面向高分辨率遥感影像分类的多尺度光谱-空间-语义特征融合  

Fusion of Multiscale Spectral-Spatial-Semantic Features for Classification of High Resolution Remote Sensing Images

作　　者：慎利[1,2] 吴林梅[1,2] 方灿明 

机构地区：[1]西南交通大学高速铁路运营安全空间信息技术国家地方联合工程实验室,四川成都611756 [2]西南交通大学地球科学与环境工程学院,四川成都611756

出　　处：《地理与地理信息科学》2017年第5期1-8,127,128,共10页Geography and Geo-Information Science

基　　金：国家重点研发计划项目(2016YFB0501403);国家自然科学基金项目(41401374);国家重点基础研究发展计划(2012CB719901);中央高校基本科研业务费专项(2682016CX079);国土资源部航空地球物理与遥感地质重点实验室开放基金项目(2016YLF10)

摘　　要：在核函数集成SVM分类框架下,提出一种融合多尺度光谱-空间-语义特征的高分辨率遥感影像分类方法。首先,以多尺度影像分割集为基础,利用潜狄利克雷分配模型对分割图斑的语义特征进行建模,并结合原始影像的光谱特征以及分割图斑内的空间均值特征,在不同分割尺度下分别开展光谱-空间-语义特征的多核函数融合分类;然后根据多数投票法原则在决策级集成多尺度分类结果,通过最小尺度下的分割影像实现像素级分类结果至面向对象分类结果的转化。不同场景和分辨率数据下开展的实验结果表明,该分类方法能够实现分类结果的自适应平滑分类,并在一定程度上提高建筑物和道路等"同谱异物"地物的区分能力,分类总体精度由基于光谱特征SVM的66.7%和63.7%提升至86.8%和87.2%。A novel approach for the classification of high resolution remote sensing images is presented by fusing multiscale spectral- spatial- semantic features via the multiple- kernel SVM classifier. Based on a series of image segmentation maps produced by the Entropy Rate Superpixel Segmentation（ ERSS） algorithm, the proposed method utilizes the latent Diridilet allocation to model the semantic feature of each segment first ly. Then, the spectral feature represented by t he original pixef s value and spa-tial feature represented by the average pixel value of each segment are combined with the learned semantic feature to conduct the spectral- spatia- semantic classification in the framework of the composite kernel SVM at each scale in scale space,followed by a majority decision fusion to integrate multiple classification results. Finally, the image segmentation map at the minimum scale is used to convert the ensemble pixel based classification results to the object based classification results. Experimental results u-sing high resolution satellite imageries with different scenes and spatial resolutions indicate that the proposed approach can a2 chieve a self- adaptive smoothing effect on the classification map, and increase the ability to distinguish the geo- objec：ts with spec-tral overlap, such as road and building.T he overall accuracy was improved from 66. 7% and 63. 7% with the conventional SVM based on the spectral feature to 86. 8% and 87.2% with the proposed method.

关 键 词：高分辨率遥感 多特征融合 核函数集成 主题模型 多尺度影像分割 

分 类 号：P237[天文地球—摄影测量与遥感]