Design Of Remote Sensing Image Analysis System Based On Deep Neural Network

Remote sensing is a comprehensive technology based on aerial photography,which combines spectroscopy,meteorology,geography,computer science and many other disciplines.Modern remote sensing technology plays an important role in the fields of marine resources development and utilization,weather analysis and forecasting,land and resource exploration,agricultural crop estimation,military reconnaissance,and smart cities.The development of satellite launching technology has promoted the research of high spectral resolution remote sensing,high spatial resolution remote sensing,thermal infrared remote sensing,synthetic aperture radar remote sensing.The acquisition of high-resolution remote sensing images become more convenient.Traditional remote sensing image interpreting methods can no longer meet the needs of remote sensing analysis in the context of big data.To extract more information in high-resolution remote sensing images,this thesis conducts a series of studies on the problems of haze removal,scene recognition,and semantic segmentation.An end-to-end system model was proposed,and transfer learning on remote sensing datasets was explored.Firstly,a training sets were built with the simulated haze image slices.Referring by the main architecture of existing deep neural networks,a network was designed and trained,which achieve an improvement in real-time performance.Secondly,with the use of significance test and AP clustering,the remote sensing image was preprocessed and sent to the offline training classification network for scene prediction.In the off-line training phase,transfer learning method help to fine-tune the pre-training models on other data sets,speeding up the convergence speed and reducing the training time.Finally,we extend the image-level classification task to pixel-level classification tasks and semantically segment remote sensing images.Two types segmentation networks based on the SegNet model was trained.The results on the test set outperformed the traditional method.The research of this thesis combined deep learning with remote sensing technology,and tried to use transfer learning to solve the data needs of deep learning.It provided a certain technical basis for the wide application of remote sensing technology,and provided ideas for deep learning in new areas.