Few-Shot Remote Sensing Scene Classification with Multi-Metric Fusion

Abstract

Few-shot remote sensing scene classification is one of the research topics in the field of computer vision and few-shot learning, aiming to classify remote sensing scene through few training samples. The current methods of few-shot remote sensing scene classification use single metric, thus the classification accuracy is affected for the features cannot be effectively extracted. Therefore, we propose multi-metric fusion networks (MMFN) to address the problem via assembling a feature map multi encoder (FMME) and relation attention networks (RAN) to extract the features effectively and improve the classification accuracy. The FMME is designed to further encode the feature map which is extracted in the embedding phase to get different meaningful features. The RAN is aiming to calculate the similarity between features via fusing results of multiple methods based on image attention mechanism. Experimental results on three remote sensing data sets show that the multi-metric fusion method can extract meaningful features and effectively improve the classification performance of few-shot remote sensing scene.