Multi-scale Local Representation Learning For Remote Sensing Image Scene Classification

Recently,with the development of remote sensing(RS)technology,the number and resolu-tion of RS images that people can obtain has increased dramatically.RS images has many advantages,e.g.,wide observing range,short cycle and abundant surface information.To obtain and process RS images more efficiently,the scene classification task of RS images has attracted more and more researcher’s attention.The essence of RS image classification is the process of mapping image’s content to semantic information labels.However,due to the complex contents within the RS images(i.e.,the targets are huge in volume and diverse in type),the classification results obtained by the CNNs cannot reach the satisfactory stage.Due to weak anti-noise ability and high computational complexity,the traditional methods are usually unable to accurately and quickly classify RS images.(1)Multi-scale attention for remote sensing scene classification: Considering that the cur-rent classification model lacks consideration of local features,our method uses the channel attention mechanism to obtain the local characterization features of remote sensing images.Through experiments,it is found that the channel-level attention mechanism in remote sens-ing image scene classification have the problem of excessive concentration of attention area.Considering the above problems,the model designed the attention multi-scale module to ex-pand the attention area of the attention mechanism,which improved the robustness of the model and improved classification accuracy.(2)Attention consistent network for remote sensing scene classification: Traditional con-volutional neural networks lack the acquisition of local features of remote sensing images,which limits the classification performance of convolutional neural networks.Considering the above problems,a parallel attention mechanism is designed to obtain the local features of remote sensing images from the two dimensions of channel and space and fuse them.In or-der to further improve the discriminability of remote sensing influence features,this method considers the relationship between remote sensing images(reduce the interclass differences and increase intra-class variations of remote scene images),the attention consistent model is desigend to unify the attention area of remote sensing images which effectively improvs the classification accuracy.(3)Global and local information transformer for remote sensing scene classification: Con-sidering the large amount of computation and overfitting of convolutional neural networks due to multiple stacking of convolutional layers,this method uses the Transformer model which shines in the field of natural language processing,in which the self-attention mech-anism can capture the global information come from RS images and reduce the amount of calculation.Taking into account the complexity of remote sensing images,this method de-signs a joint loss function,which combines the global and local features of remote sensing images to be used in remote sensing image classification tasks.Experiments show that this method has excellent performance for remote sensing image scene classification tasks.The experimental results demonstrate that our proposed methods can obtain more represen-tative feature and get better classification performance.