ABSTRACT

The small sample problem has always been a serious challenge in hyperspectral image (HSI) classification. In order to obtain satisfactory results when the training samples are insufficient, the information around the training samples should be fully utilized. In this paper, we focus on small sample learning and propose a novel composite kernel learning network (CKLNet) for HSI classification. First, principal component analysis and extended morphological analysis are utilized to extract features. Then, we introduce generalized kernel method into deep learning technology. The spatial-spectral composite kernel learning (SSCKL) module is developed to construct discriminative and robust spatial-spectral generalized kernel features. In the process of constructing kernel features, the deep correlation information between samples is extracted simultaneously. The kernel hyperparameters in SSCKL are automatically learnt through backpropagation, thus avoiding the need to spend a lot of time on cross-validation. Finally, inspired by U-Net, a global-local feature extraction (GLFE) module is designed to extract spatial features of different scales. A set of classification probability maps can be obtained by the 1$\times$1 convolutional layer in the GLFE module. Experimental results on three widely used datasets demonstrate the effectiveness of the proposed CKLNet.

摘要

小样本问题一直是高光谱图像（HSI）分类中的一个严峻挑战。为了在训练样本不足时获得满意的结果，应充分利用训练样本周围的信息。在本文中，我们专注于小样本学习，并提出了一种用于 HSI 分类的新型复合核学习网络 (CKLNet)。首先，利用主成分分析和扩展形态分析来提取特征。然后，我们将广义核方法引入到深度学习技术中。开发了空间光谱复合核学习 (SSCKL) 模块以构建具有判别性和鲁棒性的空间光谱广义核特征。在构建内核特征的过程中，同时提取样本之间的深层相关信息。SSCKL 中的内核超参数是通过反向传播自动学习的，从而避免了在交叉验证上花费大量时间的需要。最后，受 U-Net 的启发，设计了一个全局-局部特征提取（GLFE）模块来提取不同尺度的空间特征。一组分类概率图可以通过 1$\times$GLFE 模块中的 1 个卷积层。在三个广泛使用的数据集上的实验结果证明了所提出的 CKLNet 的有效性。