卷积神经网络CNN（Convolutional Neural Networks）具有强大的特征提取能力，应用于高光谱图像特征提取取得了良好的效果，双通道CNN模型能够分别提取高光谱图像的光谱特征和空间特征，并实现了特征的决策级融合。局部二值模式LBP（Local Binary Patterns）是一种简单但有效的空间特征描述算子，能够减轻CNN特征提取的压力并提高分类精度。为了充分利用CNN的特征提取能力及LBP特征的判别能力，提出一种双通道CNN和LBP相结合的高光谱图像分类方法，首先，采用1维CNN（1D-CNN）模型处理原始高光谱数据提取深层光谱特征，同时采用另一个1D-CNN模型处理LBP特征数据进一步提取深层空间特征，然后，将两个CNN模型的全连接层进行连接，实现深层光谱特征和空间特征的融合，并将融合特征输入到分类层中完成分类。实验结果表明，该方法在Indian Pines数据、Pavia University数据及Salinas数据上能够分别取得98.54%、99.73%、99.56%的分类精度，甚至在有限数量的训练样本条件下也能取得较好的分类效果。

The classification of hyperspectral image remains a challenging task because of the complexity of spectral and spatial structures, high dimensionality, and strong correlation between adjacent bands. The combination of spatial and spectral information can provide significant advantage in terms of reducing the uncertainty of the samples because the same object has different spectrums and objects with the same spectrum in a hyperspectral image. The Local Binary Pattern (LBP) has also been introduced for spatial-domain feature extraction and classification of hyperspectral images as a simple but powerful texture descriptor. More recently, deep learning has been proven to be a preferable way to extract nonlinear high-level features because of its hierarchical learning framework. The combination of LBP features and the CNNs can lessen the workload of CNNs because of the discrimination capacity of LBP features. In this paper, a novel classification method combining DC–CNN and LBP features, called LBP Dual-Channel CNN (LBP–DC–CNN), is proposed.In LBP–DC–CNN, original hyperspectral data and LBP features are processed in a DC–CNN framework. On the one hand, original hyperspectral data is fed into a 1D–CNN model to extract original spectral features. On the other hand, LBP features are fed into an identical1D–CNN model to extract hierarchical spatial features further. Next, the fully connected layers of the two 1D–CNN models in the DC–CNN framework is concatenated into a fused layer, thus completing the fusion of spectral features and spatial features. Finally, the fused layer is fed into a softmax layer to conduct classification.(1) The OAs of LBP–DC–CNN are better than those of LBP–CNN and DC–CNN, which validate the feature extraction capacity of the CNNs and the advantage of LBP features. LBP–DC–CNN provides better accuracy than that of DC–CNN, which is an advantage of LBP features compared with the spatial features extracted by 2D–CNN model. In addition, the accuracy of LBP–DC–CNN is better than that of LBP–CNN, which validates the reasonability and discriminative power of the dual-channel CNN framework.(2) The OA of LBP–DC–CNN is apparently superior to those of compared methods, which makes DC–CNN and LBP features advantageous. For the Indian Pines data, LBP–DC–CNN (i.e., 98.54 %) yields approximately 2% higher accuracy than the DC–CNN (i.e., 96.68%)and approximately 4% higher accuracy than the LBP–CNN (i.e., 94.74 %). For the University of Pavia data, LBP–DC–CNN (i.e., 99.73 %) yields approximately 1 % higher accuracy than the DC–CNN (i.e., 98.74 %) and approximately 4 % higher accuracy than the LBP-CNN (i.e., 95.92 %). For the Salinas data, LBP–DC–CNN (i.e., 99.56 %) yields approximately 2 % higher accuracy than the DC–CNN (i.e., 97.33 %) and approximately 5 % higher accuracy than the LBP–CNN (i.e., 94.52 %).(3) LBP–DC–CNN can improve the class-specific accuracy of some ground materials, such as Corn-notill and Soybean-mintill in the Indian Pines data, Asphalt and Bricks in the University of Pavia data, and Grapes_untrained and Vinyard_untrained in the Salinas data. LBP features aremore discriminative than spatial features extracted by 2D–CNN.Result Experiments were conducted on the Indian Pines dataset, Pavia University dataset, and Salinas dataset to verify the performance of LBP–DC–CNN compared with conventional methods. The results are as follows: