Volume XLII-2/W11
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLII-2/W11, 975-980, 2019
https://doi.org/10.5194/isprs-archives-XLII-2-W11-975-2019
© Author(s) 2019. This work is distributed under
the Creative Commons Attribution 4.0 License.
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLII-2/W11, 975-980, 2019
https://doi.org/10.5194/isprs-archives-XLII-2-W11-975-2019
© Author(s) 2019. This work is distributed under
the Creative Commons Attribution 4.0 License.

  05 May 2019

05 May 2019

DETECTION OF BUILDING ROOFS AND FACADES FROM AERIAL LASER SCANNING DATA USING DEEP LEARNING

F. Pirotti1,2, C. Zanchetta3, M. Previtali4, and S. Della Torre4 F. Pirotti et al.
  • 1CIRGEO - Interdepartmental Research Center of Geomatics, University of Padova, Viale dell’Università 16, Legnaro (PD), Italy
  • 2TESAF Department, University of Padova, Viale dell’Università 16, Legnaro (PD), Italy
  • 3DICEA Department, University of Padova, Via Francesco Marzolo, 9, 35131 Padova, Italy
  • 4Department of Architecture, Built Environment and Construction Engineering, Politecnico di Milano, 20133 Milan, Italy

Keywords: deep learning, semantic classification, Tensorflow, laser scanning, lidar

Abstract. In this work we test the power of prediction of deep learning for detection of buildings from aerial laser scanner point cloud information. Automatic extraction of built features from remote sensing data is of extreme interest for many applications. In particular latest paradigms of 3D mapping of buildings, such as CityGML and BIM, can benefit from an initial determination of building geometries. In this work we used a LiDAR dataset of urban environment from the ISPRS benchmark on urban object detection. The dataset is labelled with eight classes, two were used for this investigation: roof and facades. The objective is to test how TensorFlow neural network for deep learning can predict these two classes. Results show that for “roof” and “facades” semantic classes respectively, recall is 84% and 76% and precision is 72% and 63%. The number and distribution of correct points well represent the geometry, thus allowing to use them as support for CityGML and BIM modelling. Further tuning of the hidden layers of the DL model will likely improve results and will be tested in future investigations.