Research On The Extraction Of Buildings From High-resolution Remote Sensing Images Based On Deep Learning Methods

There is very rich feature information in high spatial resolution remote sensing images.At present,how to quickly and accurately extract the feature information that people need from high spatial resolution remote sensing images is one of the hotspots in the application of remote sensing technology..Since the composition of most cities has a certain complexity,the extraction of urban building information has become the difficulty and hot spot of high spatial resolution remote sensing image processing.Building information plays a very important role in urban planning and management,urban construction and other related work,and is a kind of basic geographic information data.The improvement in accuracy and efficiency of high spatial resolution remote sensing image building extraction technology is conducive to the in-depth application of high spatial resolution remote sensing image in urban planning and management,smart city construction and other fields.At present,in the practical work of processing the extraction of building information from high spatial resolution remote sensing images,traditional objectoriented classification methods and shallow machine learning methods are usually used.Object-oriented classification method requires manual selection of segmentation parameters and special selection,and the classification process requires a lot of manpower and time cost;building information extraction based on shallow machine learning method is difficult to get good extraction due to the shallow structure of the model effect.Deep learning technology is an extension of machine learning technology and an emerging technology that has developed rapidly in recent years.It has been successfully applied in many fields such as image recognition,speech recognition and automatic driving.Deep learning methods can automatically learn and extract shallow layers of images With deep features,the use of these features for automatic classification of remote sensing images,this method has achieved certain results and has greater application prospects.In order to obtain better results of automatic extraction of building information,and to promote the rapid,accurate,and automatic extraction of high spatial resolution remote sensing image building information and successfully apply it to actual projects,this paper carried out high spatial resolution based on deep learning methods Research on Building Information Extraction from Remote Sensing Images.The main research contents and conclusions of the paper are as follows:(1)The UNet ++ network type widely used in the field of medical image processing has been researched and implemented,and the building data set has been used for training and building information extraction experiments.The experiment shows that UNet ++ has a higher level in the original network state Compared with the typical U?Net,the model accuracy and semantic segmentation ability have greater advantages in building information extraction;(2)Pyramid segmentation network(Pyramid Scene Parsing Network,PSPNet)can realize the understanding ability of global scene context information,but its network structure is not simple,the parameter is large,the calculation complexity is large,and the dense convolutional network Dense Convolutional Network(Dense Net)has the characteristics of making full use of features and reducing the number of model parameters.Dense Net is introduced into PSPNet to improve the network model and compress the model.The results show that the optimized network model parameter amount is only 0.68 of the original model parameter amount,which greatly reduces the calculation amount of the model;through training and testing,it shows that the optimized network model reduces the model parameter amount while ensuring that the network model is in the building Information extraction accuracy;(3)Aiming at the problems of excessive calculation of the model and too many model parameters in the Deep Labv3 + network model,study and introduce the lightweight network model Moblienetv2 to optimize and improve it;in order to make the data set meet the training requirements of the network model,experiment The data set used has been data enhanced.After improvement,the final parameter of the model improved by Moblienetv2 network is only 10 mb,and the verification set MIo U and test set MIo U of the network model have reached more than 90%;(4)The building datasets of some areas in the main urban area of Kunming were produced,and the building extraction experiments were carried out using UNet ++,New?PSPNet,and the improved Deep Labv3 + model.The experimental results show that the improved Deep Labv3 + model extracts the building accuracy and The robustness of the model is optimal,and its MIo U on the building dataset reaches 0.86.