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, 1855–1859, 2019
https://doi.org/10.5194/isprs-archives-XLII-2-W13-1855-2019
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLII-2/W13, 1855–1859, 2019
https://doi.org/10.5194/isprs-archives-XLII-2-W13-1855-2019

  05 Jun 2019

05 Jun 2019

CLASSIFICATION OF TREE SPECIES ON THE BASIS OF TREE BARK TEXTURE

L. Ganschow1, T. Thiele1, N. Deckers2, and R. Reulke2 L. Ganschow et al.
  • 1VINS 3D GmbH, Wegedornstrasse 32, 12524 Berlin, Germany
  • 2Humboldt-Universität zu Berlin, Unter den Linden 6, 10099 Berlin, Germany

Keywords: Convolutional Neural Network, Image Segmentation, Forest Inventory, Pretraining and Finetuning, Integrated Positioning System, Plant Classification

Abstract. Forest inventory is an important topic in forestry and a digital solution which works on the basis of tree images is looked for. Implementing a system which automatically classifies tree species is the overall goal. In this paper the implementation of a convolutional neural net for solving this classification problem is executed and evaluated. The objective is creating a system which works well on unseen data and deriving guidelines and constraints to guarantee good accuracy results. Images including tree segmentation and the corresponding labels are provided as training data. The tree species classification takes the segmentation results of a stereo vision based image segmentation algorithm as input. The basic idea consists of cropping the tree images into quadratic boxes before feeding them into the neural net. First, each box is classified separately and then the results are evaluated to get a classification for the whole tree. Methods for result improvement include altering box size, using overlapping boxes, artificially enlarging the training set, pretraining and finetuning. Cropping a tree image into boxes of a specific size and accumulating the single results to get a classification of the whole tree leads to an accuracy of 96.7% provided that specific constraints like minimum box number and the projected size of the tree on image plane are considered. Finally, ways to further improve performance are pointed out.