Research On Road Extraction Method Of High-Resolution Remote Sensing Imagery Based On Context Awareness And Feature Representation Reasoning

As a task of image analysis and understanding,road extraction from high-resolution remote sensing images plays an important role in many fields,such as urban planning,land management,traffic management,automatic navigation and so on.Deep convolutional neural network has gradually become the main method of road extraction due to its powerful feature representation ability.However,roads are generally blocked by vegetation,buildings,and shadows in remote sensing images,which often leads to incomplete and discontinuous roads in the road extraction results.To solve these problems,this paper proposes a method framework for road extraction from high-resolution remote sensing images.On the one hand,aiming at the problem of encoder-decoder structure in road extraction task,a feature encoder HR-Road Net suitable for road extraction task is proposed,which aims to retain complete narrow road information and rich spatial details.On the other hand,aiming at the problem that roads are partially occluded in remote sensing images,a road extraction network CA-HR-Road Net is proposed to improve the accuracy of road extraction and enhance the integrity and continuity of the extraction results.Finally,aiming at the problem that roads are completely occluded in remote sensing images,a post-processing network RIRNet is proposed to optimize and repair the initial extraction results.It can not only connect the fractured roads,but also eliminate errors such as omission and misclassification.Specifically,the following studies were carried out:(1)Road feature representation with using high-resolution network.Most of the existing encoder-decoder structures have problems such as loss of narrow road information and irrelevant noise information introduced by skip connections.Inspired by the principles of high-resolution network,this paper proposes an encoder for road extraction,namely HR-Road Net.The model contains three parallel branch streams with different resolutions,the high-resolution branch can retain intact narrow road information and rich spatial details,and the low-resolution branch can extract the effective deep semantic information.In addition,different branch streams interact with each other through feature fusion.The results of quantitative experiments show that,compared with the backbone networks such as Res Net and UNet,the proposed HR-Road Net can obtain the best accuracy of road extraction and more complete and smooth extraction results.It achieved 76.92% on F1 and 62.50% on Io U on the Massachusetts Dataset,75.46% on F1 and 60.59% on Io U on the Deep Globe Dataset,and 77.58%on F1 and 63.37% on Io U on the CH6-CUG Dataset.The results of qualitative experiments show that the proposed model can not only avoid the loss of narrow roads,but also eliminate irrelevant noise information.It is proved that the proposed HR-Road Net can be better applied to the road extraction task.(2)Road information extraction with coupling local and global context.As for the problem that roads are partially occluded in remote sensing images,this paper proposes a road extraction model CA-HR-Road Net.The model uses the HR-Road Net model as the main feature encoder and includes the Multi-scale Feature Representation Module and the Coordinate Attention Module.The Multi-scale Feature Representation Module combines multi-scale features with residual learning to take full advantage of multi-scale features,thereby efficiently extracting multi-scale local context information and modeling the dependencies between roads and background environments.The Coordinate Attention Module,as a lightweight global context awareness algorithm,can effectively capture the long-range dependencies in channel and spatial dimensions and model the correlation between different road objects.The results of experiments show that,compared with other related models,the proposed CA-HR-Road Net can obtain the highest extraction accuracy on different data sets.It achieved78.19% on F1 and 64.19% on Io U on the Massachusetts Dataset,76.79%on F1 and 62.33% on Io U on the Deep Globe Dataset,and 77.92% on F1 and 63.83% on Io U on the CH6-CUG Dataset.The results of qualitative experiments show that the proposed model can obtain road extraction results with better completeness and continuity.Therefore,the effectiveness and generalization of the CA-HR-Road Net are demonstrated.(3)Road network reasoning with considering direction-connectivity.As for the problem that roads are completely occluded in remote sensing images,this paper proposes a lightweight post-processing network RIRNet.The model is based on encoder-decoder structure and includes the Information Reasoning Module and the Road Direction Reasoning Task.The Information Reasoning Module can reason the spatial information relationship between different cows or columns from different directions to reason and repair the road fracture effectively.The Road Direction Reasoning Task indirectly enhances the reasoning ability of the postprocessing model and optimizes the initial road extraction result through the mutual constraint and promotion among the multi-tasks.The results of quantitative and qualitative experiments show that the proposed RIRNet can achieve excellent post-processing effect.It can achieve better accuracy improvement on all road extraction models,and can effectively repair the broken road segments,and deal with errors such as omission,misclassification and noise,which proves the effectiveness and generalization of the RIRNet model in post-processing optimization.At the same time,the proposed road extraction framework(CA-HR-Road Net +RIRNet)can obtain the best accuracy of road extraction on different data sets.It achieved 78.44% on F1 and 64.53% on Io U on the Massachusetts Dataset,77.68% on F1 and 63.51% on Io U on the Deep Globe Dataset,and78.22% on F1 and 64.23% on Io U on the CH6-CUG Dataset,which proves the excellence and generalization of the framework.