Research On Remote Sensing Image Classification Based On Deep Learning

In the field of Image Recognition,remote sensing images have always been main issue.They are also applied in many domains,such as Geographic Mapping,Environmental Monitoring,Feature Recognition and so on.However,many problems are still exisited in the recognition technology of remote sensing images.On the one hand,Applying deep learning models to classify remote sensing images can not match the characteristics of remote sensing images very well.On the other hand,remote sensing images have more complex features and these complex features are main reason to a lower classification accuracy than traditional images.Therefore,the following work has been carried out:In order to focus on the content of remote sensing image classification and deep learning,current research status and research progress are reviewed systematically.At the same time,we analyzed and summarized these achievement.To verify the superiority of a new model(VGG-X)designed by us,we also compared it with two classic CNNs based on the WHU-RS19 and UCMerced datasets.In order to transform the improved model(VGG-X)to process pixel-level classification,the FCN rules must be utilized.After studying the FCN model in depth,FCN-VGG-X is desiged in this phase.To our disappointment,it is weak to process pixel-level classification with FCN-VGG-X in our experiments.However,this defect can be amended.With the extended CRFs,we find the pixel classification accuracy can be improved.Conclusions:(1)Image-level recognition: On the WHU-RS19 dataset,the improved deep learning model achieves 95.12% classification accuracy,0.69% and 2.72% higher than state-of-the-art methods,GoogLeNet and VGG-16 respectively;On the UCMerced dataset,it also achieves a classification accuracy of 92.51%,which is 0.62% and 1.17% higher than GoogLeNet and VGG-16 respectively.These datas both show that a new constructed CNN can improve the performance on the related datasets.(2)Pixel-level recognition: the extended binary function of CRFs improves the ability of recognition in this process,especially for recognition on the shadow.The experiments show that our two optimization schemes have certain significance in the two fields of remote sensing image classification.