Abstract: Deep learning is able to extract high-level features from input data via layer by layer abstraction. In recent years, it has been widely used in hyperspectral image classification. Most of the existing deep learning-based feature extraction methods for hyperspectral images belong to supervised learning models, which require a large number of labeled samples in the training process, but it is difficult and time-consuming to label hyperspectral images pixel by pixel. Therefore, we propose an unsupervised deep learning model based on inter-spectral contrast learning in this paper. It can extract features by modeling the relationship between different spectral bands without annotation of samples. Specifically, because different spectral channels of hyperspectral image depict the response degree of the same object in different electromagnetic spectrum, there must be a feature space, which makes the spectral information of different channels have similar characterization. Inspired by this, we first divide the high-dimensional spectral information into two groups, and then extract the features of each group using multi-layer convolution operations. Finally, the features extracted from different samples are compared and a contrastive loss function is constructed to optimize the model parameters. To test the performance of the proposed model, it was applied to a hyperspectral image classification task and validated on three commonly used data sets, including Houston 2013, Pavia University and WHU-Hi-Longkou. Experimental results show that using only 10 training samples in each class, the proposed unsupervised learning model can obtain better classification performance than the commonly used unsupervised models such as principal component analysis and auto-encoder.

Key words: unsupervised learning, deep learning, hyperspectral image, feature extraction