A Hierarchical Prototypical Network for Few-Shot Remote Sensing Scene Classification.

Few-shot learning (FSL) aims at making predictions based on a limited number of labeled samples. It is a hot topic in many fields such as natural language processing, computer vision and more recently, remote sensing. In this work, we focus on few-shot remote sensing scene classification which aims to recognize unseen scene categories at training stage from few or even a single labeled sample at test stage. Although considerable progress has been achieved in this topic, less attention has been paid to leveraging the prior structural knowledge. In this paper, we learn transferable visual features by introducing the class hierarchy which encodes the semantic relationship between the classes. We build on a prototypical network and we define hierarchical prototypes that allow us to encode the different levels of the hierarchy. Experiments conducted on the remote sensing NWPU-RESISC45 dataset demonstrate that the proposed hierarchical prototypical network acts as a regularizer and leads to better performance than the original network in the context of few-shot remote sensing scene classification.