McMaster and Planet Labs build self-supervised AI to map urban tree carbon from LiDAR

A team led by McMaster University, in collaboration with Environment and Climate Change Canada and Planet Labs, has released a self-supervised AI framework that estimates tree-level above-ground biomass (AGB) in urban areas from airborne LiDAR and optical imagery. Published on arXiv, the system uses a dual-stream cross-attention network trained on rule-based pseudo-labels to semantically segment buildings, needleleaf trees, and deciduous trees. On independent test tiles, it achieves global mean precision of 0.86, recall of 0.83, and Dice score of 0.84. Individual tree crowns are delineated using multiscale watershed segmentation, and AGB is estimated via a crown area–height power-law proxy calibrated to species-specific allometry from 21,921 inventory trees.

For a landscape of 810 km² in Ontario, Canada, the framework processed leaf-off airborne LiDAR at 8–10 pulses per square meter and orthophotography at 0.16–0.20 m resolution from 2018 and 2023. On a held-out test set of 90,726 trees, the model achieved an R² of 0.609 for AGB prediction when using inventory crown geometry and 0.570 under operational segmentation—highlighting crown delineation as the primary uncertainty source. Aggregated to 30 m resolution, the estimates yielded total AGB stocks of 1.73 Tg in 2018 and 1.81 Tg in 2023, with densities up to ~140 Mg/ha along the Niagara Escarpment and a net carbon gain of 39 Gg C over five years. The framework uses standard provincial data, requires no manual annotation, and generates deep-ensemble uncertainty maps that guide assignment of uncertain crowns to a pooled allometric equation. The resulting public bitemporal crown-level AGB database enables high-resolution urban carbon accounting for trees outside forests.