University of Natural Resources and Life Sciences, Vienna (BOKU) - Research portal

Go directly to the nagivation.
Go directly to the content.

Navigation/Search

Gewählte Master / Diploma Thesis:

Julia Maschler (2018): Tree Species Classification Using Hyperspectral and Laser Scanning Data.
Master / Diploma Thesis - Institut für Vermessung, Fernerkundung und Landinformation (IVFL), BOKU-Universität für Bodenkultur, pp 101. UB BOKU obvsg FullText

Data Source: ZID Abstracts

Abstract:: Knowledge of the distribution of tree species within a forest is key for multiple economic and ecological applications. It is usually acquired through time-consuming and thereby expensive field work. This thesis evaluates the suitability of two hyperspectral datasets (Visible and Near-Infrared (VNIR), shortwave infrared (SWIR)) combined with laser scanning data for an object-based Random Forest classification of 699 manually delineated and segmented tree crowns (13 tree species). Further, it assesses the automation potential of the classification by using two segmentation algorithms for delineating the tree crowns. The overall classification accuracy (OA) of the classification based on all three datasets and manual delineations was 93.7 %. Due to a position deviation between all datasets, only one hyperspectral dataset could be used for segmentation. The classification based on manual delineations on the VNIR data outperformed the corresponding SWIR-based classification (OA = 91.0 vs. 84.1 %), which is why VNIR bands were used as input for the segmentation. The subsequent classifications of data based on mean shift and watershed segmentation and VNIR-derived variables yielded high and similar results (OA = 89.4 v. 88.7 %). The Random Forest models from this step were applied to the whole study site. The classification reliability, i.e. the dominance of votes for the finally assigned class over all other class votes, was highest for Scots pine, followed by the highly abundant European beech. This thesis shows that highly accurate tree species classification can be carried out with hyperspectral data on their own and in combination with laser scanning data. Further, the results obtained for segmented tree crowns indicate a high automation potential of the method.; Beurteilende(r): Atzberger Clement; 1.Mitwirkender: Immitzer Markus