Australia-Wide 30 m Machine Learning-Derived Canopy Height Models Composites: Best Pick and Median 

This dataset is part of the OzTreeMap project and provides two new 30 m spatial resolution canopy height products for continental Australia: the best-pick canopy height model (pick-CHM) and the median canopy height model (med-CHM). Both products represent estimates of vegetation canopy height across Australia and were developed to improve the accuracy and consistency of existing large-scale canopy height models, which were generated by researchers to represent canopy heights from variable time periods ranging from 2007 until 2020. The best-pick and median products are composites and derive from an extensive validation of the 4 original CHMs. Each product is provided as a single-band GeoTIFF raster in the Australian Albers (EPSG:3577) coordinate reference system, with 30 m spatial resolution and float32 data type. These datasets support applications in forest structure monitoring, carbon accounting, and ecosystem assessment across Australia. 

We at TERN acknowledge the Traditional Owners and Custodians throughout Australia, New Zealand and all nations. We honour their profound connections to land, water, biodiversity and culture and pay our respects to their Elders past, present and emerging. This research was conducted under the ANU-CSIRO Collaborative Research Agreement entitled ‘Spatial and vertical distributions of individual tree and shrub canopies across Australian ecosystems’. We thank Drs Libby Pinkard, Steve Roxburgh, and Glenn Newnham (all from CSIRO Environment) for their continued support. 

These datasets provide better Canopy Height Model estimates as validated versus an extensive point cloud datasets, therefore, these could improve downstream modelling works where the original datasets are used. 

A total of 26,987 LiDAR and photogrammetry point cloud tiles (1–4 km² each) were obtained from the Elevation and Depth (ELVIS) and Terrestrial Ecosystem Research Network (TERN) open repositories, representing a 5% stratified sample designed to match the distribution of Australia’s 16 vegetation structure classes (Scarth et al., 2019). For each tile, a 0.5 m canopy height model (CHM) was generated using the pit-free algorithm (Khosravipour et al., 2014), and individual tree crowns were delineated with the Dalponte segmentation algorithm (Dalponte & Coomes, 2016) using vegetation-specific optimized parameters (Pucino et al., 2025, under review).

The resulting point-cloud-derived CHMs served as reference data for evaluating the vertical accuracy of four publicly available satellite-based machine-learning or deep learning-derived CHMs: (1) Lang et al. (2023); (2) Liao et al. (2020); (3) Potapov et al. (2021); and (4) Tolan et al. (2024). All datasets were co-registered and resampled to 30 m resolution. Pixel-wise error metrics were computed, and a combined score defined for each vegetation class which publicly available dataset is the most accurate. Water bodies are masked using Digital Earth Australia Waterbodies dataset. Three new continental-scale 30 m CHMs were then produced: (i) a pixel-wise median composite; (ii) a vegetation-class-specific best-pick composite; and (iii) a deep-learning CHM derived from a multi-layer perceptron (MLP - not publicly available). Note: this document's Start Date and End Date indicate the nominal dates of the datasets we tested, not the publication dates of their associated articles.