Land Cover Classification With High Resolution Remote Sensing Images Using Deep Domain Adaption

Land cover classification based on high-resolution remote sensing images is of great significance to territorial spatial planning.However,the current classification methods mostly focus on traditional machine learning algorithms such as support vector machines and decision trees.However,this methods cannot be applied to semantic segmentation tasks with different data sources or in different study areas.Based on deep learning technology can efficiently and accurately provide ground object classification information,enduster remote sensing interpretation with more intelligence and automation,and provide dynamic data information for land and resource departments to supervise land use.However,there are few studies on large-scale and high-precision land cover classification.However,pixel level groud trurh is time-consuming and laborious,which not only has sufficient quantity,but also has high requirements for annotation quality.Therefore,it is of great significance to implement model migration by applying the source domain data with annotated information to the unannotated target domain.In this study,we finshed the land cover classification of Pleiades high-resolution images by a seires improved fully convolutional neural network.Then we adopted deep DA to segment RGB remote sensing images from different satellites and locations semantically,so that the same model can also realize the discrimination of land cover classification in the field of lack of labeled data.The main contents are as follows:(1)Firtly,we built a set of land cover classification data by using Pleiades HRRS of Zhuhai which acquired in February 2020.Then explored The applicability of the improved semantic segmentation models,on high-resolution remote sensing images,such as UNet,PSPNet and Deep Lab.The accuracy of the VGG16＿UNet with the most parameters in the above data sets is 92.51%,which is 1.37% and 0.67% higher than that of Res50＿UNet and Res101＿UNet.Resnet50 and Res Net101 have deeper network architecture than VGG16,which can effectively reduce the number of network parameters and reduce overfitting without much loss of accuracy.With same backbone network,the accracy of Unet is highest but Deeplabv3 Plus has fewer parameters and shortest training time.Then,five improved semantic segmentation models were used to classify and predict the images of Jinwan District in Zhuhai City,and the large-scale and refined land cover classification mapping was completed.(2)Secondly,Compared with the direct use of Deep Lab V3 Plus across domains,the classification accuracy of Deep Lab V3 Plus with Domain adaption is improved by5.91% in the Zhuhai Pleiades image to the Zhuhai Tianditu,and by 8.35% in the Zhuhai Pleiades image to the Guangzhou Tianditu.The accuracy of forest land,agricultural land and water surface improved more,while the classification accuracy of buildings changed little.After the domain adaption,the classification accuracy of bare land decreased.The domain adaption method can achieve the model migration without requiring the strict alignment of the image features of the two domain,which greatly saves the time cost of annotation.The experimental results show that the domain adaption algorithm is effective in the cross-domain application of remote sensing image classification,but the classification accuracy needs to be improved.