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

发布时间：2018-06-08 22:04

本文选题：高分辨率遥感 + 多特征融合　；参考：《地理与地理信息科学》2017年05期

【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%。
[Abstract]:In the framework of kernel function integrated SVM classification, a high resolution remote sensing image classification method which combines multi scale spectral spatial semantic features is proposed. First, based on the multiscale image segmentation set, the semantic features of the segmented graph spot are modeled by the didikray distribution model, and the spectral features and segmentation of the original image are combined. The spatial mean feature of the image spot is used to carry out the multi kernel function fusion classification of spectral space semantic features at different segmentation scales, and then integrate the multi-scale classification results at the decision level according to the majority voting principle, and realize the transformation of the pixel classification result to the object oriented classification result through the segmentation image under the minimum scale. The experimental results under the same scene and resolution data show that the classification method can realize the adaptive and smooth classification of the classification results, and to some extent improve the distinguishing ability of the "same spectrum" objects, such as buildings and roads, and the overall precision of the classification is raised from 66.7% and 63.7% based on spectral characteristic SVM to 86.8% and 87.2%..
【作者单位】：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;西南交通大学地球科学与环境工程学院;
【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)
【分类号】：TP751

【相似文献】