Counting trees in Dutch forests: Using AHN LiDAR data for National Forest Inventory

By Thijs Koop

Abstract
The Dutch National Forest Inventory (NFI) has not used remote sensing for managing its national forests while there is nationwide Light Detection and Ranging (LiDAR) data called ‘Actueel Hoogtebestand Nederland’ (AHN) available since the late 1990s. The main objective of this thesis was to assess the usability of AHN LiDAR data for forest inventory applications. The availability of high-resolution RICOPTER LiDAR data for four forest study areas with different tree species provided the opportunity to extensively research the individual tree detection (ITD) and tree height accuracy by using AHN LiDAR data. Two existing algorithms for individual tree detection were applied, the raster-based Local Maxima Filter (LMF) and the point cloud-based Li2012 algorithm. The ITD and tree height accuracy of AHN was validated by manually selecting treetops in the RICOPTER LiDAR data. The accuracy was evaluated with the accuracy metrics F-score for ITD and mean error plus gradient of the regression line and residuals for tree height. Next, aiming at the potential of AHN for NFI the robustness of results for different areas and computation time of the algorithms by using AHN was assessed. For coniferous forests, both algorithms produce similar ITD results by using AHN LiDAR data. The found F-scores range from 58.6% to 84.5%, where trees with conical shapes show the best potential. For deciduous forests Li2012 performed better than LMF, although the found F-scores hardly show potential ranging from 13.9% to 63.5%. Also, both algorithms detect more trees in the area with more space between the crowns. Regarding tree height, both methods underestimated for all areas, although Li2012 (-0.07 m to -1.24 m) less profound than LMF (-0.85 m to -2.26 m). The overall tree height error of Li2012 was 0.8 m smaller than LMF. The detailed assessment provided a more in-depth understanding of the contribution of ground level and treetop level to total tree height error. The ground level contribution was relatively small but can both under- and overestimate compared with the RICOPTER derived ground elevation. For the treetop level, it was concluded that Li2012 results in better estimations. Also, the precision in the form of smaller residuals and stable gradients of the regression lines make Li2012 more suitable to correct for known errors. It was found that regarding computation time Li2012 takes on average 1956 times as long as LMF. However, because Li2012 outperforms the LMF approach in terms of ITD, more accurate tree height estimations and smaller residuals, it is recommended to invest extra computational power for higher accuracies. Considering AHN for NFI, Li2012 can better be implemented for deriving tree height of individual trees than LFM. It was concluded that AHN can be used for specific NFI applications, especially when the focus is on dominant coniferous trees with conical shape.

Keywords: AHN; LiDAR; Forest Management; National Forest Inventory; Individual Tree Detection; Tree Height.