Abstract: Canopy height is a critical metric for assessing forest biomass, productivity, and ecological condition. Consequently, its rapid and accurate estimation remains essential for evaluating terrestrial ecosystem processes, dynamics, and services. In recent years, efforts to predict canopy height by integrating remote sensing technologies and artificial intelligence (AI) from regional to global scales have proliferated. This study evaluates the vertical accuracy of multiscale canopy height models (CHMs) in Andean Patagonia, assessing the performance of global (the High−Resolution Canopy Height Model of the Earth, HRCH; and the Global Map of Tree Canopy Height, GMTCH) and regional (the Nationwide Scale Native Forest Structure Maps, NNFS; and the Canopy Height Mapper, CH−GEE) canopy height products at spatial resolutions ranging from 1 m to 30 m. Model−derived vertical accuracy was quantified by benchmarking estimates against
in situ dominant height references, which were calculated from forest inventory plots spanning representative native stands along altitudinal gradients. We assessed model performance using three criteria: observed−versus−predicted regression analysis, mixed−effects modelling of prediction errors, and canopy height profiling. Overall, mean absolute error (MAE) values ranged from 5.6 m (HRCH) to 7.9 m (GMTCH), while root mean square error (RMSE) values spanned from 6.6 m (HRCH) to 9.9 m (GMTCH). Compared with dominant tree−height field measurements, HRCH, CH−GEE, and NNFS tended, on average, to overestimate canopy height, whereas GMTCH exhibited marked underestimation. Our results indicate that HRCH, followed by CH−GEE, achieves the best performance and calibration, exhibiting moderate agreement with field data. In contrast, canopy height predictions from GMTCH and NNFS failed calibration and displayed poor predictive power. Prediction errors were strongly CHM dependent and systematically influenced by elevation, with all CHMs exhibiting increasing overestimation along the altitudinal gradient. GMTCH showed the greatest sensitivity to elevation, indicating reduced predictive accuracy in complex mountainous terrain. In contrast, HRCH was comparatively insensitive to topographic variability, whereas CH−GEE and NNFS displayed intermediate but more heterogeneous responses. Forest type was also a significant source of systematic prediction error, yielding differences from substantial overestimation in low−stature
Nothofagus antarctica shrublands to consistent underestimation in tall, structurally complex
Nothofagus dombeyi forests. HRCH, NNFS, and CH−GEE tended to overestimate canopy height at dominant heights up to approximately 20 m, after which errors shifted toward underestimation. Conversely, GMTCH systematically underestimated canopy height across the full range of dominant heights. None of the evaluated CHMs achieved the ±2 m accuracy threshold commonly considered suitable for operational forest height assessment. Nevertheless, the performance of HRCH and CH−GEE suggests that these products can provide valuable regional−scale information where high−resolution local CHMs or extensive airborne LiDAR coverage are unavailable. Overall, our results provide realistic expectations of the capabilities and limitations of current−generation CHMs and support their application in forest monitoring and ecological assessments across data−limited temperate mountain forests.
Key words: accuracy assessment, mountain forests, multisource remote sensing, tree height estimation