Road Centerline Extraction And Weakly-Supervised Road Segmentation In High Resolution Remote Sensing Images

High-resolution remote sensing images acquired by airborne or spaceborne sensors are the main resources for surface observation,environmental monitoring,and surveying and mapping,etc.How to quickly and accurately extract the objects of interest from the images is the key issue in remote sensing image interpretation.Roads are important man-made objects,which are widely used in many aspects of social life,such as vehicle navigation,traffic management,map updating,geological disaster emergency,humanitarian assistance,etc.High-resolution remote sensing images are the main resources for road extraction,but the occlusion of vehicles,trees and buildings makes automatic road extraction very difficult.How to improve the efficiency and accuracy of road extraction is a current hot research topic.This article focuses on how to reduce the degree of manual intervention,improve the efficiency of road tracking and reduce the training cost of road segmentation model in the road extraction process.The main innovations are listed as follows:(1).This paper presents an improved semi-automatic road tracking algorithm based on adaptive circular template matching.The algorithm automatically calculates the optimal circular template radius,improves the template matching rules,reduces the number of algorithm parameters,and finally reduces the cost of manual intervention and improves the efficiency of road centerline tracking.In this paper,the radius of the circular template is estimated automatically using enhanced morphological gradient map.In the process of searching road center points,the angles between the road center points are employed,which makes the algorithm have better anti-noise ability.The algorithm has fewer parameters,and the meanings of the parameters are more intuitive.It can effectively solve the noise cast by trees,vehicles and buildings with simple parameter adjustment which improves the adaptability and practicability of the algorithm.(2).We propose a model to automatically identify the road center points using deep convolutional neural network(DCNN)and designed a tracking method named Deep Window to make the road extraction process fully automatic,which greatly improves the efficiency of road tracking.Deep Window uses a sliding window based on the DCNN decision guidance to directly extract the road network from the image without prior road segmentation.In this algorithm pipeline,the patch-based DCNN model is not only used for the search of the road center points,but also for the road center point detection during the tracking process.Furthermore,we propose an algorithm for estimating the road direction in a patch using the Fourier spectrum to determine the initial tracking direction for the tracking algorithm.The DCNN model is trained by point annotations which greatly reduces the training cost comparing to the fully-supervised models.The road extraction results are similar to those of the latest fully-supervised algorithms.At the same time,experiments show that the DCNN model is highly robust to the accuracy of point annotations,which makes the algorithm more applicable.(3).A weakly-supervised road segmentation method based on point annotations is proposed in this paper,and achieves performances similar to those of the fullysupervised algorithm.The method integrates deep neural network model and heuristic algorithms(such as support vector classifier,multi-scale and multi-directional Gabor filter,active contour model etc.)to achieve pretty good results.The powerful feature representation of deep neural network and the interpretability of the heuristic algorithms ensure the effect of our algorithm.This method indicates a novel methodology for road segmentation in remote sensing images,which is to train the model by pointing out the roads instead of marking all the road areas,just like the way people teach children.(4).Finally,we designed and released a patch-based road center annotation tool which can be used to manually label road center points as well as background patches.The tool outputs image patches and the corresponding road center Gaussian masks.The tool can also be used to mark the center point of any linear or strip objects.In addition,we also released a manually-sampled road center data set with more than 280 thousand samples.This work provides tool and dataset for researchers on the related issues.