Research On Feature Extraction And Shadow Information Restoration Of Urban Areas Based On High Resolution Optical Remote Sensing Image

Remote sensing images have long been an important means for urban analysis.Especially in the past ten years,satellite technology has developed rapidly around the world.Satellites equipped with high-resolution optical imagers have been launched one after another,which has greatly promoted the development of urban remote sensing.Currently,urbans are expanding rapidly and road networks are renewed and iterated quickly.The use of high-resolution remote sensing images can quickly and efficiently extract road networks,which is of great significance to urban planning,smart city construction,traffic management,and disaster emergency analysis.The launch of sub-meter optical satellites makes it possible to use optical remote sensing images for road network analysis in urban areas.However,the current research on the extraction of the landcovers in urban areas from high-resolution optical images is still in the development stage,and there are several problems that need to be solved urgently:(1)The interference of the same object of different spectrum and the same spectrum of different objects.Such as roads from different regions may be paved with different materials like cement,asphalt or road bricks.Thus,different roads cannot be extracted using the same spectral information.In addition,building roofs and roads are likely to be composed of the same material,which can easily interfere with road extraction.(2)The occlusion of buildings,vegetation and shadows.There are dense buildings in urban areas,and roads and buildings are built adjacent to each other.In remote sensing images,buildings,street trees and shadows will block the road to a certain extent,causing the incompletion of road extraction.Among them,building shadows are widespread in remote sensing images on a large scale,causing great interference to urban remote sensing.If the shadow can be extracted and the information of shaded region can be accurately compensated in the preprocessing stage of remote sensing image,it will greatly promote the development of urban remote sensing.At present,there are several obvious problems in the research of shadows from remote sensing images:(1)The characteristics of shadows under different lighting conditions are very different.The same shadow extraction strategy can hardly meet the shadow extraction in different environments,and the shadows are very vulnerable to water bodies.(2)In the shadow information compensation process,where the information and what strategy are used to compensate the loss information has not been well resolved;(3)At present,for ordinary images with training data,Generative Adversarial Network(GAN)can generate deshadow images well.But remote sensing images cannot obtain datasets with/without shadows at the same location,resulting in deep learning models cannot been well applied in remote sensing image shadow information recovery.In response to the above problems,we proposed a new strategy in road extraction to solve the problem of road breaks.In shadow removal,we combined with the multi-spectral information from remote sensing image and illumination model to extract shadows and compensate the information in shaded region.In addition,we also applied the GAN to train the existing shadow datasets and used it for shadow removal of UAV images.The work of this paper includes the following four parts:1.Traditional road extraction algorithms,which focus on improving the accuracy of road surfaces,cannot overcome the interference of shelter caused by vegetation,buildings,and shadows.In this paper,we extract the roads via road centerline extraction,road width extraction,broken centerline connection,and road reconstruction.We used a multiscale segmentation algorithm to segment the original images,and applied feature extraction to get the initial roads.Then,the Fast Marching Method(FMM)algorithm is employed to obtain the boundary distance field and the source distance field,and the branch backing-tracking method is applied to acquire the initial centerlines.Road width of each initial centerline is calculated by combining the boundary distance fields,before Tensor Voting algorithm is applied for connecting the broken centerlines to gain the final centerlines.Finally,the final centerlines are matched with their corresponding road widths,and the final roads are reconstructed.2.Under different illumination conditions,the spectral characteristics of shadows have a large dynamic range.It is difficult to extract shadows under different illumination environments with the same criterion,and shadows are easily disturbed by water bodies.To solve the above problems,we used the ratio of direct light intensity and ambient light intensity to define the shadow intensity,and applied the Fractal Net Evolution Approach(FNEA)to weaken the influence of high-reflectance landcovers in the shaded region.Then use the Normalized Difference Water Index(NDWI)and Near Infrared(NIR)bands to enhance shadows and eliminate water interference.3.Aiming at the problem that the existing shadow information recovery algorithms cannot accurately recover shadows from remote sensing images with complex background,we proposed an algorithm that used the nearest neighbor information to compensate the shadow information based on the illumination model.In shadow detection,an improved shadow index(ISI)is proposed.ISI enhances shadow features by combining NIR bands and YCb Cr space,and then reconstructs ISI through segmented objects obtained from the mean-shift algorithm.The reconstruction of ISI can greatly reduce the interference of high-reflectance landcovers in the shadowed region and improve the integrity of the shadow edge.In shadow compensation,we proposed an object-based method which always use adjacent non-shadow objects to compensate for shadow objects.In addition,we also proposed a Dynamic Penumbra Compensation Method(DPCM)to define the range of penumbra and compensate penumbra accurately.4.For the problem that remote sensing images cannot obtain shadowless labels,the GAN(Generative Adversarial Network)network is used to train ordinary image datasets containing non-shadow labels,and the obtained model is used for shadow removal of UAV remote sensing images.On the basis of the existing shadow datasets,we train two GAN networks to generate shadow mask images and non-shadow images.First,the shadow mask images are generated by the first GAN,and the shadow masks combined with the shadow image are used as the inputs of the second GAN to generate non-shadow shadow images.For the above training,we save the optimal parameters of GAN for shadow removal of UAV remote sensing images.In summary,we optimized the existing road extraction algorithms that cause incomplete road extraction when the road is occluded,and in-depth study of the extraction methods of optical remote sensing images in urban areas under different illumination conditions and the information compensate methods in complex backgrounds.We also simultaneously explored the GAN-based methods in generating non-shadow UAV remote sensing images.Our work is of great significance for the use of optical remote sensing images to analyze the ground features in urban areas,especially the study of shadow information compensation,which is an important prerequisite for remote sensing images to be used in urban remote sensing applications.