Abstract: In view of the problem of dimension “disaster”, insufficient use of spatial information and features in hyperspectral images, combined with the recent advances in deep learning, manifold learning and multiscale spatial features, a TSNE and multiscale sparse auto-encoder neural network hyperspectral image classification algorithm is proposed. TSNE algorithm is used to reduce the hyperspectral image, and then multiscale spatial information extraction is carried out for each pixel’s neighborhood. A sparse auto-encoder network is trained by using the spatial spectrum joint information and classified by softmax classifier, which reduces computational complexity and improves classification accuracy. Experiments on the data of Indian Pines and Pavia University show that the proposed algorithm has better classification results than the other five algorithms.

Key words: hyperspectral image, deep learning, multiscale spatial feature, manifold learning

摘要： 针对高光谱图像存在维数“灾难”、特征以及空间信息利用不足的问题，结合深度学习、流形学习及多尺度空间特征的最新进展，提出了一种TSNE和多尺度稀疏自编码网络的高光谱图像分类算法。利用TSNE算法对高光谱图像进行降维，再对每个像元的邻域进行多尺度空间特征提取，利用加入空谱联合信息的像元训练稀疏自编码网络模型并通过softmax分类器进行分类，减少计算复杂度，提高分类精确度。通过对Indian Pines及Pavia University两组数据进行实验，结果表明，提出的算法与其他五种算法相比分类效果更好。

关键词: 高光谱图像, 深度学习, 多尺度空间特征, 流形学习