Accurate tree counts are essential for restoration, yet conventional methods struggle across heterogeneous landscapes. A research team led by the University of Copenhagen has developed a deep-learning framework that combines multisensor satellite imagery to estimate tree abundance, and applied it to Rwanda and Tanzania to count 5.25 billion trees. The results show that more than half of these trees occur outside areas classified as tree cover, a finding that could reshape how governments plan and monitor restoration.
Global tree-planting pledges have increased sharply since the Paris Agreement, creating demand for reliable baselines showing how many trees already exist and where they occur. Field inventories are costly and spatially limited, while repeated acquisition of submeter satellite or UAV imagery is difficult to scale. Object-detection methods also struggle when crowns overlap, occupy fractions of pixels, or lack clear boundaries. Coarser approaches based on vegetation indices, climate, or topography may smooth local variation and underrepresent trees in nonforest systems.
The team published (DOI: 10.34133/remotesensing.0970) the study on August 18, 2026, in the Journal of Remote Sensing. The framework uses PlanetScope, Sentinel-1, and Sentinel-2 imagery to estimate tree numbers from spatial, spectral, and textural patterns within image patches, rather than detecting every crown individually. This targets a restoration-monitoring challenge: producing consistent, spatially detailed tree-count baselines across forests and nonforest landscapes without relying on expensive, repeatedly acquired submeter imagery.
The framework estimated roughly 213 million trees in Rwanda and 5.04 billion in Tanzania after calibration. Only 48.8% fell within the WorldCover tree cover class; the other 51.2% occurred across shrublands, grasslands, croplands, built-up areas, wetlands, sparse vegetation, and mangroves. Unlike forest-focused inventories or coarse statistical extrapolations, the approach captures trees across heterogeneous landscapes. Its image-level regression avoids dependence on explicit crown boundaries, allowing tree numbers to be inferred even when individual crowns are not fully resolved at 3-m pixel resolution. This provides a more direct, spatially explicit baseline for restoration planning and monitoring.
The model 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. In architecture 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 that Sentinel-1 plus Sentinel-2 reached an MAE of 18.42 and r² of 0.77, comparable with the three-sensor composite at 18.8 and 0.77, respectively.
Training used about 12 million tree-crown centers from diverse landscapes, assembled from manual interpretation of submeter imagery and semi-automated extraction from airborne and UAV LiDAR canopy-height models. The national predictions were then bias-corrected using independent National Forest Inventory measurements, with country- and land-cover-specific bootstrap calibration. Final estimates indicated median densities of 37 trees per hectare in Rwanda and 42 in Tanzania, demonstrating national applicability across contrasting ecosystems.
The researchers emphasized the framework's practical value, noting that it "can guide the planning and optimization of resource allocation for tree restoration efforts at national scales." They further argue that repeatable satellite-based baselines could help governments track restoration outcomes, identify appropriate planting areas, and better account for trees growing beyond conventional forest boundaries.
The framework could support restoration programs by establishing repeatable national baselines, tracking tree gains and losses, and guiding planting toward ecologically appropriate locations. Future work must improve performance in dense closed-canopy forests, where crown overlap and signal saturation reduce sensitivity, and incorporate understory or 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 satellite sensors at larger scales.
