Aerial discrete return LiDAR (Light Detection And Ranging) technology (ALS – Aerial Laser Scanner) is now widely used for forest characterization due to its high accuracy in measuring vertical and horizontal forest structure. Random and systematic errors can still occur and these affect the native point cloud, ultimately degrading ALS data accuracy, especially when adopting datasets that were not natively designed for forest applications. A detailed understanding of how uncertainty of ALS data could affect the accuracy of derivable forest metrics (e.g. tree height, stem diameter, basal area) is required, looking for eventual error biases that can be possibly modelled to improve final accuracy. In this work a low-density ALS dataset, originally acquired by the State of Minnesota (USA) for non-forestry related purposes (i.e. topographic mapping), was processed attempting to characterize forest inventory parameters for the Cutfoot Sioux Experimental Forest (north-central Minnesota, USA). Since accuracy of estimates strictly depends on the applied species-specific dendrometric models a first required step was to map tree species over the forest. A rough classification, aiming at separating conifers from broadleaf, was achieved by processing a Landsat 8 OLI (Operational Land Imager
) scene. ALS-derived forest metrics initially greatly overestimated those measured at the ground in 230 plots. Conversely, ALS-derived tree density was greatly underestimated. To reduce ALS uncertainty, trees belonging to the dominated plane were removed from the ground dataset, assuming that they could not properly be detected by low-density ALS measures. Consequently, MAE (Mean Absolute Error) values significantly decreased to 4.0 m for tree height and to 0.19 cm for diameter estimates. Remaining discrepancies were related to a bias affecting the native ALS point cloud, which was modelled and removed. Final MAE values were 1.32 m for tree height, 0.08 m for diameter, 8.5 m2
for basal area, and 0.06 m for quadratic mean diameter. Specifically focusing on tree height and diameter estimates, the significance of differences between ground and ALS estimates was tested relative to the expected 'best accuracy'. Results showed that after correction: 94.35% of tree height differences were lower than the corresponding reference value (2.86 m); 70% of tree diameter differences were lower than the corresponding reference value (4.5 cm for conifers and 6.8 cm for broadleaf). Finally, forest parameters were computed for the whole Cutfoot Sioux Experimental Forest. Main findings include: 1) all forest estimates based on a low-density ALS point cloud can be derived at plot level and not at a tree level; 2) tree height estimates obtained by low-density ALS point clouds at the plot level are highly reasonably accurate only after testing and modelling eventual error bias; 3) diameter, basal area, and quadratic mean diameter estimates have large uncertainties, suggesting the need for a higher point density and, probably, a better mapping of tree species (if possible) than achieved with a remote sensing-based approach.
Borgogno Mondino, Enrico; Fissore, Vanina; Falkowski, Michael J.; Palik, Brian. 2020. How far can we trust forestry estimates from low-density LiDAR acquisitions The Cutfoot Sioux experimental forest (MN, USA) case study. International Journal of Remote Sensing. 41(12): 4549-4567. https://doi.org/10.1080/01431161.2020.1723173.