融合高程信息与随机森林模型的改进区域合并算法  

Improved region merging algorithm combining elevation information with random forest model

作　　者：王荣康 熊俊楠[1,2] 唐浩然 涂才森 宋南霄 WANG Rongkang;XIONG Junnan;TANG Haoran;TU Caisen;SONG Nanxiao(Southwest Pertoleum University School of Civil Engineering and Geomatics,Chengdu 610500,China;Xizang Autonomous Region Key Laboratory of Satellite Remote Sensing and Application,Xizang Autonomous Region,Lhasa 851400,China;Sichuan Institute of Metal Geologic Survey,Chengdu 611730,China;Southwest Pertoleum University School of Geoscience and Technology,Chengdu 610500,China)

机构地区：[1]西南石油大学土木工程与测绘学院,四川成都610500 [2]西藏自治区卫星遥感与应用重点实验室,西藏拉萨851400 [3]四川省金属地质调查研究所,四川成都611730 [4]西南石油大学地球科学与技术学院,四川成都610500

出　　处：《测绘通报》2025年第4期114-119,共6页Bulletin of Surveying and Mapping

基　　金：国家重点研发计划(2023YFC3006701);四川省科技厅重点研发项目(2024YFHZ0134)。

摘　　要：随着面向对象影像分析的广泛应用,图像分割在遥感图像处理中起着重要作用。目前很多影像分割算法都是基于区域合并方法,但此类方法普遍面临特征尺度局限、依赖单一光学影像特征及参数设置上的主观性问题,限制了分割效果。针对这一问题,本文提出了一种顾及高程特征合并策略的机器学习区域合并方法,使用基于随机森林(RF)的机器学习模型,同时辅助高程特征合并策略,通过计算邻接区域的特征矩阵作为输入特征变量,构建区域合并分类器,将区域合并问题转换成0和1的分类问题。试验结果表明,融合0.5 m空间分辨率的高程特征的区域合并算法,取得了较为优秀的分割结果,其F1、精确率、召回率、交并比分别达到90.5%、89.98%、91.02%、82.64%;与未使用高程特征相比,本文提出算法有效提升了分割精度,分别提升约3.4%、6.8%、1.1%、6.2%;同时高程特征重要性占比达32.5%,比光学特征重要性高约7%。With the wide application of object-oriented image analysis,image segmentation plays an important role in remote sensing image processing.At present,many image segmentation algorithms are based on region merging method,but these methods generally face the problem of limited feature scale,relying on a single optical image feature and subjective parameter setting,which limits the segmentation effect.To solve this problem,this paper proposes a machine learning region merging method that takes into account the elevation feature merging strategy.In this paper,the machine learning model based on random forest(RF)is used to assist the elevation feature merging strategy,and the region merging classifier is constructed by calculating the feature matrix of the adjacent region as the input feature variable.Transform the region merging problem into a classification problem of 0 and 1.Experimental results show that the region merging algorithm with 0.5 m spatial resolution elevation features achieves excellent segmentation results,with F1,accuracy,recall and crossover ratio reaching 90.5%,89.98%,91.02%and 82.64%,respectively.Compared with no elevation features,the proposed algorithm effectively improves segmentation accuracy.It is increased by about 3.4%,6.8%,1.1%and 6.2%respectively.Meanwhile,the importance of elevation features reached 32.5%,which is about 7%higher than that of optical features.

关 键 词：机器学习 高程特征 区域合并 图像分割 尺度变量 

分 类 号：P208[天文地球—地图制图学与地理信息工程]