Tampere University of Technology

TUTCRIS Research Portal

Tree species classification using structural features derived from terrestrial laser scanning

Research output: Contribution to journalArticleScientificpeer-review

Details

Original languageEnglish
Pages (from-to)170-181
Number of pages12
JournalISPRS Journal of Photogrammetry and Remote Sensing
Volume168
DOIs
Publication statusPublished - 1 Oct 2020
Publication typeA1 Journal article-refereed

Abstract

Fast and automated collection of forest data, such as species composition information, is required to support climate mitigation actions. Recently, there have been significant advances in the use of terrestrial laser scanning (TLS) instruments, which facilitate the capture of detailed forest structure. However, for tree species recognition the structural information from TLS has mainly been used to complement spectral information. TLS-only classification studies have been limited in size and diversity of plot forest types. In this paper, we investigate the potential of TLS for tree species classification. We used quantitative structure models to determine 17 structural tree features. These features were computed for 758 trees of five tree species, including two understory species, of a 1.4 hectare mixed deciduous forest plot. Three classification methods were compared: k-nearest neighbours, multinomial logistic regression and support vector machine. We assessed the potential underlying causes for structural differences with principal component analysis. We obtained classification success rates of approximately 80%, however, with producer accuracies for three of the five species ranging from 0 to 60%. Low producer accuracies were the result of a high intra- and low inter-species variability. These effects were, respectively, caused by a high size-dependency of the structural features and a convergence of structural traits across species as a result of the individual tree position in the forest canopy and shade tolerance. Nevertheless, the producer accuracies could be improved through sensitivity vs. specificity trade-offs, with over 50% for all species being obtainable. The high intra -and low inter-species variability complicate the classification. Furthermore, the classification performance and best classification method greatly depend on its targeted application. In conclusion, this study proves the added value of TLS for tree species classification but also shows that TLS opens up potential for testing and further development of ecological theory.

Keywords

  • Machine learning classifiers, Quantitative structure model, Structural tree features, Terrestrial laser scanning, Tree species classification

Publication forum classification

Field of science, Statistics Finland