The bigger, the better? Sample size effects in drone-estimated forest height

Abstract Accurate tree height measurement is crucial for assessing forest health and management strategies, as it directly correlates with biomass and carbon storage capabilities. Thus, drones are increasingly used in forest remote sensing due to their high spatial resolution and operational flexibility. They support individual tree detection and detailed structural mapping, yet the reliability of tree height estimates remains influenced by multiple factors. We present a meta-analysis of 36 case studies to evaluate how environmental complexity and sample size influence tree height estimation accuracy. Forest environments, characterised by heterogeneous canopy structures, occlusions, and species diversity, were associated with higher errors and more significant variability. Conversely, plantations and urban sites yielded lower and more consistent errors, reflecting their structural simplicity and spatial regularity. Although lidar and image-based methods performed similarly in forests (1.58 m vs. 1.69 m mean error), lidar clearly outperformed image matching in plantations (0.44 m vs. 0.91 m). Regarding sample size, mean errors were not consistently lowest in larger datasets, but studies with more than 300 trees exhibited the lowest variability, indicating more stable performance. These findings highlight that both environmental structure and sample size affect the robustness of drone-based height estimation. We recommend the implementation of standardised workflows that account for environmental and technical limitations, including terrain complexity, sensor configuration, and weather conditions. Although drones provide considerable benefits, their successful application depends on careful mission planning and grounded operational expectations. Addressing these challenges through improved mission planning and methodological consistency is essential to ensure the robustness and scalability of drone applications in long-term forest monitoring.