Continental-scale mapping of forest tree density in North America using remote sensing and deep learning with uncertainty quantification

Accurate, spatially consistent estimates of tree density remain elusive at continental scales, limiting our ability to assess forest structure, carbon stocks, and biodiversity. Existing global assessments have relied on simplified statistical models and sparse, heterogeneous ground data that are insufficient to capture nonlinear ecological interactions and spatial variability. To address these limitations, we integrated more than 600,000 harmonized ground-based forest inventory plots with satellite-derived vegetation indices, climate surfaces, soil properties, and topographic covariates to develop a deep learning framework for high-resolution mapping of tree density across North America. We evaluated four modeling approaches—generalized linear models (GLMs), ridge regression (RR), random forest (RF), and a feedforward neural network (FFNN). Among all models tested, the FFNN achieved the highest predictive accuracy (RMSE = 344.8; R² = 39.53%), and was used to produce a wall-to-wall tree density map at 3 km resolution for the continent. We estimated that the total number of forest trees with diameter at breast height (DBH) ≥ 10 cm across North America ranges from 339 to 514 billion, substantially lower than the widely cited estimate of 603 billion trees reported by Crowther et al. (2015). When smaller stems were included (no DBH threshold), totals more than doubled, reaching 738 billion to 1.12 trillion trees. We quantified uncertainty using Monte Carlo (MC) Dropout, generating pixel-level error estimates and confidence intervals. Spatial patterns reveal high tree densities in boreal and temperate forests, intermediate densities in mixed broadleaf regions, and relatively low densities in deserts, Mediterranean systems, and tundra. Compared to the global GLM-based benchmark by Crowther et al. (2015), our deep learning framework achieves markedly higher predictive accuracy, aligns more closely with national forest inventory statistics, and provides explicit uncertainty quantification, supporting applications in carbon accounting, biodiversity modeling, and ecosystem monitoring at scales through region specific calibration and validation.