High Resolution Remote Sensing Image Classification Based On Attention Guided Convolutional Neural Network

As a large-scale surface monitoring method,remote sensing is of great significance for geographic national conditions monitoring,environmental change research,military target identification,and sustainable development planning.Compared with low and medium resolution remote sensing images,high resolution remote sensing images contain more details of ground objects,such as,richer texture,shape,topology and adjacency information,which can provide objective and reliable information for intelligent interpretation task.In recent years,although the classification of high-resolution remote sensing images and the segmentation of typical surfaces based on convolutional neural networks have made great progress,there are still some urgent problems to be solved.Firstly,in large and wide remote sensing images,there are some common problems,such as serious sample imbalance,"large intra-class differences,small inter-class differences" and difficulty in obtaining dense multi-scale information,which lead to low overall classification accuracy,especially in small and less samples.Secondly,the complex boundary of ground object and the lack of spatial structure information,global context information and boundary information lead to the incoherent segmentation boundary and serious sawtooth phenomenon.Therefore,in view of the above problems in high-resolution remote sensing image classification based on convolution neural network,this paper aims to carry out relevant research.The main research contents and results of this paper are as follows:(1)Aiming at the problems of imbalanced high-resolution remote sensing image samples,"large intra-class differences,small inter-class differences" and difficulty in acquiring dense multi-scale features,the high resolution remote sensing image classification method based on the attention-guided multi-scale spatial and channel information joint are proposed in this paper,named PCASPPNet.This method includes a parallel structure composed of channel attention module,spatial attention module and atrous spatial pyramid pooling(ASPP),which alleviates the low utilization rate of input features and the neglect of some useful information in ASPP module,and assists in guiding the aggregation of multi-scale spatial and channel information to obtain dense multi-scale features.For the Vaiheigen and GID datasets,the experimental results show that the classification accuracy of PCASPPNet is significantly improved compared with multiple classic methods,especially in the classification of small-scale targets and less samples.(2)Verify the response mechanism of different attention modules in remote sensing images,and design ablation experiments to explore the information aggregation ability of different attention modules.The results show that CCAM only responds to the semantic information on the "criss-cross path" of the marked point,but RCCAM can obtain the relevant information of the marked point in the whole image by recursing CCAM twice,and PAM captures the similarity semantic information and long-distance dependence of the same category by directly establishing the relationship between a certain pixel and other pixels in the remote sensing image.Therefore,in terms of global context information aggregation capability,PAM is the best,RCCAM is the second,and CCAM is the worst.In addition,CAM clearly responds to different categories by simulating the dependency between different channels.(3)Aiming at the problem of edge pixels being prone to misclassification caused by insufficient spatial structure information and boundary information of ground features,we design a typical ground feature fine segmentation network based on two parallel branches composed of boundary extraction and semantic segmentation,named GAFSNet.In this network,the semantic segmentation branch is used to obtain the discriminative characteristic of the ground features which for clarifying what the ground features are.The edge detection branch is used to obtain the accurate position and boundary features of the ground features which for clarify where the ground features are.The "what-where" joint learning method improves the fine-grained representation of the network,and then overcomes the problems of incoherent segmentation boundary and serious sawtooth phenomenon.Meanwhile,the pointrend module is used to improve the classification accuracy of baseline method DeeplabV3+and FPN at the boundary based on iterative subdivision strategy.For WHU datasets,the improved baseline method can adaptively render anti-aliasing high-quality segmentation results.In addition,compared with the baseline network and the improved baseline method,GAFSNet has been proved to achieve better results in the fine segmentation of buildings.