Now, thanks to an international collaboration involving Scion, the University of Gottingen and the University of Applied Science and Arts (HAWK) in Germany and led by Dr Stefano Puliti at the Norwegian Institute of Bioeconomy Research (NIBIO), a benchmark dataset for forestry is publicly accessible.
Research papers associated with the two-year project have recently been published. However, the impact of the ‘For-Instance’ dataset extends far beyond the academic realm. More than simply a collection of labelled data, the dataset marks a significant turning point in 3D deep learning for forestry by providing researchers and AI specialists with the raw material they need to train and test their models on real-world forest data for the first time.
At a time when data is often considered the most valuable commodity, this dataset breaks down the barriers to entry for AI adoption in forestry. Instead of spending months collecting and labelling data themselves, data scientists at New Zealand companies and industry service providers now have access to a valuable resource that can accelerate their own innovative AI-driven solutions for forestry. Scion’s team lead for Remote Sensing and GIS, Grant Pearse, says the dataset is “game-changing” for New Zealand’s forestry industry as it will unlock a multitude of possibilities.
“We contributed labelled data from radiata pine to empower researchers and industry to develop and test models specifically for point clouds collected from our commercial radiata forests. The data are labelled in such a way that models can be trained to identify and segment individual trees as well as the stems, live branches and woody branches of each tree within the lidar point clouds – this directly aligns the AI models with the needs of New Zealand’s forestry sector.
“But what makes this dataset even more powerful is its benchmarking aspect.” In machine learning and AI, benchmark datasets play a pivotal role in advancing research. They provide a standardised way to evaluate different algorithms and approaches. The FOR-Instance dataset not only offers labelled data, but also predefines how the data should be split for training and evaluation of AI models. This ensures that researchers and industry can make fair and objective comparisons between their models and others in the field, enhancing the credibility of their work.
As part of the upcoming ForestTECH 2023 event, Grant Pearse, Scion’s Team Leader for Remote Sensing & GIS will be outlining to local foresters as part of the NZ leg of the annual series on 14-15 November, opportunities for using one of the largest datasets for high resolution land-cover mapping (a digital twin of New Zealand’s productive forest estate at national scale). Full details on the ForestTECH 2023 programme can be found here.
Source: Scion, FIEA