National tree-planting pledges have surged since the Paris Agreement, but governments lack reliable, spatially explicit baselines of how many trees already exist and where. A new deep-learning framework, developed by a University of Copenhagen-led team and published (DOI: 10.34133/remotesensing.0970) in the Journal of Remote Sensing, addresses that gap by estimating tree abundance from multisensor satellite imagery across entire countries. Applied to Rwanda and Tanzania, the system counted approximately 5.25 billion trees—about 213 million in Rwanda and 5.04 billion in Tanzania after calibration.
The most consequential finding is where those trees are. 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 means conventional forest-focused inventories and coarse statistical extrapolations systematically miss more than half of the trees in these landscapes, undermining restoration planning and carbon accounting that depend on accurate baselines.
The framework combines PlanetScope, Sentinel-1, and Sentinel-2 imagery into a 20-band composite resampled to a 3-meter grid, analyzing 96 × 96 meter patches with a hybrid UNet-ResNet50 architecture that predicts one tree-count value per patch. Rather than detecting individual crowns—which fails when crowns overlap or occupy fractions of pixels—the model infers tree numbers from spatial, spectral, and textural patterns. In comparisons, UNet-R50 achieved a mean absolute error of 17.32 and an r² of 0.73, outperforming ResNet50, ResNeXt50, SwinT, and ViT. Sensor-ablation tests showed Sentinel-1 plus Sentinel-2 alone reached an MAE of 18.42 and r² of 0.77, comparable to the three-sensor composite, suggesting a more affordable path to national monitoring.
Training drew on roughly 12 million tree-crown centers across 2,832 sample areas spanning about 608,483 hectares, assembled from manual interpretation of submeter imagery and airborne and UAV LiDAR canopy-height models. National predictions were bias-corrected against independent National Forest Inventory plots, yielding median densities of 37 trees per hectare in Rwanda and 42 in Tanzania. The full study is available at https://spj.science.org/doi/10.34133/remotesensing.0970.
The researchers argue the framework “can guide the planning and optimization of resource allocation for tree restoration efforts at national scales,” and that repeatable satellite-based baselines could help governments track restoration outcomes, identify appropriate planting areas, and account for trees beyond conventional forest boundaries. Limitations remain in dense closed-canopy forests, where crown overlap and signal saturation reduce sensitivity, and understory or smaller woody plants are not yet captured. 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. For restoration programs that must verify progress against billion-tree pledges, the ability to produce consistent, spatially detailed counts without repeated submeter imagery acquisition marks a practical shift from extrapolation to observation.
