The Research On Adaptive Road Extraction Method Of Remote Sensing Image Based On Deep Learning

How to extract information from remote sensing image quickly,efficiently and intelligently has always been one of the research focuses of remote sensing image interpretation.However,due to the influence of noise and complex surface environment(such as tall buildings,trees,etc.),traditional remote sensing image interpretation methods have low extraction accuracy and low automation.In recent years,deep learning technology has developed rapidly,and some progress has been made in the research of remote sensing image information extraction based on deep learning.However,at present,information extraction from remote sensing image is mainly carried out at a specific spatial scale.In different spatial scales,the image features of terrain will change.Single depth learning model is difficult to adapt to different spatial scales of information extraction.To solve the above problems,proposing an adaptive remote sensing image extraction method based on depth learning,and the road is taken as an example.Specific research work is as follows:(1)Intelligent Identification of Spatial Scale of Remote Sensing ImageBased on the hierarchical definition of online remote sensing image,a quantitative method of spatial scale of remote sensing image is established.Taking the sky map as tile data source,a sample library for spatial scale identification of online remote sensing images is established.By means of in-depth learning and convolution neural network(CNN),the intelligent recognition of spatial scale of remote sensing image is realized.Through the research,it is found that the accuracy of the identification level of 14-18 tiles is more than 90%,but the accuracy of the identification level of 13 tiles can only reach 73.14%,which can not be effectively applied to adaptive road extraction.To solve this problem,an intelligent spatial scale recognition method based on ensemble learning simple voting method is proposed.The image is decomposed into several sub-blocks according to tile size,and then the spatial scale of each tile sub-block is distinguished by using depth learning model.Finally,the most distinguished level of computer is used as the spatial scale of image prediction.Experiments show that the recognition accuracy of this method reaches 100%,which lays a foundation for adaptive road extraction.(2)Research on Deep Learning Method of Road Extraction at Different Spatial ScalesFirstly,typical samples containing road information are selected,road surface is manually drawn,and road sample databases of different levels are constructed and perfected.On this basis,road extraction depth learning models of different spatial scales are obtained by training the sample databases.In remote sensing image classification,in order to extract pixel-level information,the U-Net model is introduced into road extraction.According to the experiment,it is found that the U-Net model has a good effect on road extraction of remote sensing images at levels 17 and 18,but the effect on road extraction of remote sensing images at levels 16 is general.Therefore,some improvements are made to the model and then road extraction is carried out,which achieves better results.(3)Research on Adaptive Road Extraction of Remote Sensing ImageOn the basis of intelligent recognition of spatial scale of remote sensing image and adaptive road extraction method,a decision tree for adaptive road extraction research of remote sensing image is established.Firstly,the input image is judged intelligently in spatial scale,and the depth learning model is extracted by calling a specific level of road according to the result of the judgement.Practice proves that the method in this paper has high accuracy in road extraction on all spatial scales,and its research methods and technical routes have certain theoretical significance and practical value.