Remote Sensing Image Scene Classification Based on Attention Mechanism and Feature Fusion
Aiming at the problems of limited training samples and insufficient representation of the relationship between ground objects and background in the remote sensing image scene classification task,a feature fusion method for remote sensing image scene classifica-tion is proposed.Firstly,a convolutional neural network(CNN)pretrained on ImageNet dataset is employed as a feature extractor.Addi-tionally,an attention mechanism is introduced to highlight the spatial location information,strengthen the contextual relationship,and en-hance the feature expression.Subsequently,the enhanced convolution features are fused with the fully connected layer features and used for scene classification.Thereby,the proposed method obtains a fused feature expression with good discriminability,which is demonstra-ted through experiment on the UC Merced dataset.Later,it is applied to the land use classification task of Gaofen-2 satellite images.The overall classification accuracy reaches 92.83％,and its performance is on a par with that of other advanced methods.
万方数据
维普
