电子学报 ›› 2012, Vol. 40 ›› Issue (4): 780-787.DOI: 10.3969/j.issn.0372-2112.2012.04.026

• 学术论文 • 上一篇    下一篇

基于SSMFA与kNNS算法的高光谱遥感影像分类

王立志, 黄鸿, 冯海亮   

  1. 重庆大学光电技术及系统教育部重点实验室,重庆 400044
  • 收稿日期:2011-08-23 修回日期:2011-11-21 出版日期:2012-04-25
    • 通讯作者:
    • 黄鸿
    • 基金资助:
    • 国家自然科学基金 (No.61101168); 重庆市科技攻关重点项目 (No.CSTC2009AB2231); 重庆市自然科学基金 (No.CSTC2009BB2195)

Hyperspectral Remote Sensing Image Classification Based on SSMFA and kNNS

WANG Li-zhi, HUANG Hong, FENG Hai-liang   

  1. Key Laboratory on Opto-Electronic Technique and Systems,Ministry of Education,Chongqing University,Chongqing 400044,China
  • Received:2011-08-23 Revised:2011-11-21 Online:2012-04-25 Published:2012-04-25

摘要: 为了研究高光谱影像数据的维数约简和分类问题,提出了一种基于半监督边际费希尔分析(SSMFA)和kNNS的高光谱遥感影像数据分类算法.该方法利用有标记数据和无标记数据的信息获得数据的内在流形结构,通过SSMFA将高光谱数据从高维观测空间投影到低维流形空间,然后利用邻域内多个近邻点的信息通过kNNS分类器对低维空间中的数据进行分类.在Urban、Washington和Indian Pine数据集上的分类识别实验表明,该方法能够较为有效地发现高维空间中数据的内蕴结构,在每类随机选取4,6,8个有类别标记的样本10个无类别标记的样本的情况下,该方法的总体分类精度能够比MFA+kNNS提高0.8%~2.5%,比MFA+kNN提高2.8%~4.5%,比其他算法提高4.0%~7.0%,分类精度有了明显的提高.

关键词: 高光谱影像, 地物分类, 图嵌入框架, 最近邻

Abstract: In order to explore dimensionality reduction and classification in hyperspectral remote sensing image,an algorithm based on semi-supervised marginal Fisher analysis(SSMFA) and k-nearest-neighbor simplex(kNNS) is proposed in this paper.First,the data are projected from a high-dimensional space onto low-dimensional space by SSMFA combined with the information of different classes.Then,classification is performed under the kNNS classifier by using a few neighbors from each class.The experimental results on the Urban data set,Washington DC Mall data set and Indian Pine data set show the effectiveness of the proposed algorithm,when i(i=4,6,8) labeled samples and 10 unlabeled samples of each class are randomly selected for training and 100 samples of each class for testing,the overall accuracy of our proposed algorithm is improved by 0.8%-2.5%,2.8%-4.5% and 4.0%-7.0%,respectively,as compared with MFA+kNNS,MFA+kNN and other methods.

Key words: hyperspectral images, land cover classification, graph embedding framework, nearest neighbor

中图分类号: