基于元学习的小样本遥感图像分类

为改善传统分类算法在小样本遥感图像分类上效果差的缺陷,提升模型的快速学习能力,提出融合迁移学习和元学习的小样本分类算法.设计基于长短期记忆网络的元学习器,通过门控结构拟合网络参数更新方式最下化损失下界,具有自动学习分类器参数更新方式的机制,相比于传统方法,能够有效扩展优化算法的搜索空间;考虑样本的跨类别知识转移和训练时...

Few-shot remote sensing image classification based on meta-learning

To improve the shortcomings of traditional classification algorithms in the classification of few-shot samples of remote sensing images,and to improve the rapid learning ability of the model,a few-shot sample classification algorithm combining transfer learning and meta-learning was proposed.A meta-learner based on LSTM was designed,which optimized the lower bound of loss by fitting the network parameter update mode through gated structure with the mechanism of automatic learning classifier parameter update mode.Compared with the traditional method,the search space of the optimization algorithm was effectively expanded.Considering the cross-category knowledge transfer and training time of samples,the idea of transfer lear-ning was used,the data of different categories were mapped to the same feature space,and the classifiers that had been under-gone representational training were trained for meta-training,so that the classifiers better grasped the overall characteristics of the categories and accelerated the training process of meta-learning.Experimental results verify the superiority of the algorithm and provide solutions for the classification of few-shot small samples of remote sensing images.