,this paper proposes a segmentation method(PCSLIC-MW)to improve the superpixel and marker watershed,including feature fusion,superpixel initial segmentation,and control marker watershed segmentation.In the phase of superpixel segmentation,a new distance measure calculation rule is proposed,which combines color space,spatial position information and phase consistency texture feature.And then the gray value of each patch is calculated after superpixel segmentation,image reconstruction after segmentation,and morphological extension technology is used to extract local minimum(H-minima)so as to control the number of segmentation regions.The over-cutting produced by the traditional mathematical morphologic watershed segmentation algorithm is optimized and improved.The reconstructed image is conducted by Gaussian filter,and then the control marker watershed algorithm is used to re-segment the reconstructed image.For experiment,ZY3-02 satellite image and airborne aerial image are adopted to verify the proposed method,the precision and recall rate are used to evaluate the segmentation accuracy,and the results are compared with those of other segmentation methods to prove the segmentation effectiveness of the proposed method.

关 键 词：高分辨率影像 分水岭分割 超像素 资源三号 

分 类 号：TP79[自动化与计算机技术—检测技术与自动化装置]