摘要： 高光谱图像具有较高的空间分辨率，蕴含着丰富的空间光谱信息，近年来被广泛用于城市地物分类中。在高光谱图像分类过程中，空间光谱特征的提取直接影响着分类精度；传统的高光谱图像特征提取方法只利用了4或8邻域的像素进行简单卷积处理，因而丢失了大量的复杂、有效信息；卷积神经网络（CNN）虽然可以自动提取空间光谱特征，在保留图像空间信息的同时，简化网络模型，但是，随着网络深度增加，网络分类产生退化现象，而且网络间缺乏相关信息的互补性，从而影响分类精度。该工作引入CNN自动提取空间光谱特征，并且针对CNN深度增加所导致的退化问题，设计了面向地物分类的高光谱特征融合残差网络。首先，为了降低高光谱图像的光谱冗余度，利用PCA提取主要光谱波段；然后，为了逐级提取光谱图像的空间光谱特征，定义了卷积核为16，32，64的低、中、高3层残差网络模块，并利用64个1×1的卷积核对3层特征输出进行卷积，完成维度匹配与特征图融合；接着，对融合后的特征图进行全局平均池化(GAP)生成用于分类的特征向量；最后，引入具有可调节机制的Large-Margin Softmax损失函数，监督模型完成训练过程，实现高光谱图像分类。实验采用Indian Pines，University of Pavia和Salinas地区的高光谱图像来验证方法有效性，设置批次训练的样本集为100，网络训练的初始学习率为0.1，当损失函数稳定后学习率降低为0.001，动量为0.9，权重延迟为0.000 1，最大训练迭代次数为2×104，当3个数据集的样本块像素分别设置为25×25，23×23，27×27，网络深度分别为28，32和28时，3个数据集的分类准确率最高，其平均总体准确率（OA）为98.75%、平均准确率（AA）的评价值为98.1%，平均Kappa系数为0.98。实验结果表明，基于残差网络的分类方法能够自动学习更丰富的空间光谱特征，残差网络层数的增加和不同网络层融合可以提高高光谱分类精度；Large-Margin Softmax实现了类内紧凑和类间分离，可以进一步提高高光谱图像分类精度。
关键词：高光谱图像分类；深度学习；残差网络；Large-Margin Softmax
Abstract：Hyperspectral images contain a wealth of feature information, and they have been widely used in urban feature classification in recent years. In the process of hyperspectral image classification, the extraction of spatial spectral features directly affects the classification accuracy. Traditional hyperspectral image feature extraction methods only use 4 or 8 neighborhood pixels for simple convolution processing, thus losing a lot of complex and effective information. Convolution neural network (CNN) can automatically extract spatial spectral features and retain the same spatial information of the image, and the network model is simplified. However, with the increase of network depth, the network classification will degenerate, and the network lacks complementarity of relevant information, which will affect the classification accuracy. In this paper, a hyperspectral residual network for feature classification is designed for the degradation problem. Firstly, define the residual network module of the low, medium and high three-layer structure with convolution kernels of 16, 32, and 64. Then, convolve the 3-layer output features with 64 1×1 convolution kernels to complete the dimension matching and feature map. Next, the global average pooling (GAP) of the feature map is generated to generate the feature vector for classification. Finally, the Large-Margin Softmax objective function is introduced to achieve hyperspectral image classification. The experiments were performed using hyperspectral images from the Indian Pines, University of Pavia, and Salinas regions. The primary bands of the hyperspectral image were extracted by PCA. With the sample set of batch training being 100, the initial learning rate being 0.1, the momentum being 0.9, the weight delay being 0.000 1, and the maximum number of training iterations being 2×104, when the sample sizes of the three data sets are set to be 25×25，23×23 and 27×27, the network depth is 28，32 and 28, the classification accuracy of the three data sets is the highest, and the average overall accuracy OA is 98.75%, the average accuracy AA is 98.1% and the average Kappa coefficient is 0.98. The experimental results show that the classification method based on residual network can get more affective features. It can improve the classification accuracy with the increase of the number of residual network layers and the fusion of complementary information of different network layer outputs; Large-Margin Softmax achieves intra-class compactness. Separation between classes further improves classification accuracy.
Key words：Hyperspectral image classification; Deep learning; Residual network; Large-Margin Softmax
[1] Fauvel M, Tarabalka Y, Benediktsson J A, et al. Proceedings of the IEEE, 2013, 101(3): 652.
[2] Chen Y, Nasrabadi N M, Tran T D. IEEE Transactions on Geoscience and Remote Sensing, 2011, 49(10): 3973.
[3] Zhang H, Li J, Huang Y, et al. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2014, 7(6): 2056.
[4] Krizhevsky A, Sutskever I, Hinton G. ImageNet Classification with Deep Convolutional Neural Networks. NIPS 2012: Neuram Information Processing Systems, 2012.
[5] Chen Yushi, Jiang Hanlu, Li Chunyang, et al. IEEE Trans Geoscience and Remote Sensing, 2016, 54(10): 6232.
[6] Hu W, Huang Y, Wei L, et al. Journal of Sensors, 2015(2): 1.
[7] Zhao W Z and Du S H. IEEE Transactions on Geoscience and Remote Sensing, 2016, 54 (8): 4544.
[8] Zhong Z, Li J, Luo Z, et al. IEEE Transactions on Geoscience and Remote Sensing, 2018, 56(2): 847.
[9] Liu W, Wen Y, Yu Z, et al. Large-Margin Softmax Loss for Convolutional Neural Networks. Proceedings of 33rd International Conference on Machine Learning, 2016: 507.
[10] Chai S, Liu H, Gu Y, et al. IEEE Transactions on Geoscience and Remote Sensing, 2017, 55(8): 4775.
[11] Zhang X, Zhang H, Zhang Y, et al. IEEE Transactions on Image Processing, 2016, 25(3): 1033.
[12] Song Weiwei，Li Shutao，Fang Leyuan，et al. IEEE Transactions on Geoscience and Remote Sensing, 2018, 56(6): 3173.
