Remote Sensing Information On Extraction And Feature Analysis Based On Deep Neural Network

As the increasing amount of domestic satellite data and the large demand of geospatial information services,how to extract the interesting information efficiently and accurately is a hot topic with important theoretical significance and broad application prospects.The traditional methods are based on amount of feature analyses,which not only consumes manpower,material and financial resources,but also has great limitations of the features' expression ability and processing efficiency.Due to the low efficiency and poor stability of existing algorithms,we proposed an "Encoder-Perception-Decoder" convolutional neural network structure aiming at achieving high efficiency and accuarte information results.We constructed the perception module using dilated convolution and full convolution units.The main contents include as follow:(1)Encoder-Perception-Decoder Convolutional Neural Network:Modeling and AnalysisBased on the Encoder-Decoder(ED)network structure,we proposed"Encoder-Perception-Decoder" network structure for information extraction.Depth feature can be definied by convolutional neural network model to visualize different deep convolutional characteristics.Meanwhile,we definied the relative and absolute depth within the numbers of pooling layers of the network architecture.As the assosiation of remote sensing image resolution and the actural object size,we proposed an optimal relative depth model for information extraction.Through experimental comparison and analysis,the appropriate depth interpretation model can effectively improve the recognition and expression ability of remote sensing imagery.(2)Multi-Scaled end-to-end Encoder-Perception-Decoder Convolutional Neural Network:Cloud Extraction MethodAccording to the problem of heterogeneous appearances with low interclass variation and noise interference in the process of cloud detection,based on the above network structure,we propose a multi-scaled end-to-end encoding-perception-decoing convolutional neural network cloud extraction method.It is based on the theory of full convolution without the limitation of inputs' size,the network structure is extended from the with to longitude,integrating the expansion results of different scales on the task extraction results,and then increases the scope on the upsampling process.The constraints parts reduce the impact of edge aliasing problems due to small size.We used ZY-3 satellite images to verify the validity and robustness.The results show that the method we proposed can reduce the impact of noise interference,the overall accuracy of the compared methods,which are OTSU,FCN-8S,UNET and Deeplab,ours can improve on 9.76%,3.08%,1.59%and 1.48%,meanwhile,the spatial depth feature of cloud has better linear separability(3)Refined Residual Encoder-Perception-Decoder Convolutional Neural Network:Road Extraction MethodAiming at the difficulty of road information interpretation caused by the complex spatial structure of high-resolution remote sensing images and the problem of road occlusion,we propose a refined residual encoder-perception-decoder convolutional neural network method for road extraction.It is included residual convolutional neural network algorithm and post process.Fistly,based on the residual connection unit and encoder-perception-decoder structure,the network is trained to obtain the task features automatically.Secondly,the post process consists of math morphology and tensor voting algorithm for repairing the broken road,especially for the straight lines.To verify the efficiency of the proposed method,we achieve the experiment on two datasets,compared with others,ours can improve the OA on 1.86%,0.85%and 1.55%with other methods which are CNN,UNET and GL-Dense-U-NET.Above all,the method we proposed can enhance roads' topological performance efficiently.In conclusion,the encoder-perception-decoder network structure we proposed is used for information extraction from remote sensing images.The method owns versatility and universality in some degree.In the background of bigdata,deep convolutional neural network can gradually improve the information extraction accuracy,which can greatly reduce the manual extraction workload of remote sensing image information and shorten the information extraction cycle of remote sensing imagery.The study provides technical support for the daily quality inspection of domestica satellites remote sensing imagery and the normal monitoring of thematic tasks,and has reference for related research.