Remote Sensing Image Classification Based On Feature Pyramid Network

Data and information obtained from remote sensing images can be used in various fields.Neural networks are currently the mainstream technology for remote sensing image classification.However,there is still much room for improvement in accuracy and performance.Remote sensing image classification has become a current research focus.Most neural network models obtain less information on the spatial location and detailed features of remote sensing images,and cannot achieve excellent results in remote sensing image classification.The multi-feature hollow spatial pyramid pooling network model proposed in this paper makes up for the above shortcomings and significantly improves The accuracy of remote sensing image classification.The research work of this paper is as follows:(1)Improve the classification accuracy of remote sensing images: due to the high requirements for the spatial characteristics of remote sensing images,the semantic information of the high network layer is relatively rich,but the spatial information of the high network layer is relatively small,so it is easy to ignore the location information of the target;The spatial information of the low network layer is rich,the location information of the target is accurate,but the semantic information is less;Therefore,high and low network layers cannot be directly used for prediction and identification;Because the feature pyramid model can use both low network layer and high network layer,the idea of feature pyramid network is added to the neural network proposed in this paper.Combined with the characteristics of rich spatial information in the low network layer and rich semantic information in the high network layer,the results are predicted and identified.(2)Improve the accuracy and performance of remote sensing image classification: by improving the Atrous Spatial Pyramid Pooling module,this paper proposes a Multilevel Atrous Spatial Pyramid Pooling network module.The network module can capture more levels of semantic information in remote sensing images,help to recover the detailed information of targets,and shorten the training time by reducing the amount of parameters.(3)Improve the classification accuracy of remote sensing images: Based on the feature pyramid,this paper proposes a residual two-way channel network module combining top-down and bottom-up,which can integrate the features of different network layers,obtain more feature information and improve the final classification results.This paper conducts experiments on the data set of the public CCF remote sensing image classification and recognition competition,compares the Multi Feature Atrous Spatial Pyramid Pooling network proposed in this paper with the FPN,Deep Lab V3+,SPNet and HANet network models,and analyzes the evaluation indicators.Finally,It proves that the accuracy of the network model proposed in this paper is significantly improved compared with the latter.