Hi-tech tree accounting: 3D models, LiDAR & machine learning

By on 21 September, 2016

A subset of LiDAR scan over Barmah-Millewa forest, left to right: (1) forest height in grey shades, (2) individual tree health (green – healthy, orange – declining, red – dead) (3) individual trees indicated by different colours.


The past sixty years have seen sizeable areas of forest in Australia experience dieback, driven mostly by drought and high temperatures. In order to develop initiatives to restore these valuable natural resources back to health, there is a need to quantify not only the distribution of trees, but also specific tree characteristics such as health, species and size.

This 2-part article was written by Iurii Shendryk and originally published in the Aug/Sep 2016 Issue of Position magazine.

Iurii Shendryk, a PhD student from the University of New South Wales (UNSW) has found that LiDAR technology offers an ideal method for quantifying such comprehensive characteristics of forests.

As part of the Geospatial Analysis for Environmental Change (GAEC) lab at UNSW, Shendryk was tasked with developing such an accounting system of trees and set out for the Murray River to do just that.

There, in the Barmah-Millewa forest near the border between Victoria and New South Wales, is one of the largest colonies of River Red Gums, themselves one of the largest Eucalypt species. Shendryk and the GAEC lab developed algorithms to accurately classify the health of these individual trees and investigated the spatial relation of the forest’s health to flooding events in the area.

Characterising forests from the air

One of the most prominent remote sensing tools used in forest studies at present is airborne LiDAR, which represents a perfect tool for individual tree extraction. The type and density of forest as well as the algorithm used, tends to dictate the success of individual tree extraction from LiDAR scans.

The majority of existing algorithms use top-down algorithms and work best for trees with a distinct top. This is usually effective for coniferous, cone-shaped trees such as pines, however most broad-leaf trees including eucalypts are asymmetrically-shaped and often have complex structures, thus requiring a bottom-up algorithm for segmentation.

To get stories like this delivered to your mailbox every week, subscribe to our weekly newsletter. 

Therefore GAEC team proposed a bottom-up algorithm for extraction of individual eucalypt trees. It is a stepwise procedure, whereby first tree trunks are identified based on the spatial arrangement of points in the lowest part of the LiDAR point cloud. Secondly, the points representing individual tree trunks are labelled within the point cloud. Thirdly, a 3D model is built by connecting all points within a certain radius. Finally, based on the spatial connectivity of points to the labelled ones using a so-called `random walks algorithm’, the point cloud is segmented into individual trees.

The detailed description of Shendryk’s use of the ‘random walks algorithm’ was recently published in Remote Sensing of Environment (2016).


Individual tree segmentation procedure.


The application of such an algorithm was possible thanks to high LiDAR point densities of 10-40 points per square metre, and was evaluated against an extensive field-measured dataset of tree positions. The LiDAR dataset used in this study was acquired in a full-waveform mode using a RIEGL LMS-Q560 system by Airborne Research Australia (ARA), and conducted along 17 flight lines covering 103 square kilometres, some 14% of the Barmah-Millewa forest.

The ground measurements consisted of more than 1,000 tree position readings using a Leica Viva GS08 Plus RTK GNSS rover, which were essential for optimisation and evaluation of the segmentation algorithm. The differential capabilities of GNSS were necessary to achieve sub-metre accuracies of tree positions in a forest with density of up to 700 trees per hectare. Overall, Shendryk’s algorithm allowed correct segmentation of 68% of the trees in this complex forest site. The proposed algorithm is expected to produce even higher accuracies when applied to LiDAR scans collected over forest plantations or coniferous species.

This is the first instalment in a 2-part exclusive article. Part-2 will be published in the coming weeks and will explore what Iurrii’s survey was able to find about the health of the trees in the area and the wider potential for natural resources in general. Subscribe to our free weekly newsletter to make sure you catch part-2.


You may also like to read:



Sign up now to stay up to date about all the news from Spatial Source. You will get a newsletter every week with the latest news.

How spatial data saves time and money
MetroMap gives access to super high-resolution and accurate ...
Bennett and Bennett expands service capabilities
Bennett and Bennett selects NavVis VLX mobile scanning and d...
Q&A with FIG President, Rudolf Staiger
Staiger reflects on his four years at the helm of FIG and th...
CASA releases RPAS roadmap for the next decade
The roadmap sets out the Authority’s intended approach to ...
OGC joins new Metaverse Standards Forum
The Forum aims to foster open standards collaboration betwee...