On December 13, 2021, Yao Sun defended her Ph.D. Thesis entitled "Large-scale LoD1 Building Model Reconstruction from a Single SAR Image".
Thesis: Large-scale LoD1 Building Model Reconstruction from a Single SAR Image
Abstract: Three-dimensional (3-D) building models are widely used in public and commercial sectors for environmental researches and location-based services. For the past three decades, 3-D building reconstruction has been a hot topic in remote sensing, however, there is limited information on building models on regional and global scales. Synthetic Aperture Radar (SAR) data have been employed for modeling buildings due to their imaging capability regardless of the time or weather conditions. In addition, complete global coverages of TerraSAR-X/TanDEM-X stripmap mode data have been acquired since 2012, providing great potential as a data source for global building reconstruction. However, building interpretation from SAR data is highly challenging. Due to the side-looking geometry and one-band radar sensors, urban structures are clearly visible in SAR images but are difficult to distinguish from each other. Although extensive research has been carried out on building reconstruction using SAR data, to date, no single study investigates large-scale building reconstruction from a single SAR image. This dissertation addresses large-scale Levels of Detail (LoD)-1 building models reconstruction from a single SAR image. Considering the characteristics of buildings in SAR images, building footprints are introduced as complementary data, and deep neural networks are employed for large-scale reconstruction. The work is developed in three stages. First, building footprints must be registered to SAR images for supporting SAR image interpretation. Therefore, a framework is developed that automatically registers building footprints to a corresponding SAR image. Second, the employment of deep learning methods requires training data. Therefore, an accurate Digital Elevation Model (DEM) is introduced to generate individual building areas in a SAR image, and a segmentation network is proposed for predicting building areas on a large scale. The extracted building segments are then employed for LoD1 model reconstruction. Third, to reconstruct buildings in larger areas, more training data are needed. However, accurate DEMs are unavailable in most cases. Therefore, the LoD1 building reconstruction problem is reformulated as a bounding box regression problem so that height data from multiple sources can be employed to generate bounding boxes of buildings. A regression network is proposed and examined for four study sites using both TerraSAR-X spotlight image and stripmap mode images. To the author's best knowledge, this is the first study investigating individual buildings in single SAR images on a large scale and the first study applying deep learning for individual building analysis using SAR images. The proposed algorithms have great potential to be applied on a regional and even global scale.