RS-MetaNet: Deep Metametric Learning for Few-Shot Remote Sensing Scene Classification

Training a modern deep neural network on massive labeled samples is the main paradigm in solving scene classification problem for remote sensing, but learning from only few data points remains challenge. Existing methods few-shot sensing are performed sample-level manner, resulting easy overfitting of learned features to individual and inadequate generalization category segmentation surfaces. To solve this problem, should be organized at task level rather than sample level. Learning tasks sampled family can help tune algorithms perform well new that family. Therefore, we propose simple effective method, called RS-MetaNet, resolve issues related real world. On one hand, RS-MetaNet raises by organizing training metaway, it learns learn metric space classify scenes series tasks. We also loss function, balance loss, which maximizes ability model maximizing distance between different categories, providing categories with better linear planes while ensuring fit. The experimental results three open challenging sets, UCMerced_LandUse, NWPU-RESISC45, Aerial Image Data, demonstrate our proposed method achieves state-of-the-art cases where there 1 ~ 20 samples.