The latitudinal forest-tundra ecotone (LFTE) represents the largest ecological boundary on Earth. Trees within this region take unusual growth forms, often presenting as short, stunted, and deformed, while exhibiting complex spatial patterns involving small, isolated individuals or patches. This complexity makes canopy height modelling challenging and likely means that canopy height models (CHMs) trained on forests in lower latitudes and extrapolated into the LFTE may not represent forests in this region accurately. This study presents a CHM within the LFTE by upscaling regional airborne laser scanning (ALS) data using satellite imagery to improve upon existing CHM products. Sentinel-1/2, PlanetScope, and Pleiades 1A/1B -based models were produced using multiple machine and deep learning algorithms. Estimating maximum canopy height using PlanetScope and an artificial neural network outperformed three globally extrapolated CHM products, with an RMSE half that of the next best model (RMSE = 2.13 m), and a data distribution that closely matched field measured height values. Proportional biases in model predictions were corrected using orthogonal regression, but wall-to-wall mapping displayed substantial overprediction. Therefore, wall-to-wall prediction surfaces were constrained using an attention augmented U-Net + ResNet-50 encoder backbone for binary semantic segmentation, capturing forest area well but with suboptimal F1-scores, highlighting the challenge in finding the balance between over- and under- prediction of sparse forest area. Overall, this study demonstrates that regional upscaling of ALS height using PlanetScope satellite imagery outperforms multiple pre-existing global CHM products regarding height magnitude, using a methodology replicable in other ecotonal or sparsely forested ecosystems.

The latitudinal forest-tundra ecotone (LFTE) represents the largest ecological boundary on Earth. Trees within this region take unusual growth forms, often presenting as short, stunted, and deformed, while exhibiting complex spatial patterns involving small, isolated individuals or patches. This complexity makes canopy height modelling challenging and likely means that canopy height models (CHMs) trained on forests in lower latitudes and extrapolated into the LFTE may not represent forests in this region accurately. This study presents a CHM within the LFTE by upscaling regional airborne laser scanning (ALS) data using satellite imagery to improve upon existing CHM products. Sentinel-1/2, PlanetScope, and Pleiades 1A/1B -based models were produced using multiple machine and deep learning algorithms. Estimating maximum canopy height using PlanetScope and an artificial neural network outperformed three globally extrapolated CHM products, with an RMSE half that of the next best model (RMSE = 2.13 m), and a data distribution that closely matched field measured height values. Proportional biases in model predictions were corrected using orthogonal regression, but wall-to-wall mapping displayed substantial overprediction. Therefore, wall-to-wall prediction surfaces were constrained using an attention augmented U-Net + ResNet-50 encoder backbone for binary semantic segmentation, capturing forest area well but with suboptimal F1-scores, highlighting the challenge in finding the balance between over- and under- prediction of sparse forest area. Overall, this study demonstrates that regional upscaling of ALS height using PlanetScope satellite imagery outperforms multiple pre-existing global CHM products regarding height magnitude, using a methodology replicable in other ecotonal or sparsely forested ecosystems.

Aboveground biomass (AGB) mapping is fundamental to tropical forest carbon monitoring, yet national-scale estimation remains challenging because field plots are sparse and model performance is often sensitive to predictor choice and validation design. This study assessed whether geospatial embeddings improve national AGB mapping in Guyana when combined with environmental and topographic predictors. Predictor selection was undertaken using repeated grouped resampling at the plot-cluster level, and model performance was evaluated across 100 independent train–test repeats. Three final random forest models were compared. The environmental baseline model (Env + SRTM-derived elevation; 8 predictors) achieved a mean R² of 0.179, an RMSE of 148.5 Mg/ha and a relative RMSE of 36.1%. A retained 8-predictor model combining environmental variables with a selected embedding subset (Env + Emb*) improved performance slightly, with a mean R² of 0.189, an RMSE of 147.6 Mg/ha and a relative RMSE of 35.9%. The best performance was obtained with a 22-variable full-stack model combining environmental, topographic and embedding predictors, after all Sentinel-2 predictors had been eliminated during feature selection; this model achieved a mean R² of 0.203, an RMSE of 146.3 Mg/ha and a relative RMSE of 35.5%. Across models, isothermality, a measure of how day-to-night temperature variation compares to annual temperature variation, and precipitation of the coldest quarter were consistently the most influential predictors. Mean ensemble coefficient of variation, representing relative model disagreement, ranged from 0.336 to 0.361. These results indicate that geospatial embeddings provide useful complementary information, but predictive performance remained modest overall, with the best model explaining only about one-fifth of plot-level AGB variance. The resulting maps are therefore best interpreted as broad-scale decision-support products rather than high-precision local estimates of AGB.