基于形态学属性剖面的高光谱影像集成分类

传统高光谱遥感影像逐像素分类方法未考虑像元之间的空间关联性且泛化性能较低。形态学属性剖面是表征影像空间结构的有效方法,同时集成学习可显著提升分类算法的泛化能力。为了在高光谱影像分类中充分利用影像的空间信息并提高分类的稳定性,提出一种基于形态学属性剖面高光谱遥感影像集成学习分类方法。首先,用主成分分析和最小噪声变换进行特征提取,并借助形态学属性剖面获取影像的多重空间特征;然后用极限学习和支持向量机的方法进行分类;最后将多个分类结果以多数投票的方式集成。区别于已有集成学习方法,综合考虑了不同特征提取和不同分类方法的联合集成,并将形态学属性剖面引入其中以充分利用影像的空间信息。采用AVIRIS和ROSIS两组高光谱数据检验该方法的分类性能,实验结果表明该方法可获得高精度和高稳定性的分类结果,总体精度分别达到83.41%和95.14%。

Ensemble Classification for Hyperspectral Imagery based on Morphological Attribute Profiles

The traditional pixel\|wised classification methods is lack of hyperspectral image (HIS) ignore spatial information and generalization ability,resulting in a big limit of classification performance.Morphological attribute profiles is an effective method to express spatial information of remote sensing image.In the meanwhile,ensemble learning machines possess high ability of power generalization so that to improve classification stability.An ensemble method based on morphological attribute profiles is proposed for hyperspectral image in order to make full advantage of spatial information to improve stability.Firstly,principal component analysis and minimum noise fraction are utilized for feature extraction,and then morphological attribute profiles operations are carried out on the first a few features which have most information of the image;Secondly,supporting vector machines and extreme learning machines are used as base classifiers for their good classification performance.Finally,the results of each base classifier are combined by 〖JP3〗the way of majority voting.Compared with other ensemble method,different feature extraction algorithm and different base classifiers are both combined to form the joint integrated model.In addition,spatial information deriving from morphological attribute profiles is introduced to improve classification accuracy. Experiment on AVIRIS data set and ROSIS data set respectively illustrate that the method can obtain better classification performance in terms of precision and stability,and the overall accuracy of them reach to 83.41% and 95.14%,respectively.