Multi-Scale Interaction Prototypical Network For Few-Shot Remote Sensing Scene Classification

Shiji Pei, Yijing Wang, Jingjing Ma, Xu Tang, Yuqun Yang

Abstract: Few-shot remote sensing scene classification (FSRSSC) aims to make the model quickly adapt to new scenes with a small amount of annotation data. The large intra-class variance and high inter-class similarity in remote sensing (RS) scenes make this task more challenging. To this end, we propose a multi-scale interaction prototypical network, which pays attention to capturing multi-scale information of images during model learning, and then generates a prototype representation of mixed query information through a feature interaction module, thereby enhancing the rapid learning ability of the model, so as to reducing of the within-class and between-class variance ratio in RS scenes. The positive experimental results on UC-Merced and NWPU datasets demonstrate the effectiveness of our model in FSRSSC.