Knowledge Distillation Based Semi-supervised Hyperspectral Image Classification

Deep learning based models have achieved great progress in hyperspectral image classification. However, the lack of labeled data increases over-fitting risk and decreases the performance of the model. To tackle this issue, we propose a semi-supervised architecture, KDSemi, for better exploring the unlabeled data in hyperspectral images. Specifically, we employ a segmentation model to assign labels to every pixel. Unlike classification models, the segmentation model exports both labeled and unlabeled pixels. The unlabeled ones are used to develop an implicit reconstruction loss to learn in a knowledge distillation manner. We evaluate our model on three popular datasets. Experiments verify that our KDSemi achieves competitive results with SOTAs.