摘要: |
提出了一种融合监督分类与非监督分类结果的高光谱遥感影像分类新方法——众数赋值分类法。采用ISODATA 非监督分类方法对高光谱遥感影像进行分类, 并对非监督分类结果的图斑进行标记,同时用最大似然法(ML)和支持向量机(SVM)法进行监督分类, 然后以监督分类结果对非监督分类后各斑块进行类别赋值。方法是: 统计每个非监督分类斑块中由监督分类所获得的各类别像元数及所占比例, 将非监督分类斑块的类别赋予所占比例最高的监督分类结果的类别, 最终获得高光谱图像分类结果。研究表明: (1)非监督分类类别数量大于10 时, 其与ML 分类结果融合的总体分类精度和Kappa 系数均较监督分类法的分类结果好; (2)ML 和20 个类别的ISODATA 分类结果融合的总体精度最高, 为87.35%, 比单独ML 的总体精度高约2 个百分点; (3)SVM 和10 个类别的ISODATA 分类结果融合的总体精度提高最大, 较SVM 的总体精度提高近3 个百分点; (4)随着非监督分类类别数量的增多, 分类结果的总体精度呈现由低到高再到低的变化过程。 |
关键词: 高光谱图像 黄河口湿地 监督分类 非监督分类 众数赋值 |
DOI:10.11759/hykx20141011009 |
分类号: |
基金项目:国家海洋局国家海域管理技术重点实验室开放基金项目(201205); 国家海洋公益性行业科研专项资助项目(201005011-3) |
|
A hyperspectral image classification method based on maximum assignment |
|
Abstract: |
In this paper, a new image classification method of hyperspectral remote sensing, named maximum assignment classification was proposedusing the advantages of unsupervised classification to improve the classification results of supervised classification. Firstly, the unsupervised classification (ISODATA) was used to classify the hyperspectral remote sensing images, and the map spot was marked on the classification result. Then the maximum likelihood algorithm (ML) or support vector machine (SVM) algorithm was used to supervise classification. Finally the supervised classification results were used for the unsupervised classification map spot type assignment: by counting the pixels number and proportion of various categories in each map spot, the map spot was assigned by the highest proportion of categories, eventually getting the hyperspectral image classification map. It was found that(1) when the unsupervised classification number was higher than 10, the overall accuracy and Kappa coefficient of the fusion result with ML were better; (2) when the ML and ISODATA with 20 kinds of classifications were combined, the overall accuracy reached the maximum value of 87.35%, which was about 2%higher than the result obtained by ML alone; (3) when SVM and 10 kinds of classifications were combined, the overall accuracy increased to the maximum value, compared with the overall accuracy of SVM increased by nearly 3%; (4) with the increasing number of unsupervised classification, the overall accuracy of classification results changed from low to high and then to low. |
Key words: hyperspectral remote sensing image wetland of the Yellow River Delta supervised classification unsupervised classification maximum assignment |