Change Detection Between Airborne Laser Scanning And Photogrammetric Data

As a research branch of change detection,3D change detection is of great significance to national economic development,environmental protection,and national defense.It is widely used in land use/land cover change detection,geographic information updating,and deformation monitoring of industrial products,environment and disaster monitoring,3D model updating,object tracking,vegetation growth monitoring and biomass assessment etc.In many cases,archived 3D geographic information data and newly acquired data are obtained by different technologies and exhibit different attributes.It is of great theoretical and practical significance to study the change detection techniques between 3D data of different modalities.However,the existing 3D change detection methods mainly aims at the same type of 3D data.When these methods are applied to multimodal 3D change detection,a large number of false alarms or omissions often occur.This thesis focuses on change detection between airborne laser scanning data and photogrammetric data.The research background and significance of multimodal 3D change detection are summarized.The related work of 3D data acquisition,representation and registration is reviewed.The related work of 3D semantic segmentation in remote sensing and computer vision is reviewed.The research challenges of 3D multimodal change detection are analyzed.The current methods and existing research gaps of 3D change detection are summarized.The main innovations are as follows:1.To investigate whether it is feasible to use photogrammetric point cloud as a replacement for laser scanning point cloud,a robust method for quantitative evaluation of photogrammetric point cloud and DSM is proposed.The quality of photogrammetric data is evaluated at the whole block level and local patch level.A large number of square patches are extracted from the ground points and non-ground points.The patch-based evaluation shows the overall distribution of dense matching errors in the study area.This paper also designs quality measures for assessing the accuracy and precision of photogrammetric data at the block level.The distribution of dense matching errors,the number of GCPs,the impact of GCP weights,the impact of oblique images on the dense matching errors,and the dense matching errors on different object types are studied.In addition,the distribution of gross errors in photogrammetric point clouds is also analyzed.This chapter provides practical recommendations and methods for photogrammetric quality control and quality evaluation.2.Aimed at detection building changes between laser scanning data and dense matching data,an unsupervised change detection method is designed.An adaptive object extraction method is applied to extract ground,roof and vegetation from laser scanning points.Considering the different properties between laser scanning data and photogrammetric data,an unsupervised method for building change detection is applied.Firstly,the objects are extracted from the laser scanning points.The heightened or demolished buildings are detected by comparing the roof segments with photogrammetric point clouds.Then,newly-built buildings are detected by comparing the dense matching points with laser scanning points in the remaining unknown areas.Experiments on laser scanning data and dense matching data from different epochs show that the proposed method can detect building changes effectively and achieves high accuracy.3.A Siamese neural network called“PSI-DC”is proposed to detect changes between multimodal point clouds.After obtaining the initial change positions,a method for delineating building change boundaries is designed.Firstly,point clouds and orthoimages of different modalities are converted into images.The images are then fed into light-weighted PSI-DC to quickly detect the initial change positions.The precise boundaries of changed buildings are then delineated by pixel-wise change detection.This“coarse to fine”method is not only robust to point cloud noise and scene complexity,but also generates heightened or lowered labels for each changed pixel.The proposed PSI-DC model is compared with the other Siamese architectures and a feedforward architecture,and the effects of different input configurations and different network structures on the change detection accuracy are compared.After obtaining initial change map,a post-processing method based on morphological operations is proposed to optimize the results at the minimum cost to obtain individual object-based building changes.Experiments are implemented on Rotterdam data of 15.4 km~2.Experimental results were quantitatively evaluated at the patch level,pixel level and object level,respectively to demonstrate the effectiveness.4.In order to achieve“end-to-end”change detection and avoid the pre-processing and post-processing steps,a deep learning model targeted for integrated semantic segmentation and change detection is proposed.The model takes PointNet++as the backbone network which takes features from laser scanning points and the neighboring dense matching points as the input.Considering the output,the laser point cloud is classified into six categories,which contain both the semantic labels of unchanged points,and the change label of the changed objects.The experiments are implemented on the Rotterdam dataset.The results show that the classification accuracy of PointNet++is better than that of PointNet as a whole.The classification accuracy of three types of unchanged classes(Terrain2Terrain,Building2Building and Vegetation2Vegetation)are better than the that of three changed classes(Heighted building,lowered building and other).