The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Download
Publications Copernicus
Download
Citation
Articles | Volume XLII-2/W13
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLII-2/W13, 43–50, 2019
https://doi.org/10.5194/isprs-archives-XLII-2-W13-43-2019
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLII-2/W13, 43–50, 2019
https://doi.org/10.5194/isprs-archives-XLII-2-W13-43-2019

  04 Jun 2019

04 Jun 2019

LOCAL VERSUS GLOBAL VARIATIONAL APPROACHES TO ENHANCE WATERSHED TRANSFORMATION BASED INDIVIDUAL TREE CROWN SEGMENTATION OF DIGITAL SURFACE MODELS FROM 3K OPTICAL IMAGERY

C. Kempf1,2, J. Tian2, F. Kurz2, P. d’Angelo2, and P. Reinartz2 C. Kempf et al.
  • 1TUM, Department of Ecology and Ecosystem Management, Freising, Germany
  • 2DLR, Department of Photogrammetry and Image Analysis, Oberpfaffenhofen, Germany

Keywords: segmentation, individual tree crown delineation, aerial stereo imagery, digital surface models

Abstract. Detection and delineation of forest trees in airborne observational data has been under study for decades, starting with images. With the advent of 3D point cloud generation techniques, much research has been spent for point cloud segmentation. From a cost perspective, aerial images are still advantageous. In this paper, two individual tree crown segmentation approaches for digital surface models are compared. Both methods attempt to enhance the drawbacks of watershed segmentation in unmanaged forests by applying a variational technique, locally to a watershed segment or globally to the image, respectively. The preprocessing by means of local histogram equalization that is necessary to harness the globally applied technique simultaneously improves the performance of the feature detection, while resulting boundaries are distorted. In contrast, the approach that uses the locally applied technique does not perform local histogram equalization prior to feature detection. It produces better localized boundaries in cases where detection is correct, but has a significantly lower rate of detection.