Deep Self-Supervised Pixel-Level Learning for Hyperspectral Classification

The task of hyperspectral classification plays an important role in decision-making processes. It is mainly solved applying supervised deep learning techniques that require numerous well-annotated pixels. Recently developed methods based on contrastive learning attempt to alleviate this dependency on abundant labels. To learn representations, these techniques distinguish between views of an image patch, which could be unfavorable for downstream tasks requiring dense pixel predictions. To overcome this challenge, the current work exploits contrastive learning at the pixel-level to learn robust feature representations for classifying hyperspectral imagery with fewer labels. The method starts with the Self-Supervised Learning (SSL) phase that consists of sampling two augmented views of the same image patch to compute feature maps. These are subsequently used to carry out the pixel discrimination task for feature learning. Afterwards, the learned representations are transferred to solve the classification using different amounts of labels. Experiments with only one-tenth of the labels show that the proposed method achieves better classification results than well-established methods, including those trained on Image-Net.