论文部分内容阅读
高光谱遥感图像为地物的精确分类带来了机遇,但也面临着一些挑战,高光谱遥感图像分类中所面临的一个挑战是如何处理高的光谱维数和小的样本数目之间的矛盾,目前几乎全部采用降维方法来缓解这一矛盾。集成学习的出现和选择性集成概念的提出为解决这一问题提供了新的研究思路,基于这一思想提出了基于波段分组和分类器集成的方法。在高光谱遥感图像的原始光谱空间根据波段之间的相似性信息对光谱波段进行分类,从每类中随机抽取一个波段形成新的光谱组,并依靠限制不同光谱组中相同波段的数目增加不同光谱组之间的差异程度,将新的光谱组作为训练分类器的特征子集,在特征子集训练最大似然分类器,使用简单的多数投票法合成得到最终的集成分类器。实验结果表明,使用基于波段分组和分类器集成的方法可以得到更高的分类精度。
Hyperspectral remote sensing images provide opportunities for the accurate classification of features, but also face some challenges. One of the challenges in the classification of hyperspectral remote sensing images is how to deal with the contradiction between high spectral dimension and small sample size At present, almost all adopt the method of dimensionality reduction to alleviate this contradiction. The emergence of integrated learning and the concept of selective integration provide a new research idea for solving this problem. Based on this idea, a method based on band grouping and classifier integration is proposed. In the original spectral space of hyperspectral remote sensing images, the spectral bands are classified according to the similarity information of the bands, a band is randomly selected from each class to form a new spectral group, and the number of the same bands in different spectral groups is increased by different limits The new spectral group is regarded as a subset of the training classifier, the maximum likelihood classifier is trained on the feature subset, and the final integrated classifier is synthesized by using the simple majority voting method. Experimental results show that using the method based on band grouping and classifier integration can get higher classification accuracy.