Study On Building Change Detection Using Time-series UAV Images Combining Both Two And Three Dimensional Information

To comprehend and control promptly on the layout of urban regions,environmental changes,illegal construction,illegal land usage,ground greening,lake distribution and road layout is important for urbanization and sustainable development.Timely and accurate urban change detection(CD)is of great significance to economic construction,environmental monitoring.Changes in urban region are mainly the changes of buildings including construction or dismantlement in local area.The conventional on-site reconnaissance and observation methods are unable to meet the needs of urban planning regulators monitoring the rapid development of urban areas,especially urban fringe areas.In recent years,low-altitude remote sensing technology of unmanned aerial vehicles(UAV)has developed rapidly,which is cost effective,high flexibility,safe,high efficiency,and high spatial resolution.It is very suitable for dynamic monitoring hot spot area of rapid development and for providing relief of emergencies or disasters.It also can be used for rapid discovery of illegal land usage,automated digital city,geographic information updates,mining monitoring,engineering monitoring,and archaeology.UAV images have detailed structure information,the image texture is discontinuous,and spectral resolution is low.The spectra of different objects in UAV images usually overlaps with each other,making the difference of intra-class increase and the distinction level of inter-class decrease.In addition,the phenomenon of "same object different spectrum" and "same spectrum different object" is severe.Apply traditional two-dimensional CD methods for satellite images directly to UAV data usually cannot acquired desired CD results.Successful CD requires high relative co-registration accuracy of multi-temporal images,while the absolute position of images is of less concern.Because of the low flight height,small coverage,and the large number of UAV images,co-registration between multi-temporal UAV images is more difficult.Using only small number of ground control points(GCPs)to geo-reference UAV images using aerial triangulation bundle adjustment(BA)could achieve global co-registration of multi-temporal UAV images.However,the co-registration accuracy of images cannot meet the need for CD.Local image area around buildings may suffer huge misalignment.Even if dense GCPs are used in BA,there still be some deformation in local area.What's more,surveying large number of GCPs is time-inefficient and cost-ineffective,making it unsuitable for dynamic monitoring.Base on the analysis of conventional UAV image processing methods and the current three-dimensional building CD methods,this thesis proposed a combined two and three-dimensional object-oriented urban building CD method for time-series UAV images.The proposed method uses the united bundle adjustment(UBA)for time series UAV images spatial co-registration and the relative radiometric consistency processing method for spectrum consistency processing.After spatial and spectrum consistency preprocessing for time series UAV images,the three-dimensional elevation data of multi-temporal images of the survey area are acquired using dense image matching.Using an object-oriented change detection method,the images,the stereo model,and the matched point cloud are combined to obtain initial change candidates,then the combined two and three-dimensional object-oriented building CD method is conducted using three-dimensional elevation,edge,shape,spatial relations and other features to acquire CD results of the survey area.The main contents of this thesis are as follows:(1)Spatial and spectrum consistency processing for time series UAV imagesAn important prerequisite for image CD is that the images must have a uniform coordinate system.In order to obtain the uniform coordinate system of time-series UAV images in the same region,this thesis propose a new strategy to perform BA called united bundle adjustment for time-series UAV images.Time-series UAV images are matched with each other to obtain a unified tie points net.One BA process is performed on this unified tie points net,placing each image in the same coordinate system.Hence the orthophotos and dense matching point cloud will achieve high co-registration accuracy,which will meets the need for registration accuracy of CD.Spectral information plays an important role in CD and image classification,thus radiometric consistency processing is essential.Image scene is often complex in ultra-high resolution UAV image for it usually contain many distinguishable objects,and radiation difference between images usually does not have an overall linear correlation.Therefore,this thesis studies a novel relative radiometric consistency processing method using two-dimensional radiometric distribution feature,which is suitable for high-resolution remote sensing image change detection.(2)UAV image segmentation method using building height,edge,and shape characteristicsBecause of the low flying height,the ground sample distance of UAV images reaches centimeters.The sensor on-board UAV is usually a consumer-level digital camera with only RGB bands,thus the spectral resolution is low.The phenomenon"same object different spectrum" and "same spectrum different object" prevails.It is difficult to obtain desired results using a pixel-level CD method.Object-level methods require object-oriented image segmentation.It is important for object-oriented CD method that the image segments express the real information.Therefore,this thesis studies an image segmentation method suitable for UAV images based CD which takes the building characteristics into account.In addition,there are rich geometric features in UAV images.How to use these features has been a difficult problem in high resolution image processing.There are usually long straight edges and more regular geometric shapes for buildings,this thesis tries to extract and apply these features for CD.(3)A combined two and three-dimensional object-oriented building change detection methodAfter UBA,the exterior orientation parameters of all UAV images are within the same coordinate system.The 3D point cloud of different temporal images is obtained by the dense image matching and the digital surface model(DSM)is constructed.This thesis utilizes UAV imaging geometry information and stereo models and take into account the image spectrum and elevation information to obtain building construction,building dismantlement,and building height change information.Since the ground objects are very complex and there may exist matching errors in dense matching results,this thesis first detect huge buildings in survey area using height information and apply CD on them to acquire high-level-confidence results.Apply the combined two and three dimensional information CD on none-mosaicked ortho images and height ortho images to acquire initial changes,then eliminates "pseudo-change" using the spatial relations,the edge features,shape features,the spectral features,imaging geometry,the 3D point cloud/DSM extracted from the image,and two and three dimensional information cross-validation to obtain more accurate building change results.