University of Twente Student Theses

Login

Building outline delineation and roofline extraction: a deep learning approach

Golnia, Mina (2021) Building outline delineation and roofline extraction: a deep learning approach.

Link to full-text:https://ezproxy2.utwente.nl/login?url=https://library.itc.utwente.nl/login/2021/msc/upm/golnia.pdf
(only accessible for UT students and staff)
Abstract:Nowadays, many authorities are attempting to address complex urban and environmental issues through digital technology. Following the achievements of deep learning and the availability of remote sensing data, the interest in developing automatic and robust techniques to generate accurate and up-to-date building mapping models, which are fundamental for constructing 3D models or urban digital twins, is rapidly increasing. Deep learning proved helpful in recognising urban objects and structures and extracting the buildings' geometrical characteristics. Having all in mind, in this research, we propose a methodology to automatically extract the building rooflines, namely, Eave, Ridge and Hip lines, which are the prerequisites for 3D building models with Level Of Detail 2 (LOD2) using a CNN-based deep learning technique. Our strategy combines two stages; first, predicting a binary building mask that will be added as an input layer in the second stage- roofline extraction. In both stages, the Unet-Resnet network architecture with 51 and 101 layers are adopted and fine-tuned to find the optimal solutions. The proposed method is tested in Enschede, the Netherlands, using the 25cm orthorectified aerial (RGB) images in 2018. Both networks are also tested using the normalised Digital Surface Model (nDSM) to improve the results. Unet-Resnet 101 performs better in both stages, reaching an average F1-score (the harmonic mean between precision and recall) of 0.68 for binary building mask prediction and 0.55 for rooflines extraction. The results improve to 0.85 and 0.66 for binary building mask prediction and roofline extraction, respectively. A class-wise evaluation is also applied to clearly understand the model's behaviour for each class of rooflines. Accordingly, an average F1-score of 0.81, 0.55 and 0.32 is achieved for eave, ridge and hip lines, correspondingly. The precision (correctness) and recall (completeness) values for eave lines prediction do not deviate much (0.82 and 0.81). In contrast, the ridge and hip classes have a higher recall (0.61 ridges, 0.51 hips) than the precision value (0.49 ridge, 0.23 hip). Having predicted the lines, they are then simplified using the Douglas-Peucker simplification algorithm, with a tolerance of 0.5m. Regarding our investigation results, the proposed method can be effectively used to automatically extract building roof structures and linear elements, which can be generalised to any type of roof. Besides, the model is able to extract inner walls, which is a big challenge in the building segmentation field. However, it is recommended to use higher resolution images and a larger amount of training data with more variety in building types in future studies.
Item Type:Essay (Master)
Faculty:ITC: Faculty of Geo-information Science and Earth Observation
Programme:Geoinformation Science and Earth Observation MSc (75014)
Link to this item:https://purl.utwente.nl/essays/88990
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page