基于混合光谱特征的建筑物高分影像分类样本筛选

摘要

The screening programs of random samples and interested samples are designed.Sample tiles of typi-cal buildings are used to obtain the minimum sample data of the main land types and the influence factors of high precision by K-means unsupervised classification.We obtained spectral variations of buildings that were misclassified and omitted by binary clustering,and analyzed clustering segmentation rules of the high spatial res-olution remote sensing image from the spectral statistical view.Inherent relationship is found between the surface spectral characteristics and the spectral distribution clustering results.An effective method is proposed to solve the screening of the building’s mixed spectral region of interest.The study can help the understanding of de-pendence of high-resolution image classifier’s performance on the spectral distribution of sample tiles prior scene preferably,and improveing the classification accuracy of the buildings.%设计了随机样本、感兴趣样本筛选方案，选择典型建筑物样本瓦片通过K-means非监督分类获得核心地类的最小样本信息及提高精度的影响因素。利用二值聚类获取建筑物的错分漏分光谱变化，从光谱统计角度分析高空间分辨率影像的聚类分割规律，发现地表真实像元的光谱特征与聚类结果类别光谱分布的内在联系，提出了解决建筑物混合光谱感兴趣区筛选的有效方法。研究表明，该方法可以更好的了解高分影像分类器性能对样本瓦片先验场景光谱分布的依赖程度，进而提高建筑物的分类精度。