Building Edges From High Spatial Resolution Remote Sensing Imagery Using Richer Convolution Features Network

Buildings are the prime elements of urban landscape and one of the most important and frequently updated parts of urban geographic database.Fast and accurate extraction of building edges from high spatial resolution remote sensing images is the basic requirement for detecting building changes.It is also of great significance for urban information updating,urban planning and management.In recent years,the development of deep learning provides a new technology for fast and accurate extraction of building edges from high-resolution remote sensing images.In this paper,Richer Convolutional Features(RCF)network is introduced into the field of high spatial resolution remote sensing image processing and analysis for the first time.In order to improve the generalization ability of network and the accuracy of building edge extraction,three aspects of data set construction,network retraining and edge probability map post-processing are studied.In addition,this paper starts with the evolution of RCF network,and deeply analyses the contribution of network structure changes to high-level feature information extraction.In this paper,a data set for building edge detection is constructed based on the most peripheral constraint extraction algorithm,and the RCF network is trained again.The trained RCF-building network can extract the probability map of building edge from high resolution remote sensing images.In the post-processing of the probability map of building edge,a new method is proposed based on the principle of geometric morphology analysis.The edge probability map thinning algorithm realizes the fine extraction of building edges.The experimental results show that the data set constructed in this paper can produce a better extraction model.The final experimental results show that the RCF-building edge detection learner based on RCF network constructed in this paper can extract the building edges pertinently,and can maintain a high level of building edge extraction.The accuracy of this method is at least 5% higher than that of other three typical networks.In addition,compared with the non-maximum suppression thinning algorithm provided by BSDS500 database benchmark evaluation,the proposed edge probability map thinning algorithm improves the accuracy and recall of edge feature information,and has better visual effect.Starting with the structure of RCF network,this paper qualitatively and quantitatively discusses the influence of its unique mixed output mode on edge extraction,which provides a new idea for future network improvement.In this paper,the best data sets,models and post-processing methods are combined to extract building edges from high-resolution remote sensing images,and thematic maps of building edges are generated.The comparison between thematic maps and ground truth shows that the experimental results in this paper are practical.