MSSPG-AL: Few-Shot Hyperspectral Image Classification with Active Learning Updated Multi-Scale Superpixel Graph Fusion

Deep learning has been widely used in the field of hyperspectral image (HSI) classification, but existing classification methods generally require a large number of labels. With rarely labeled samples, most deep learning methods have the problem of overfitting. Although few-shot learning has developed in this direction in recent years, many methods are weak in exploring the relationships between samples in the current scene. Graph-based methods have advantages in this respect, but there are some problems in graph learning, such as lack of label information guidance and multi-scale information loss. To solve these problems, in this paper, we propose a new multi-scale superpixel graph fusion method (MSSPG), and the graph structure is dynamically optimized by combining active learning (MSSPG-AL). By incorporating class relationships and multi-scale information, false connections in the graph structure can be iteratively eliminated. Experiments on three real hyperspectral data demonstrate that our method can achieve remarkable performance in the few-shot HSI classification.