ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Download
Publications Copernicus
Download
Citation
Articles | Volume IV-2/W5
ISPRS Ann. Photogramm. Remote Sens. Spatial Inf. Sci., IV-2/W5, 103–110, 2019
https://doi.org/10.5194/isprs-annals-IV-2-W5-103-2019
ISPRS Ann. Photogramm. Remote Sens. Spatial Inf. Sci., IV-2/W5, 103–110, 2019
https://doi.org/10.5194/isprs-annals-IV-2-W5-103-2019

  29 May 2019

29 May 2019

EXTRACTION OF VINEYARD MACROSTRUCTURE FROM SUB-OPTIMAL SEQUENCES OF AERIAL IMAGERY

A. Finn, A. Melville-Smith, and R. Brinkworth A. Finn et al.
  • School of Engineering, University of South Australia, Mawson Lakes, South Australia, SA 5095

Keywords: unmanned aerial vehicle, 3D point cloud, structure from motion, vineyard

Abstract. Remote sensing techniques can be used to identify and classify vine properties such as row width, height, cover-fraction, missing segments and leaf area density, providing opportunities to visualise plant vigour as a spatial function of vineyard geography. This information may then be integrated into decision support tools to improve vineyard management practices. An algorithm for identifying vines from a sequence of overlapping aerial images and then estimating their properties is described. The image stacks were obtained from visible and long wave infrared cameras carried by an unmanned aerial vehicle (UAV). Structure from motion (SfM) was used to create 3D thermal and colourised point clouds, from which the underlying topography of the surface terrain was extracted. The surface topographic model was obtained using bounded data query nearest neighbour calculations, which were reduced to computationally manageable levels using Kd-trees that recursively partitioned the point clouds by progressively separating them into binary trees. This allowed the point clouds to be classified in terms of their hue, saturation, surface temperature and height relative to surface topography using Lloyd’s unsupervised k-means clustering. Individual samples were then associated using Gaussian probability density functions normalised by cluster statistics. The algorithm was evaluated against ground truth obtained using aerial data in terms of its accuracy and robustness using a combination of real world conditions that included high shadowing, poor contrast and UAV flight paths and camera settings that delivered sub-optimal SfM performance.