摘要
高光谱影像中的标记样本往往有限,即使利用大量的训练样本,分类结果也会出现大量斑点状的误分点。利用JSEG分割的方法生成了同质区,以获取大量未标记样本的标签,并在光谱特征的基础上进行了多特征提取,提高了类别辨识度;利用最大投票原则,对分类图和分割图进行融合,将分割斑块内的类别众数作为该斑块的类别。实验证明,最大投票融合的方法减少了斑块状的误分点,大大提高了分类精度。
Marked samples in hyperspectral images are often limited,which poses a huge obstacle to their classification.Even with a large number of training samples,there will also be a lot of speckled misclassification points in the classification results.In this paper,we generated the homogenous region by using JSEG segmentation method to obtain a large number of unlabeled samples,and performed multi-feature extraction on the basis of spectral features to improve the class identification.And then,we used the principle of maximum voting to merge the classification map and the segmentation map,and took the category mode within the segmentation plaque as the category of the plaque.The experimental results show that the method of maximum voting fusion reduces the plaque-like misclassification points and greatly improves the classification accuracy.
出处
《地理空间信息》
2020年第5期20-25,I0005,共7页
Geospatial Information
基金
国家自然科学基金资助项目(41271420/D010702)。
关键词
JSEG分割
同质区
最大投票融合
半监督分类
JSEG segmentation
homogeneous region
maximum voting fusion
semi-supervised classification
