AI model counts individual trees from satellite imagery

Researchers from the University of Copenhagen and Harvard University developed a deep learning method that detects large individual trees in 3-meter PlanetScope imagery and converts satellite views into tree-by-tree maps. The approach could improve climate, carbon, and biodiversity monitoring by making scattered trees and fragmented landscapes easier to measure at scale.

Why it matters: - Trees influence climate, biodiversity, rural livelihoods, and the bioeconomy, but many global monitoring systems still treat forests as broad areas instead of individual trees. - The new approach could help count, locate, and monitor trees inside and outside forests across farms, savannas, drylands, and fragmented landscapes. - Better tree-level mapping could improve climate science, forest management, biodiversity assessment, restoration tracking, and carbon accounting.

What happened: - Researchers from the University of Copenhagen and Harvard University reported a new deep learning framework in Journal of Remote Sensing on April 30, 2026. - The method detects large individual trees in 3-meter PlanetScope imagery. - The study is documented under DOI 10.34133/remotesensing.1049. - The source URL for the paper is the full paper.

The details: - The model uses an anchor-free design that represents each tree crown as a Gaussian heatmap instead of a fixed detection box. - The heatmap lets researchers extract crown centers and generate binary tree cover maps. - Training combined PlanetScope imagery, airborne light detection and ranging data-derived canopy height models, GEDI data, and satellite-based location embeddings. - The training set included about 14 billion tree points across about 1,030,000 square kilometers. - The training inputs used PlanetScope imagery from 2018 to 2022 and LiDAR sources from 17 countries. - A U-Net model with a ResNet50 encoder produced both a heatmap and a spatial uncertainty map. - The model reached fractional cover R² of 0.81 against aerial LiDAR. - The system showed balanced detection metrics across biomes. - Satellite Contrastive Location-Image Pretraining, or SatCLIP, improved generalization. - Fine-tuning with manual labels further sharpened predictions. - The heatmap output can be used for either tree counting or tree cover mapping.

Between the lines: - The main shift is conceptual as much as technical: trees are no longer treated only as continuous green canopy but as countable objects. - Gaussian crowns help the model handle crown-size variation, noisy pseudo-labels, and imperfect alignment between LiDAR and satellite imagery. - The researchers framed the system as a scalable framework, not a final global tree census. - The work also shows where the field still struggles, especially in regions with limited airborne LiDAR coverage. - Threshold selection across regions remains a technical hurdle. - Performance drops where small crowns or dense forests blend together at the current ground sampling distance.

What's next: - The approach could be expanded with better imagery, more LiDAR data, and targeted manual labels. - The method is designed to transfer to future satellite missions with higher spatial and radiometric quality. - Remaining gaps in South America, Africa, and Asia could limit near-term global coverage until more training data become available. - Researchers may use the system to surface overlooked tree resources in agricultural land and other human-managed landscapes.