Automatic Building Extraction From High-Resolution Optical Remote Sensing Imagery Using Deep Learning

Automatic extraction of buildings is an important research direction for high-resolution remote sensing image understanding and target recognition.It has been widely-used in the field of digital city,military reconnaissance,and disaster evaluation.Automatic,intelligent,reliable and accurate remote sensing image building extraction is of great practical value and significance for basic geographic data acquisition and updating.In recent years,the rapid development of high-resolution remote sensing images has increased the spectral,structural and texture information of the ground objects.However,it has led to many problems such as foreign body with spectrum and image noise at the same time.Especially in large-scale images,it is difficult to extract buildings with high precision due to the wider size of the buildings in the real city scene and the complex color textures.At present,building extraction based on remote sensing image mainly adopts image processing or traditional machine learning methods,which requires a lot of prior knowledge of professionals.Moreover,most of the designed algorithms are only applicable to specific scenes,and the scalability cannot be achieved.Deep-learning technology has developed rapidly in recent years,especially in the field natural images dataset processing.However,related technologies are still at the initial stage of target detection in remote sensing processing.In addition,the performance of existing deep learning algorithms in processing large-scale and complex scene images still needs to be improved.Therefore,this thesis uses the idea of semantic segmentation to study a building extraction method with strong robustness and low computational complexity for arbitrarily complex real-world scenes.Such method was verified in Google Earth's open source large-scale public dataset as well.The main contents and conclusions are as follows:This study first presents an end-to-end multi-level converged fully convolutional network structure,called Dilated Convolutional Networks(DCNs).The network effectively synthesizes the different levels of semantic information and position information in the images by using dilated convolution method.It constructs a series of different sizes of receptive fields to capture spatial context information of different scales,and improve the extraction of building characteristics and learning abilities.In addition,the network structure can make any size of the images uncut or deformed as network input,and get the final prediction result directly,which is not only easy to use,but also saves computing time.In the performance comparison test,rationality of the network structure has been proved.And compared with the existing mainstream deep learning methods,the algorithm has great advantages in accuracy,robustness and efficiency.Besides,in order to solve the shortages of training samples caused by the lack of public remote sensing annotated datasets,this study propose a model called Conditional Segmentation Generative Adversarial Networks(CSGAN),which is suitable for semantic segmentation of remote sensing images based on the idea of generative adversarial networks.This model adopts deep convolutional semantic segmentation as a generation model,and calculates the regression loss on a pixel-by-pixel basis based on the label prediction probability map generated by the input images.As a global loss statistics method that provides autonomous learning for generation model,the discriminant model performs higher-order statistics on differences and achieves data augmentation under unsupervised conditions.It has been verified that this algorithm can ensure high building extraction accuracy and maintain the integrity of objects under limited training samples.Furthermore,comparing with non-adversarial networks,this model can effectively reduce over-algorithm and over-fitting phenomenon.