An efficient hyperspectral image classification method for limited training data

Hyperspectral image classification has gained great progress in recent years based on deep learning model and massive training data. However, it is expensive unpractical to label hyperspectral data implement constrained environment. To address this problem, paper proposes an effective ghost module spectral network for classification. First, Ghost3D adopted reduce the size of parameter dramatically by redundant feature maps generation with linear transformation. Then Ghost2D channel-wise attention used explore informative representation. For large field covering, non-local operation utilized promote self-attention. Compared state-of-the-art methods, proposed approach achieves superior performance three sets fewer sample labelling less resource consumption.