Hyperbolic prototypical network for few shot remote sensing scene classification

Recently, in the computer vision and machine learning (ML) communities, a growing interest has been directed to similarity measures operating in hyperbolic spaces due to the geometric properties of these spaces which make them very suitable for embedding data with an underlying hierarchy. These hyperbolic spaces, although increasingly adopted, have received limited attention in the remote sensing (RS) community despite the hierarchical nature of RS data. The objective of this study is therefore to examine the relevance of hyperbolic embeddings of RS data, in particular when addressing the few -shot remote sensing scene classification problem. We adopt hyperbolic prototypical networks as a meta -learning approach to embed scene images along with a feature clipping technique to ensure a more numerically steady model. We then examine whether hyperbolic embeddings provide a better representation than Euclidean representations and better reflect the underlying structure of scene classes. Experimental results on the NWPU-RESISC45 RS dataset demonstrate the superiority of hyperbolic embeddings over their Euclidean counterparts. Our study provides a new perspective by suggesting that operating in hyperbolic spaces is an interesting alternative for the RS community.