Enhancing few-shot remote sensing image semantic annotation through spectral super-resolution guided learning

Semantic annotation of remote sensing imagery aims to automatically assign a semantic category to each pixel based on a limited set of manually labeled samples. While supervised deep segmentation networks have achieved remarkable success in this task, their performance heavily relies on abundant pixel-level annotations, which are often prohibitively expensive to obtain. In practical scenarios, the scarcity of annotated samples leads to severe overfitting, significantly degrading generalization ability. To address this challenge, we propose a spectral super-resolution enhanced deep annotation framework that advocates enhancing the spectral representation of the input imagery as a complementary direction to network architecture design, with the goal of improving pixel discriminability in few-shot annotation scenarios. Specifically, motivated by recent advances in transforming RGB images into hyperspectral representations using deep neural networks, we design a heterogeneous multi-task learning framework that jointly trains an image annotation network and a spectral super-resolution network in an end-to-end fashion, with a shared spectral reconstruction module. Furthermore, a joint optimization strategy is integrated with a semi-supervised clustering algorithm to fully exploit the limited supervision signals from scarce labeled data. Extensive experiments on four benchmark datasets demonstrate that our method consistently outperforms state-of-the-art few-shot annotation approaches.