Research On Road Extraction From High-resolution Remote Sensing Image With Deep Learning

Accurate road maps are of great importance in a wide range of applications.In the era of big data,the development of remote sensing technology has brought opportunities and challenges to road extraction.Due to the diversity of the road and background,traditional hand-craft features-based extraction methods are impractical in terms of accuracy and efficiency.In recent years,deep learning represented by convolutional neural networks has shown great potential in remote sensing applications,which offer a promising avenue for road extraction.In this article,we track the latest research of deep learning applied in remote sensing and computer vision at home and abroad;and we focus on extracting road from high-resolution remote sensing images based on deep learning,including fully supervised learning-based road semantic segmentation and topologic tracing as well as weakly supervised learning-based road surface extraction.First,considering the limitations of the mainstream road extraction methods,especially on topological connectivity and completeness,in this article,we propose a road extraction algorithm that combines the advantages of semantic segmentation and topologic tracing solutions to achieve simultaneously road surface and road centerline extraction.This algorithm contains three main steps: road boosting segmentation,road multiple starting points tracing,and the fusion process,which aims to improve the topological connectivity and completeness of road extraction.Regarding road segmentation,the proposed algorithm exceeds the state-of-the-art segmentation methods by 7% for road connectivity.With respect to road centerline extraction,the algorithm outperforms the state-of-the-art tracing methods by 40% for road completeness.Second,considering most existing road extraction methods employed fully supervised learning and relied on large amounts of manually annotated data.As the annotation of road surface data is time-consuming and labor-intensive,the scribbles(such as road centerlines)are more accessible.Inspired by the latest advances of weakly supervised learning in computer vision fields,in this article,we propose a weakly supervised deep learning algorithm for road surface extraction.By utilizing road attributes and super-pixel segmentation,the road segmentation proposal masks are generated from scribbles.Then,a two-branch neural network is trained with proposal masks and auxiliary edge information for road surface segmentation.This algorithm outperforms the classic scribble-supervised segmentation method by 20% for the intersection-over-union(Io U)indicator and also exceeds recent related research work by 4%.It can be an important step in automatic road extraction research as it learns from sparse scribbles without the need of densely labeled road surface annotations.These innovative algorithms proposed in this article are expected to promote intelligence and automation of road extraction,and would provide good prospects for theoretical research and practical applications.