Research Of Scene Classification Method For High Resolution Remote Sensing Images Based On Deep Feature Representation

With the rapid development of advanced remote sensing technology,a considerable amount of high-resolution remote sensing images have been applied to natural disaster monitoring,urban planning,ecological environment assessment,and so on.Due to the wide applications of remote sensing images,it is of importance to interpret high-resolution remote sensing images automatically and accurately.As one of the interpretation bases for remote sensing images,remote sensing scene classification has attracted increasing attention.Different from the traditional pixel-level and object-level classification research of remote sensing images,remote sensing scene classification is a scene-level interpretation task.The purpose of scene classification is to assign scene semantic labels(school,railway station,etc.)to remote sensing images according to human understanding.Therefore,how to obtain scene features of high-resolution remote sensing images and perform scene-level semantic understanding has become a hot issue in the current remote sensing image interpretation work.In more recent years,extensive efforts have been made to develop remote sensing scene classification methods and the effect of scene classification has been greatly improved.However,the existing remote sensing scene classification algorithms may not perform well for classifying high-resolution remote sensing images with highly complex spatial structures and extremely rich scene information.By analyzing the imaging characteristics of high-resolution remote sensing images,and summarizing the previous work systematically,this paper studies scene classification methods based on the deep feature representation.The major research contents and innovations include:(1)Aiming at the problem that the deep features of convolutional neural networks cannot fully describe the characteristics of remote sensing scenes,this paper proposes a remote sensing scene classification method based on multi-scale deep feature representation(MDFR).The algorithm first combines the feature map selection and region representation algorithm to generate local features of scenes,then obtains global features of scenes based on the fully connected layers,finally develops a feature-level fusion method(element-wise addition)to fuse the local and global features.Considering the local and global information of remote sensing scenes,the algorithm improves the representation ability of deep features for convolutional neural networks.The experimental results of public data sets have indicated that the proposed method can effectively improve the classification accuracy of remote sensing scenes.(2)Due to the within-class diversity and between-class similarity for remote sensing scenes,the discrimination of deep features for convolutional neural networks is lower.To address the problem,a semi-supervised center loss remote sensing scene classification algorithm(SSCL)is proposed.Based on the fact that there are few labeled remote sensing scene samples,the algorithm improves the center loss to a semi-supervised form with the unlabeled scene samples,and constructs an end-to-end scene classification framework.By the fusion of scene information for labeled and unlabeled samples,the discrimination of deep features for convolutional neural networks is significantly enhanced.Experimental results have demonstrated the proposed method improves the classification performance of remote sensing scenes.