University of Twente Student Theses


Planar roof structure extraction from Very High-Resolution aerial images and Digital Surface Models using deep learning

Kenzhebay, Meruyert (2022) Planar roof structure extraction from Very High-Resolution aerial images and Digital Surface Models using deep learning.

[img] PDF
Abstract:Roof structure reconstruction is one of the more recent and active research directions in urban-related studies. Roof geometry information is needed for the generation of 3D models, which are used for applications such as solar potential estimation and telecommunication installation planning, wind flow simulations for pollutant diffusion analysis, etc. Given the advance in remote sensing technologies and the machine learning field, particularly deep learning, the prospects of deriving the roof structure information accurately and efficiently are promising. Many approaches for extracting roof structure have been proposed; however, there are still issues with output regularization, false detection and misclassification, and low computational efficiency, which leaves room for further improvement. In our study, we attempt to address these issues by proposing deep learning FCN-based methods for extracting roof structure from aerial imagery and Digital Surface Models (DSM) in the form of joined inner and outer rooflines directly in a regularized vector format. We develop and compare two roof structure extraction methods. The methodology and implementation details of both models are identical, with the exception that one of them has frame field learning branches for inner rooflines and outer rooflines. Frame field is a 4-D PolyVector field that helps to extract more regularized building boundaries with the correctly detected corners. The methodology is comprised of outer and inner rooflines segmentation, vectorization and post-processing. The approach was evaluated using pixel-level IoU metric and line-level PoLiS, PrecisionPoLiS≤0.5, RecallPoLiS≤0.5 and F-scorePoLiS≤0.5 metrics on both outer and inner rooflines. The experimental study area is the Stadsveld – ‘t Zwering neighbourhood of Enschede, Netherlands. According to our experiments, both models showed quite good performance in extracting building roof structures. The frame field learning model slightly outperformed the no-field model on inner rooflines segmentation with an IoU value of 0.35 and a little worse than the no-field model on outer rooflines, 0.37. However, the no-field model performed better than frame-field learning on PoLiS distance with values of 3,5 m and 1,2 m for outlines and inner rooflines, respectively. Besides, the no-field model scored higher on PoLiS-thresholded F-score for outlines and inner rooflines, having, 0.31 and 0.57 respectively. The no-field model produced better visual results, with straighter walls and fewer missed inner roofline detections. It can predict buildings with common walls thanks to the skeleton graph computation. To summarize, the frame field had little impact on the findings, and the proposed no-field method is suitable for urban applications and has the potential to be improved further. Keywords: image processing, image analysis, deep learning, roof structure extraction, roof vectorization, frame field learning
Item Type:Essay (Master)
Faculty:ITC: Faculty of Geo-information Science and Earth Observation
Subject:38 earth sciences, 50 technical science in general, 54 computer science, 56 civil engineering
Programme:Geoinformation Science and Earth Observation MSc (75014)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page