| In recent decades,China has continuously increased its infrastructure construction and its road network system has expanded rapidly.In order to achieve scientific decision-making for national economic and social development,it is urgent to obtain timely information on the distribution of road networks.The use of ultra-highresolution images to extract urban,suburban and rural roads has important application value.However,extracting roads from high-resolution images usually involve two difficulties.The first is to maintain the integrity of the road surface and the straightness of the sideline.The large tree canopies and tall buildings on both sides of the road have a shadow effect,often covering the road surface,causing difficulties in road extraction.The second is to maintain the connectivity of the road network in the area,so that the road network is not missing and uninterrupted.There are still challenges in effectively extracting the road area shaded by roadside tree canopies or high-rise buildings to maintain the integrity of the extracted road area,the smoothness of the edges,and the connectivity of the road network.In the past,it was difficult to solve this problem in pixel-based analysis and object-based analysis.In response to the above problems,this paper embeds two attention modules in the Dense UNet framework,namely global attention and core attention modules,and constructs a new semantic segmentation network model,namely the Cascaded Attention Dense UNet Semantic Segmentation Model(Cascaded Attention Dense UNet,hereinafter referred to as CADUNet).First,a group of cascaded global attention modules are introduced to obtain global road information,thereby improving road connectivity and road integrity under the shadow of tree canopies and high-rise buildings.Secondly,a set of cascaded core attention modules are embedded to ensure that road information is transmitted to the greatest extent in dense blocks in the network,and further assist the global attention module to obtain more effective road information,thereby improving the connectivity of the road network.In addition,the integrity of the road areas blocked by tree canopies and high-rise buildings,internal roads in parking lots,and transportation hubs can also be restored.On the basis of binary cross entropy,an adaptive loss function is proposed for the adjustment of network parameters.This article uses two sets of public remote sensing image data sets as the research objects,the Massachusetts road data set and the Deep Globe-CVPR 2018 road data set(hereinafter referred to as CVPR data set),and experiments are carried out with the support of the hardware and software of the supercomputer center.Finally,the research conclusions of this article are as follows:(1)It is concluded from various qualitative analysis and comparison that the model in this paper is suitable for road extraction in rural,suburban,and urban areas,and its performance is better than other comparison models.At the same time,the model in this paper can effectively improve the integrity of the road surface,the straightness of the sideline and the connectivity of the road network,and the obtained road extraction results are less interrupted.The model in this paper can better solve the problem of shadows blocking roads such as tree canopy and high-rise buildings.(2)In the quantitative analysis results of the two groups of public remote sensing image road data,the model in this paper is better than other models.In the Massachusetts data set,CADUNet’s F1 score and Io U reached 77.89% and 64.12%,respectively.CADUNet’s F1 score and Io U in the CVPR data set reached 76.28% and62.08%,respectively.Generally,the performance of CADUNet in the CVPR data set is inferior to the performance in the Massachusetts dataset.This may be due to the fact that the CVPR data set has more annotation errors and higher image spatial resolution,which makes learning more difficult.From the results,the total accuracy of several models generally reaches a higher value,and there is little difference between each other.The total accuracy has limited evaluation effect on the road extraction accuracy,and the F1 score and Io U combine the precision and recall rates.Therefore,the relatively effective index for the accuracy evaluation of road extraction is the F1 score and Io U.(3)In this paper,the visual accuracy of the performance of the extraction results on multiple road types is evaluated,and it is found that the method in this paper is advanced in the extraction results of multiple road types,but in practical applications it will be affected by the quality of the training data set,it is impossible to accurately identify some rural,suburban,and urban roads.(4)In the results of the model migration experiment of World View-II and World View-III images in Zhongwei City,CADUNet trained based on two public data sets has a certain generalization ability,but the generalization ability of the model differs greatly in images with different spatial resolutions.In the World View image migration experiment without image fusion,it can be found that when the image spatial resolution is lower than the spatial resolution of the model training data,the overall road extraction effect of the two models is not good,the generalization ability is weak,and both exist.Each has its own advantages and disadvantages.CADUNet trained on the Massachusetts data set performs better on slender roads,while CADUNet trained on the CVPR data set performs better on large roads.In the World View image migration experiment after image fusion,the model migration result of CADUNet trained based on the CVPR data set can show that the model migration effect is better when the spatial resolution of the model training data is consistent with the model migration data.The generalization ability is strong and the road surface integrity is high,but the result is still room for improvement.The CADUNet results trained based on the Massachusetts data set show that when the spatial resolution of the model training data is inconsistent with the model migration data,the model’s road extraction effect is poor,and there are a lot of "missing extraction" phenomena,and the integrity of the extracted road surface is not high.(5)Through the visualization method of the model,we have further understood the working process of the model in this paper to solve the problem of canopy shading effect.In model visualization,the feature map is output layer by layer from the encoder to the decoder.By observing the feature map results of each layer,it can be found that the low-level shape and texture features of the obscured road have been learned in the first dense block of the encoder stage.The second and third dense blocks continue to learn and get the feature map of the road.The last dense block presents high-level semantic features,which cannot directly interpret the meaning of the feature map.In the decoder stage,starting from the second cascaded attention module,you can already see that the obscured road appears in the feature map after the attention module.As the decoder deepens the interpretation of the feature graphic,the polarization between the road and the surrounding background information becomes more and more obvious.It can be seen from the final prediction graph that the extraction effect of the covered road is better. |
PDF Full Text Request