Extended Subspace Projection Upon Sample Augmentation Based on Global Spatial and Local Spectral Similarity for Hyperspectral Imagery Classification
Extended Subspace Projection Upon Sample Augmentation Based on Global Spatial and Local Spectral Similarity for Hyperspectral Imagery Classification
复制标题
DOI：
10.1109/jstars.2021.3107105
复制
发表时间：
2021
影响因子：
5.5
通讯作者：
Jiaochan Hu;Xueji Shen;Haoyang Yu;Xiaodi Shang;Qiandong Guo;Bing Zhang
中科院分区：
文献类型：
--
作者：
Jiaochan Hu;Xueji Shen;Haoyang Yu;Xiaodi Shang;Qiandong Guo;Bing Zhang
波段冗余和标记样本的有限性严重制约了高光谱图像分类的发展。为了解决早期的问题，基于子空间的支持向量机等分类模型取得了一定的进展，但主要集中在降维方面，忽略了训练样本的扩充。事实上，这两个问题对于提高分类的性能同样重要，应该同时解决。因此，本文提出了一种基于全局空间和局部谱相似性的样本增强扩展子空间投影方法（GLSC），该方法兼顾了样本增强和降维两个方面。具体地说，它首先利用GLSC来扩大原始标记样本集，这使得HSIC能够获得更多的先验信息。然后将增强后的样本与原标记样本相结合，构造出更全面反映地物真实的情况的扩展子空间。最后，将原始HSI投影到子空间，并由邻域活跃度驱动的基于表示的分类器进行分类。在3个真实的高光谱数据集上的实验结果表明了该方法在HSIC任务中的实用性和有效性。
Band redundancy and limitation of labeled samples restrict the development of hyperspectral image classification (HSIC) greatly. To address the earlier issues, the classification models such as subspace-based support vector machines, which have gained a certain advance but mainly concentrate on the dimensionality reduction and ignore the augmentation of training samples. In fact, these two issues are equally important for improving the performance of classification, and should be addressed simultaneously. Therefore, this article proposes a novel method named extended subspace projection upon sample augmentation based on global spatial and local spectral similarity (GLSC) for HSIC, which takes both sample augmentation and dimensionality reduction into consideration. Specifically, it first exploits the GLSC to enlarge the original labeled sample set, which allows HSIC to obtain more prior information. Then, the augmented samples and the original labeled samples are combined to construct the extended subspace, which is more comprehensive to reflect the real situation of the ground objects. Finally, the original HSI is projected to the subspace and classified by the neighborhood activity degree-driven representation-based classifier. Experimental results on three real hyperspectral datasets demonstrate the practicality and effectiveness of the proposed method for HSIC tasks.