A novel meta-learning-based hyperspectral image classification algorithm

Aimed at the hyperspectral image (HSI) classification under condition of limited samples, this paper designs a joint spectral–spatial network based on metric meta-learning. First, in order to fully extract HSI fine features, squeeze and excitation (SE) attention mechanism is introduced into spectrum dimensional channel selectively useful features improve sensitivity information features. Second, part spatial feature extraction, VGG16 model parameters trained advance HSRS-SC dataset are used realize transfer learning knowledge, then, higher-level abstract extracted mine intrinsic attributes ground objects. Finally, gated fusion strategy connect spectral for mining more abundant information. In paper, large number experiments carried out public dataset, including Pavia University Salinas. The results show that meta-learning method can achieve fast new categories with only small labeled samples has good generalization ability different datasets.