光学 精密工程, 2018, 26 (4): 980, 网络出版: 2018-08-28
多核融合多尺度特征的高光谱影像地物分类
Fusion of multi-scale feature using multiple kernel learning for hyperspectral image land cover classification
高光谱影像 稀疏表示 多尺度 多核学习 地物分类 hyperspectral image SRC multi-scale multiple kernel learning land cover classification
摘要
对于高光谱影像地物分类问题, 为更加有效地利用像元空间信息和光谱信息, 提高地物分类精度, 提出了多核融合多尺度特征的分类方法。首先, 通过多尺度空间滤波和PCA白化, 提取出多尺度特征;接着在核稀疏表示分类器内使用多核方式对分别表示每项特征, 在分类器内实现特征自动融合, 根据子核与理想核、子核之间距离求取核组合的权重, 使用训练集所构成的字典在特征空间内对待测样本进行线性表示, 根据每类地物的重构误差确定待测像元所属地物类别。实验结果表明: 对于Indian Pines影像和Pavia University影像总体分类精度分别达到99.51%和97.96%, 较传统方法明显提高, 并且对于小样本地物识别精度也都能达到90%以上。本文算法对于高光谱影像地物具有更强的识别能力, 并且具有较强的稳定性和鲁棒性。
Abstract
In order to make full use of the spatial information and spectral information and improve the classification accuracy of hyperspectral imager, a fusion multi-scale feature with multiple kernel learning method was proposed in this paper. Firstly, multi-scale features were extracted by multi-scale spatial filtering and PCA whitening. Then multiple kernels were used to represent the multi-scale feature in the framework of kernel sparse representation classifier. The kernel weight was computed according to the CKTA between the sub-kernels and ideal kernel and the CKTA between sub-kernels. The unlabeled pixels were linearly represented by the training samples in the feature space. According to the reconstruction error of each kind of land cover, the category of unlabeled pixels was determined. The experiment results showed that the overall classification accuracy in Indian Pines images and Pavia University images reached 99.51% and 97.96%, which significantly surpassed the traditional method. The accuracy of object recognition of small sample could also reach more than 90%. It can be seen that algorithm proposed has stronger recognition ability for hyperspectral images land cover.
王庆超, 付光远, 汪洪桥, 王超. 多核融合多尺度特征的高光谱影像地物分类[J]. 光学 精密工程, 2018, 26(4): 980. WANG Qing-chao, FU Guang-yuan, WANG Hong-qiao, WANG Chao. Fusion of multi-scale feature using multiple kernel learning for hyperspectral image land cover classification[J]. Optics and Precision Engineering, 2018, 26(4): 980.