National-scale tree counting has long depended on costly field inventories or coarse statistical extrapolations that miss trees outside forests. A new study led by the University of Copenhagen offers a scalable alternative: a deep-learning framework that fuses PlanetScope, Sentinel-1, and Sentinel-2 imagery to estimate tree abundance from spatial, spectral, and textural patterns rather than individual crowns. The research, published (DOI: 10.34133/remotesensing.0970) in the Journal of Remote Sensing, matters because global tree-planting pledges have surged since the Paris Agreement, yet governments still lack reliable, repeatable baselines of how many trees already exist and where.
Applied to Rwanda and Tanzania, the model estimated roughly 5.25 billion trees—about 213 million in Rwanda and 5.04 billion in Tanzania after calibration. Only 48.8% fell within the WorldCover “tree cover” class; the remaining 51.2% occurred in shrublands, grasslands, croplands, built-up areas, wetlands, sparse vegetation, and mangroves. This finding challenges forest-focused inventories and suggests that restoration monitoring must account for trees well beyond conventional forest boundaries. The implication is significant for policy: planting programs and carbon accounting that ignore nonforest trees may misstate both baselines and gains.
The framework analyzed 96 × 96 m image patches from a 20-band composite containing four Planet optical bands, ten Sentinel-2 multispectral bands, and six Sentinel-1 radar polarization bands, all resampled to a common 3-m grid. A hybrid UNet-ResNet50 architecture predicted one tree-count value per patch, achieving a mean absolute error of 17.32 and an r² of 0.73, outperforming ResNet50, ResNeXt50, SwinT, and ViT. Sensor-ablation tests showed that Sentinel-1 plus Sentinel-2 alone reached an MAE of 18.42 and r² of 0.77, comparable with the three-sensor composite at 18.8 and 0.77—an important cost signal, since it implies fewer data sources may suffice for national-scale work.
Training used about 12 million tree-crown centers from diverse landscapes, with labels drawn from manual interpretation of submeter imagery and semi-automated extraction from airborne and UAV LiDAR canopy-height models. National predictions were bias-corrected using independent National Forest Inventory measurements with country- and land-cover-specific bootstrap calibration, yielding median densities of 37 trees per hectare in Rwanda and 42 in Tanzania. The researchers noted the framework “can guide the planning and optimization of resource allocation for tree restoration efforts at national scales,” and argue repeatable satellite baselines could help governments track outcomes, identify appropriate planting areas, and better account for trees beyond forest boundaries.
Limitations remain. Dense closed-canopy forests, where crown overlap and signal saturation reduce sensitivity, need improvement, as does incorporation of understory and smaller woody plants. The authors propose integrating forest inventories and very-high-resolution data while exploring attention-based sensor fusion, domain adaptation, and transfer learning to extend the approach to new regions, ecosystems, and sensors. The full study is available at https://spj.science.org/doi/10.34133/remotesensing.0970.
