Advisors: Markus Hollaus

Tree heights derived from airborne laser scanning and computed from a canopy height model show systematic deviations in steep terrain. A key reason is that the treetop is usually detected correctly, while the height reference below the crown is always determined vertically to the terrain surface and therefore does not correspond to the stem base. This thesis investigates this effect in the Rohrach European protected area in Vorarlberg and develops a correction for ALS tree heights that uses only features from the ALS data. The study area covers about 105 ha and has very steep terrain. 37.6 percent of the area exceeds a slope of 30 degrees. A very dense UAV LiDAR data set with about 4,380 points per square meter and an ALS data set with about 40 points per square meter are available. Reference tree heights are computed automatically from the UAV data. Individual trees are segmented, treetops are identified, and stem base points are found using a shortest-path approach following Dijkstra on a graph of point-cloud and DTM nodes. Tree height is defined as the vertical difference between treetop and stem base point. UAV and ALS trees are matched using treetops by applying a 2 m by 2 m search cylinder at the treetop. This yields 5,418 unique pairs. For the correction, a linear ensemble of Random Forest and Histogram Gradient Boosting is trained. In addition, trees in areas with high slope and in zones with strongly varying terrain elevation within a local neighborhood are weighted more strongly, and the correction values are calibrated. In a spatially separated cross-validation within the training areas, the MAE decreases from 0.97 m to 0.37 m and the RMSE decreases from 1.59 m to 0.57 m. The 99th percentile of absolute errors decreases from 6.11 m to 1.97 m and the bias is 0.002 m. In the independent test area, the MAE decreases from 1.21 m to 0.76 m and the RMSE decreases from 1.95 m to 1.16 m. The 99th percentile decreases from 7.74 m to 3.99 m and the bias decreases from 0.96 m to 0.11 m. This reduces both systematic overestimation and rare extreme errors, and ALS tree heights become more robust for use in steep terrain.