基于双通道混合模型的小样本高光谱图像分类

Small Sample Hyperspectral Image Classification Based on Dual-Channel Mixed Model

摘要: 高光谱遥感图像中含有丰富的地表类别信息，为了在高光谱图像分类中更好提取和表达光谱与空间的精细特征以及特征间的交互信息，提出了一种双通道混合模型的图像分类方法。由于遥感图像的样本标定工作难度大、成本高，为了在小样本情况下提高光谱分类效果，运用一种样本的扩充方法；利用主成分分析(PCA)对高维光谱向量进行降维，以消除数据的冗余，为深度学习做好数据准备；为了提高深度学习网络的学习能力，又提出了一种双通道混合模型，该模型采用二阶连接方式(下一层与其前两层连接)，有利于特征信息的传递；再通过双通道的设计，有利于植入不同的卷积技术，提高网络的灵活性。最后，在三个广泛使用的高光谱数据集上进行数值实验，在小样本情况下，分类精度分别达到了96.55%、98.77%、99.31%。结果表明，文中提出的算法有效地提高了精度。

Abstract: Hyperspectral remote sensing image contains rich information of land surface classification, and the rapidly increasing development of machine learning makes it an effective method to obtain land surface classification. Due to the difficulty and high cost of sample calibration of remote sensing images, in order to improve the spectral classification effect in the case of small samples, a sample expansion method is proposed; Principal component analysis (PCA) is used to reduce the dimension of high-dimensional spectral vector to eliminate data redundancy and prepare data for deep learning. In order to improve the learning ability of the deep learning network, a second-order two-channel hybrid model is proposed, in which the features of the current layer and the previous layer determine the features of the next layer together, which is conducive to the transmission of feature information. The design of dual channel is also good for the implantation of different convolution techniques, which can improve the flexibility of the network. Finally, numerical experiments are carried out on three widely used hyperspectral datasets, in which the classification accuracy achieves 96.55%, 98.77% and 99.31% respectively in the case of small samples. The results show that the proposed algorithm can effectively improve the accuracy.