A Hybrid Strategy for Generating Digital Terrain Model From Dense Digital Surface Model Using Cloth Simulation Filter and Total Variation Regularization
View/ Open
Date
2022Author
Köseoğlu, Fitnat
xmlui.dri2xhtml.METS-1.0.item-emb
Acik erisimxmlui.mirage2.itemSummaryView.MetaData
Show full item recordAbstract
A Digital Elevation Model (DEM) is a general term used to express the virtual representation of the height values of the Earth's surface. Two of the most widely used DEM types are Digital Surface Model (DSM) and Digital Terrain Model (DTM). DSM is a 3-dimensional (3D) dataset that includes the vegetation, bare earth surface and the height values of man-made objects (buildings, bridges, vehicles, etc.) built on it. DTM contains only ground height values, unlike DSM. While DSM can mostly be obtained directly from LiDAR or image matching, DTM generation is a challenging process based on filtering the DSM data to remove objects other than the terrain surface. Since generating DTM requires high accuracy and precision for application areas such as precision farming, military applications, base mapping, and transportation systems, DSM filtering approaches are still being developed.
In this thesis, a DTM extraction methodology consisting of three steps has been developed. In the first stage, the cloth simulation filtering algorithm (CSF) is utilized. During that stage, the ground index information regarding the terrain pixels of the input DSM is obtained by using both fine resolution and coarse resolution clothes. Thereafter, two initial DTMs having fine resolution and coarse resolution are generated by interpolating the pixels labeled as terrain. In the second stage, the goal is the generation of an adaptive smoothing map that acts as a critical indicator for the variational optimization framework. After calculating the differences between the fine-resolution and coarse-resolution DTMs produced, a rotation invariant two-dimensional Gaussian kernel is applied to generate the adaptive smoothing map. Finally, a modified version of the variational DTM extraction method is developed to minimize the variational cost function iteratively and generate the final DTM. Instead of using a single smoothing level, the proposed method utilized different smoothing levels that change from pixel to pixel.
The accuracy measures utilized in this thesis are based on object-based and elevation-based accuracy measures. The accuracy assessment is performed on two separate test sites (in Wales, United Kingdom, and Ankara, Turkey), and both numerical and visual results for the high resolution dense DSM datasets (LiDAR-based and UAV-based) are given. Elevation-based accuracy analysis demonstrates that the proposed DTM extraction methodology is successful against the CSF method. The best RMSE and absolute errors are consistently achieved by the proposed approach. Object-based accuracy analysis showed that the proposed method achieves a nice balance between the object-based precision and recall measures.