Self-Supervised Learning with Adaptive Distillation for Hyperspectral Image Classification
Jun Yue, Leyuan Fang, Hossein Rahmani, and Pedram Ghamisi
Hyperspectral sensors collect hundreds of spectral bands and are widely used in many fields including remote sensing, military, and healthcare. Hyperspectral images (HSI) combine spatial texture information and spectral reflectance information, and allow detecting unique features of individual objects. To assign an object category to each pixel of the image, HSI classification is performed. The accuracy of HSI classification can be significantly improved by deep learning methods with spatial-spectral feature extraction. However, training deep neural networks requires a large number of labelled samples, which are scarce in HSI.
Here, we introduce a self-supervised learning (SSL) method for HSI classification with a small number of labelled samples. SSL methods construct supervised information from unsupervised data and use it to train the model. Our method comprises two modules: adaptive knowledge distillation and data transformation in spatial and spectral domains. In the first module, adaptive soft labels are generated for unlabeled samples based on their spatial-spectral joint distance to labelled samples. To fully exploit the labelled samples, the second module rotates HSI hypercube in the spatial and spectral domains. We show that both modules are effective for improving the accuracy of the method, and their combination leads to the best results. Experiments performed on three public datasets demonstrate that the proposed method outperforms the existing state-of-the art HSI classification methods.
The code is available on GitHub.
IEEE Transactions on Geoscience and Remote Sensing, 1-13, 2021-02-22.