Research On Scene Classification Method Of Remote Sensing Image Based On AOI Semantic Information

Remote sensing image scene classification is the basic problem of remote sensing image analysis,processing and application,and an important means to reveal the spatial distribution characteristics and laws of geographical objects in the objective world.The development of aerospace technology has significantly improved the resolution of remote sensing images,providing an important data source for the study of urban scene classification.The traditional scene classification mostly relies on the underlying features of remote sensing images,and the features are designed manually,which consumes manpower and material resources,and the effect is still unsatisfactory.The development and application of deep learning technology can significantly improve the accuracy of scene classification,but in the case of complex scenes and multi-objective sample imbalance,scene classification based on deep learning still faces many challenges.Aiming at the difficulties faced in high-resolution remote sensing image scene classification,this paper develops a remote sensing image scene classification method that integrates AOI(Area of Interest)semantic information.First,an optimized NC-Net(N-type Global Context Net)model is proposed based on the baseline U-Net.The model level effectively improves the scene classification accuracy.Afterwards,the social and geographic attributes of AOI are used to integrate semantic information into image scene classification,to solve the problem of insufficient classification information for a single image scene,and to assign social semantic information to images.It can be applied to the actual urban environment research.The main work of this paper can be summarized into the following three points:(1)Summarize the basic methods of remote sensing image scene classification.This paper introduces methods based on low-level features,methods based on middle-level features,methods based on high-level features,and methods based on semantic content,and focuses on high-level features,that is,methods of deep learning neural networks.(2)Spatial Global Context Information Network.An optimization network NC-Net based on U-Net is proposed.U-Net itself has a relatively simple structure,and the scene classification accuracy is lacking.For specific problems,a multi-scale output fusion and resampling stage is proposed,the loss function of U-Net is optimized,and multi-scale loss function cascade is proposed to allow the network to learn features more fully.Obtaining context information effectively improves the accuracy of scene classification at the model level.(3)Fusion AOI semantic information approach.In the post-model processing stage,based on the scene classification map of NC-Net,the final scene classification map is obtained by using the social and geographic attribute information attached to the AOI,and the weighted fusion method is used to integrate the semantic information of multiple parties.The scene classification categories are unified with the actual needs of urban planning,and the AOI and image classification categories are mapped to the defined scene classification standards using consistent classification standards.The final classification map has good scalability and can be applied to the scene classification research of different cities.