Automated Extraction Of Road Networks In Satellite Imagery Based On Convolutional Neural Networks

Automated extraction of road networks in satellite imagery has been a focused topic to be solved.Traditional road network extraction algorithms are usually feature-based.These feature-based algorithms are vulnerable to the complicated presence of roads and the contamination on the road area in the satellite images.In this paper,we propose an automatic deep learning-based road network extraction algorithm which is able to address the mentioned problems.The main part of our algorithm are as follows:Firstly,a preprocessing method using joint filtering is proposed.The smoothing method guided by prior information of road structures can not only effectively filter the noise,but also enhance the road feature information.This method makes solid conditions for the subsequent road network procedure.Secondly,a deep learning based road network extraction algorithm is proposed.On the first hand,the algorithm proposes an automatic sample extraction mechanism to get a large number of samples.Then,the samples is trained to get a model which can predict a certain region's probability of road.Next,the model and the super-pixel segmentation algorithm are used to rapidly get a road probability response map.Finally,an adaptive threshold segmentation is applied on the map and the filtered input to get the initial road networks.Third,repairing of road pollution area mechanism based on road network topology information is proposed.The mechanism uses the road network topology information to train a model that can detect such potential road areas.A potential road probability response map is constructed through the model.Graph cut is applied on the two maps to get the final road network.Finally,the method was compared with other world-class road detection methods and achieved high performance.