Research On Classification Of Multi-temporal Remote Sensing Image Based On Deep Learning

Land cover classification is an important and widely studied field in earth observation.With the development of aerospace technology,the acquisition cost of remote sensing image is getting smaller and smaller,which promotes the rapid development of remote sensing image classification technology.Among them,high-resolution image classification based on deep learning has shown outstanding performance in practical applications such as building classification,water resource protection and urban extraction.But unlike buildings and rivers,such as plants and crops,the morphology of object was affected by season.Its present different forms in different periods in the remote sensing image,the problems with different spectrum with the spectrum of foreign bodies,classification accuracy cannot meet the actual demand,thus exploring Multitemporal classification of remote sensing technology has important research value.Recurrent Neural Network(RNN)has been successful in machine translation,speech recognition and other fields,and is gradually expanding to other fields.In this paper,the idea of RNN to extract time series is integrated into the semantic segmentation network architecture,and an efficient multi-temporal semantic segmentation network is proposed.The main work of this paper is summarized as follows:(1)At present,the traditional classification algorithm makes insufficient use of spatial and spectral information in high-resolution remote sensing images,and the feature engineering design is complicated and the generalization ability is poor,resulting in the lack of accuracy.Therefore,this paper discusses the application of deep learning network in high-resolution remote sensing image.The encoder-decoder semantic segmentation network model is proposed,which can effectively extract spatial and spectral information of remote sensing image and realize high precision automatic ground object extraction.(2)A fast deep learning semantic segmentation model called Fast-Deeplabv3+ was built for single-time image.The model is improved on the latest Deeplabv3+ model,paying more attention to the scale features and considering the balance between speed and accuracy.The experimental results show that fast-deeplabv3 + has an excellent performance on the single phase image,the classification accuracy is better than Deeplabv3+,u-net and other popular deep learning networks,and has the characteristics of small network memory consumption and Fast reasoning speed.(3)For multi-temporal images,FMTSN network is constructed to effectively process temporal features,so as to improve the expression ability of network features from three aspects of time,space and spectral information,and then find the general characteristics of ground objects.The LSTM network is innovatively applied to multi-temporal remote sensing image classification task by referring to the research of RNN in machine translation.The Patch-LSTM module employed conv-lstm unit,solved the unit memory problem and spatial information loss with Seq2 seq model.In the multi-temporal image,FMTSN achieves excellent accuracy,which greatly improves the classification accuracy of farmland and greenhouses containing green plants.(4)There are very few open remote sensing data sets,which is because the tagging of deep learning data sets requires a lot of manpower and resources.In order to improve the efficiency of labeling,this paper tries to use data recharge technology to reduce the amount of manual involvement and improve the quality of labeling.In addition,a set of remote sensing data processing processes including radiometric calibration,orthographic correction,atmospheric correction,fusion and data enhancement are explored.These pretreatment processes improve the final classification accuracy to some extent.