Researches On Deep Learning Based Methods For High Resolution Remote Sensing Image Classification

Classification of remote sensing images is an important research content in the field of earth observation.With the rapid development of sensor technique,the spatial and spectral resolution of images collected by remote sensing ground observation systems have been greatly improved.High resolution images can provide rich information of the objects on ground,but also bring new challenges to the remote sensing image classification.In addition,with the massive improvement of computing capabilities of computers and the explosive growth of data,deep learning has gradually become the mainstream method in the field of computer vision.How to use deep learning to mine the characteristics of remote sensing images to improve the classification performance has become a key issue for high resolution remote sensing technique.High resolution image classification includes scene classification and pixel wise classification.Specifically,scene classification emphasizes the understanding of image level semantic information,while pixel wise classification focuses on finer classification,which is complementary to each other.Based on the analysis of the characteristics of remote sensing images,this paper uses deep learning as the basic method to conduct research on remote sensing scene classification and pixel wise classification.The detail research content is as follows.(1)Traditional hand-crafted features have poor representation ability when handle the high resolution remote sensing images which contains a lot of complicated semantic information.To address this problem,this paper proposes a deep feature fusion based high resolution remote sensing scene image classification method.The proposed method first uses pre-trained deep convolutional neural networks to extract deep features at different levels from remote sensing images,and then uses covariance pooling to fuse the deep features.The characteristic of the proposed method is that: Higher-order statistical information contained in different level deep features can be exploited to extract more identifiable information.The experiment compares the traditional feature-based and other deep feature-based classification methods.The results show that the proposed method can achieve high classification accuracy with relatively lower computation complexity,which verifies the effectiveness of the proposed method in solving the problem of remote sensing scene classification.(2)High resolution remote sensing scene classification suffers from the problems of large scale variance and the extremely complicated background.To address this problem,this paper proposes a skip-connected covariance networks.The proposed method uses the skip connection and covariance pooling method to design a deep classification network to achieve end-to-end and high-precision classification of remote sensing images.On the one hand,the skip connection can combine the features of different scales in the deep network to effectively solve the problem of large scale variance in remote sensing images;On the other hand,covariance pooling can use higher-order statistical information in the features to extract discriminative semantic features,which can effectively alleviate the negative impact of complex backgrounds in images on classification accuracy.Experiments compare benchmark classification networks and other classification networks.The results show that the proposed skip connection covariance network can focus on the representative regions in the image and achieve higher classification accuracy.(3)Traditional deep neural network fails to make use of the global information in the high resolution remote sensing image.To address this problem,this paper proposes an attention mechanism based network for pixel wise classification.The proposed network simultaneously use global average pooling and global bilinear pooling to extract the global information of the middle layer.At the same time,the global information is used to learn the weights,and then the learned weight is used to reweigh the middle layer.Experiments have been carried out on several open data sets and the data acquired by GF-2.The results show that the proposed attention network can better maintain the edge and shape of the object better by using the context information in the image.(4)To handle the phenomenon of different objects have the same spectrum and the same objects have different spectrum in hyperspectral images,this paper proposes a covariance matrix representation and deep manifold learning based classification method.The proposed method can make full use of the spatial information of hyperspectral image as well as the correlation between the different bands of hyperspectral image.As a result,it can significantly reduce the intra class differences and increase the inter class differences.Experiments are carried out to compare the proposed method with other classic hyperspectral image classification methods.Experiment results show that the proposed method can achieve high classification accuracy even when there are only few training samples(5)Finally,based on the high-resolution remote sensing images obtained from the GF-2 satellite and the GF-5 satellite launched by China.We take Changsha City as the research area to carry out the practical application of the proposed methods.Specifically,it includes:1)By the scene classification of high resolution remote sensing images,the distribution of typical scenes(e.g.,the commercial and the residential)in Changsha is mapped.In addition,the scenes distribution changes in Changsha during different years are also analyzed;2)By the pixel wise classification of high-resolution remote sensing images,the distribution of buildings in Changsha is mapped;3)Using the pixel wise classification method for hyperspectral image,the water and vegetation in Changsha are accurately identified and mapped.