RA-ProtoNet:基于元学习的小样本遥感场景分类方法

RA-ProtoNet:Classification Based on Meta-Learning for Few-Shot Remote Sensing Scene

深度学习在解决遥感影像场景分类问题中发挥了重要作用,但在某些特定的遥感场景分类问题中,存在可训练带标签样本严重不足的情况(单类样本数少于10),造成现有的传统深度模型分类效果不理想。针对上述问题,提出一种小样本遥感场景分类方法,并构建一种基于元学习(meta-learning)训练策略的模型ResNet14-Attention-ProtoNet(RA-ProtoNet)。首先,采用预训练的深度残差网络ResNet14作为特征嵌入模块,提取遥感影像深度特征;其次,针对同类样本特征不明显会对类级(class-level)表达造成的干扰问题,在类级表达模块,采用基于双向长短期记忆网络(BiLSTM)的注意力机制强化类内样本信息,生成样本的类级特征表达;最后,利用欧氏距离度量待分类样本与类级特征之间的距离,实现分类预测。在UCMERCED、AID-30和NWPU-RESISC45等3个遥感影像数据集上,将所提方法与基于迁移学习和现有元学习方法的遥感场景分类方法进行对比实验,在5-way 5-shot条件下,所提方法的整体场景分类精度分别达到81.30%、83.29%和81.22%。实验结果表明,所提方法可以有效挖掘类内样本信息,在极小样本条件下获得更高的遥感影像场景分类精度。

Deep learning plays an important role in solving the problem of remote sensing image scene classification.However,in certain remote sensing scene classification problems,samples with labels that can be trained are severely lacking(number of singleclass samples less than 10),resulting in unsatisfactory classification using existing traditional depth models.In this paper,to solve these problems,a smallsamplesize remote sensing scene classification method is proposed,and a model called ResNet14 AttentionProtoNet(RAProtoNet)based on a metalearning training strategy is constructed.First,in the feature embedding module,the pretrained depth residual network,ResNet14,is used to extract the depth features of remote sensing images.Second,in the classlevel expression module,the problem that the features of similar samples are unremarkable and interfere in classlevel expressions is solved.For this purpose,an attention mechanism based on bidirectional long shortterm memory(BiLSTM)is used to strengthen the sample information within a class and generate classlevel feature expressions of samples.Finally,the Euclidean distance is used to measure the distances between the samples to be classified and the classlevel features for classification prediction.On three remote sensing image datasets,including UCMERCED,AID30 and NWPURESISC45,the proposed method is compared with remote sensing scene classification methods based on migration learning and existing metalearning methods.Under the fiveway fiveshot condition,the overall scene classification accuracies of the proposed method reach 81.30%,83.29%,and 81.22%,respectively.The experimental results show that the proposed method can effectively mine the sample information within a class and obtain higher classification accuracy of remote sensing image scenes under the condition of minimal samples than the other methods.