基于迁移学习和EfficientNetV2的遥感图像场景分类

Remote sensing image scene classification based on transfer learning and EfficientNetV2

针对传统遥感图像分类方法分类准确率低的问题,本文提出了一种结合迁移学习与高效缩小版神经网络第二代模型(EfficientNetV2)的遥感图像场景分类方法。首先,选取参数量较少且分类精度较高的EfficientNetV2作为基础架构;其次,通过迁移学习策略,以预训练的网络参数来初始化模型,有效避免了模型的过拟合现象;最后,在航空图像数据集(AID)和遥感图像场景数据集(NWPU45)上进行实验,结果显示,该方法在这两个数据集上的分类准确率分别达到了95.76%和94.76%,充分证明了本文方法的有效性和优越性。

In view of the low classification accuracy of traditional remote sensing image classification methods,this paper proposed a remote sensing image scene classification method based on transfer learning and an efficient scaled-down second generation of the neural network model(EfficientNetV2).Firstly,EfficientNetV2,which had fewer parameters and higher classification accuracy,was selected as the infrastructure.Secondly,the pre-trained network parameters were used to initialize the model through the migration learning strategy,which effectively avoided the overfitting phenomenon of the model.Finally,the experimental results on the aerial image dataset(AID)and the remote sensing image scene dataset(NWPU45)show that the classification accuracy of the method on these two datasets reaches 95.76%and 94.76%,respectively,fully proving the effectiveness and superiority of the proposed method.