Extracting planar roof structures from very high resolution images using graph neural networks

Wufan Zhao*, C. Persello, A. Stein

*Corresponding author for this work

Research output: Contribution to journalArticleAcademicpeer-review

12 Citations (Scopus)
35 Downloads (Pure)


Roof structure information is essential for creating detailed 3D building models. These serve in many applications that require knowledge of the roof type and geometry. Automated extraction of the roof structure from remotely sensed images is a challenge because of scene complexity and the large variety of roof top configurations. This paper introduces a fast and parsimonious parsing method to extract vectorized building rooflines and structures from very high resolution remote sensing imagery in an end-to-end trainable way. Our Roof Structure Graph Neural Network (RSGNN) method consists of two components: 1) a Multi-task Learning Module (MLM) designed for geometric primitives extraction and matching, 2) a Graph Neural Network (GNN) based Relation Reasoning Module (RRM) to reconstruct the roofline structure. We tested RSGNN against state-of-the-art computer vision methods on two data sets: the vectorizing world building data set and a custom data set acquired in Enschede, The Netherlands. It provides an increase of 0.6/1.3 and 1.2/2.1 for msAP and FH on two data sets using only half of the training time. In addition, we performed a systematic ablation study to further validate its robustness. We conclude that RSGNN automatically generates the planar roof structure that outperforms competing models in both qualitative and quantitative evaluations.

Original languageEnglish
Pages (from-to)34-45
Number of pages12
JournalISPRS journal of photogrammetry and remote sensing
Publication statusPublished - May 2022


  • Deep learning
  • Graph neural network
  • Optical remote sensing image
  • Planar roof structure extraction
  • UT-Hybrid-D


Dive into the research topics of 'Extracting planar roof structures from very high resolution images using graph neural networks'. Together they form a unique fingerprint.

Cite this