Every day, it seems that scientists publish a new technology or algorithm. Artificial intelligence and machine learning are all powerful tools that can help people worldwide and speed up research itself. One might ask whether these approaches still require some human supervision.
One labour-intensive task for ecologists and conservationists around the world is to catalogue and monitor forest habitats. Terrestrial Laser Scanning (TLS) is becoming one of the mainstream approaches to take a 3D scan of an individual tree or forest patch to extract different measurements (e.g. number of trees, tree height, tree diameter). This computationally challenging approach is prone to high error rates as it relies on algorithms determining what a tree or a tree part is in a 3D point cloud (i.e. image).
Dr Olivier Martin-Ducup from the University of Montpellier and colleagues from the University of Yaoundé, Aalto University and Tampere University compared automatic and human-associated tree measurements in a one hectare plot in the Dja Faunal Reserve in Cameroon. The researchers found that when they isolated trees on the laser scans, they reduced the measurement errors by a factor of three to ten times.

In 2018, Martin-Ducup and colleagues established a 1 ha forest plot in the Dja Faunal Reserve. The team measured 391 trees and scanned the plot in a snaking pattern. The tree, Uapaca guineensis, was the most common species, whilst Irvingia grandifolia had the largest diameter (1.2 m).
The scientists tested five fully automated pipelines using different software, algorithms and settings. The three main steps were to first localise and isolate individual trees from the LIDAR-generated 3D point cloud, segment the leaves and the wood, and reconstruct trees using quantitative structural modelling.
The scientists tested if human assistance at the different processing stages might reduce the error rates at plot, sub-plot and tree-level. The tree height, diameter, crown area, basal area and wood volume were estimated and compared with the manual measurements.

At the tree scale, the researcher found that isolating trees using human assistance reduced the error in the wood volume by a factor of ten. At the 1-ha plot scale, locating trees with human assistance reduced the error by a factor of three.
“The results showed that all the pipelines returned poorly reliable results when performed fully automatically,” Martin-Ducup and colleagues write.
“This finding highlights the risk of blindly using these automated treatments at the plot scale. However, we demonstrated that providing human assistance, even limited assistance, in critical steps in the automated pipeline methods could greatly help reduce the estimation errors.”
LIDAR scanning, machine learning and artificial intelligence are all exciting new tools that can help scientists, conservationists and land managers. But telling trees in a 3D image can still be done better by a person.