Few-shot Learning For Remote Sensing Scene Classification

Remote sensing scene image is a kind of important sensing information,which contains a large amount of semantic information,such as space,texture,structure,etc.It has a wide range of applications in the field of national economic construction.Hence,it is extremely important to understand huge and complex remote sensing images.However,compared with national images,remote sensing scene images are relatively scarce and valuable.Traditional deep learning method easily leads to overfitting and fails to learn the real data distribution under few data.Therefore,it is of great value to classify remote sensing scene images with very few samples.Few-shot learning aims to allow the model quickly generalize to new categories with a small number of samples.This paper uses deep learning and meta-learning method to deal with remote sensing scene image classification with very few samples.The main contributions of the paper are as follows:(1)We propose a few-shot classification method MACPNet(Manifold Augmentation and Calibrated Prototype Network).The method expands samples through manifold augmentation module,to alleviate the problem of insufficient data,and meanwhile fine-tune the model,so that the model can learn the adaptive feature to the new tasks.At the classifier level,we proposed a calibrated prototype classifier to alleviate the problems of large intra-class differences and high intra-class similarity of remote sensing scene images.Finally,the experiment on both NWPU-RESISC45 and AID dataset proves the effectiveness of our method.(2)We propose a few-shot classification method AMFFNet(Attention-based Multi-scale Feature Fusion Network).Firstly,the method applies several parallel atrous convolution with different rates to extract multi-scale features.Secondly,as for the feature fusion module,the method uses the channel attention mechanism to learn the importance of different scale features,so that the model can autonomously learn the scale features that have a greater impact on the classification results.Finally,the experiment on both NWPU-RESISC45 and AID dataset proves the effectiveness of our method.(3)We propose a few-shot classification method jointly learning with contrastive learning and prediction of rotation angles.Firstly,the method fully mines the internal representation information of images with self-supervised contrastive learning and predicting rotation angle,so as to alleviate the problem of lack of samples.Secondly,we integrate few-shot learning task and self-supervision task through multi-task framework,and explored the influence of different self-supervised tasks.Finally,the experiment on both NWPU-RESISC45 and AID dataset proves the effectiveness of our method.