【摘要】 图像的特征提取及研究对于遥感场景分类有很重大的意义。在研究遥感图像场景分类时,利用两种预训练卷积神经网络模型VGGNet-16和ResNet-50提取全局特征,然后采用两种特征融合策略进行特征处理。策略一先使用主成分分析(PCA)对特征降维再进行特征融合,策略二则先融合特征以后再降维,最终两种融合策略得到的特征通过随机森林分类器得到分类的结果。在开源数据集UCM和AID上的实验结果表明,两种策略均具有较好的分类能力,并提高模型的训练效率。更多还原

【Abstract】 The extraction and research of image feature have great significance for remote sensing scene classification. In this paper, two kinds of pretrain convolutional neural network models VGGNet-16 and ResNet-50 are used to extract global features, and then two kinds of feature fusion strategies are used for feature processing. In strategy one, PCA is used to reduce the dimension of features before feature fusion, while in strategy two, features are fused first and then dimension is reduced. Finally, features from two fusion strategies are classified by random forest classifier. The experimental results on UCM and AID show that the two strategies have better classification ability and improve the training efficiency of the model.更多还原