Volume XLII-1
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLII-1, 217-224, 2018
https://doi.org/10.5194/isprs-archives-XLII-1-217-2018
© Author(s) 2018. This work is distributed under
the Creative Commons Attribution 4.0 License.
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLII-1, 217-224, 2018
https://doi.org/10.5194/isprs-archives-XLII-1-217-2018
© Author(s) 2018. This work is distributed under
the Creative Commons Attribution 4.0 License.

  26 Sep 2018

26 Sep 2018

CLASSIFICATION AND REPRESENTATION OF COMMONLY USED ROOFING MATERIAL USING MULTISENSORIAL AERIAL DATA

R. Ilehag1,3, D. Bulatov2,3, P. Helmholz3, and D. Belton3 R. Ilehag et al.
  • 1Institute of Photogrammetry and Remote Sensing, Karlsruhe Institute of Technology, Germany
  • 2Fraunhofer IOSB, Ettlingen, Germany
  • 3Department of Spatial Sciences, Curtin University, Perth, WA, Australia

Keywords: Multispectral, Thermal, High-resolution RGB, LiDAR, Building outlines, Classification, Image segmentation

Abstract. As more cities are starting to experience the urban heat islands effect, knowledge about the energy emitted from building roofs is of primary importance. Since this energy depends both on roof orientations and materials, we tackled both issues by analysing sensor data from multispectral, thermal infrared, high-resolution RGB, and airborne laser datasets (each with different spatial resolutions) of a council in Perth, Australia. To localise the roofs, we acquired building outlines that had to be updated using the normalised digital surface model, the NDVI and the planarity. Then, we computed a semantic 3D model of the study area, with roof detail analysis being a particular focus. The main objective of this study, however, was to classify three commonly used roofing materials: Cement tiles, Colorbond and Zincalume by combining the multispectral and thermal infrared image bands while the high-resolution RGB dataset was used to provide additional information about the roof texture. Three types of image segmentation approaches were evaluated to assess any differences while performing the material classification; pixel-wise, superpixel-wise and building-wise image segmentation. Due to the limited amount of labelled data, we extended the dataset by labelling data ourselves and merged Colorbond and Zincalume into one separate class. The supervised classifier Random Forest was applied to all reasonable configurations of segmentation kinds, numbers of classes, and finally, keeping track of the added value of principal component analysis.