Building Extraction From High Resolution Remote Sensing Images Based On Convolution Neural Network

The rise and development of remote sensing provides massive data for the development of various fields,including urban planning,land use and so on.Remote sensing images are all inclusive and rich in ground information.The application of remote sensing images is mainly focused on the accurate extraction of ground objects.Buildings,roads or other large projects in the urban area of high resolution remote sensing images account for more than 80%.Accurate,automatic and rapid extraction of buildings provides a basis for the use of subsequent remote sensing images.At present,building detection and extraction of remote sensing images has been studied,but most of them need to be implemented manually.Some theories and methods are not mature enough to meet the actual needs.Especially in residential areas,because of the differences in building age,residence time and maintenance,the spatial distribution of buildings,the types and distribution of the land cover are complicated.In addition,the structure of the building itself is diverse.These factors cause the methods of automatic extraction difficult and the effect not that good.Therefore,the automatic extraction of buildings still has broad research prospects and important significance.In view of this phenomenon,based on the convolution neural network,an effective method of building recognition and extraction is proposed based on the high resolution remote sensing image of urban area.The main contents and results of this study are as follows:(1)Establishing a high resolution,multispectral remote sensing image training dataset for deep learning based on convolutional neural network(CNN).Using the spatial resolution from 20cm to 30cm,combined with the ground database(ground truth),the orthophoto house and other categories of samples are manually clipped and enhanced to build a large data set with 1642 samples,where the picture size is 160X160.There are 1314 training samples and 328 test samples.They are divided into the categories of buildings and others.The houses' pictures are taken as positive samples and the others as negative samples.(2)The performance of different convolution neural network models in building data sets of remote sensing images is compared in the research.Many of the existing CNN networks can achieve very high accuracy in classification of images,and practice proves that it has good scalability.Therefore,the general classification tasks will directly select one of these network connections to fine tune the network size.Since the contents of remote sensing images are different from those of ordinary pictures,especially in terms of the texture,some common convolution neural networks Alexnet,VGG,Googlenet and ResNet are not known for the effect for feature expression and extraction of remote sensing images.Therefore,we investigated these network models on this experimental data set to select the most suitable model for remote sensing images.It is limited to the number of manual set of samples,and the general CNN usually requires up to tens of thousands of data to complete the learning task,so we need to study various ways and techniques to improve the performance of network learning.We obtained an ideal model structure through a large number of experiments and fine tuning of parameters.We earned and analyzed data characteristics by using sample data to train the neural networks,.(3)The research adjusted the structure,optimized the parameters of the model,and combined selective search to segment the image.The automatic recognition and extraction of the buildings in the large range remote sensing image are realized.Study for a suitable classification network for target recognition based on feature learning.The classical support vector machine(SVM)is good at the classification of small samples,and itself is a one to one classifier.We study the situation of SVM as a classification network firstly.However,by the comparison of the classification results,a better classifier namely softmax is selected.With this model,the accurate classification of buildings and other objects is completed,with a correct rate of 100%.Combined with Selective-search method,the uniform area in the whole remote sensing image is extracted as the candidate building sample to be recognized to realize automatic recognition and extraction of buildings.Conclusion:Convolution neural network can not only identify and locate targets in small images,but also can be applied to target extraction in remote sensing images.Our research has obtained a primary building recognition network,which can obtain more powerful recognition by continuing to enter new samples,and obtain more accurate target location by increasing the probability threshold of the network output.The advantages of this method are high speed,high positioning precision,saves time and manpower.It provides a new idea for building extraction of high resolution remote sensing images.It is of great significance for future application of urban planning and building geometric modeling.