Abstract Fine-scale architectural tree models serve as an effective representation of three-dimensional plant material distributions. They can help to quantify wood volume and biomass, to estimate leaf area distributions on a detailed scale, and can be exploited for physically based modelling approaches. If architectural tree models can be derived for multiple acquisition dates, they permit the detailed investigation of phenological effects. Although promising approaches for the generation of architectural tree/forest models from terrestrial LiDAR data are available, they are often non-trivial and their application to forest plots is often difficult. This is restricting the flexibility of these reconstruction approaches especially for multi-temporal analyses. In this paper, forest models of two Larix decidua forest plots are reconstructed by making use of terrestrial LiDAR data and digital hemispherical photographs (DHP). Recent modelling strategies are enhanced and developed further in order to improve the robustness and usability of the architectural tree model reconstruction process. Raw point cloud data are directly used as input to solve both tree delineation and tree reconstruction in a single processing pipeline. This includes terrain filtering, intensity filtering, and trunk extraction. These steps are followed by a hierarchical and iterative multi-tree branch and twig reconstruction. Based on multi-temporal DHPs, various foliage states are documented. These DHPs and the reconstructed branching architectures are used to flexibly generate and update multi-temporal 3D models of foliage. In order to quantify the modelling performance with respect to various forest characteristics, a test setup based on simulated forest and acquisition geometries is build up. It can be shown, that typical sources of error in the tree reconstruction process are minimized by the proposed approach. It is possible to estimate wood volume distributions, trunk tapering and leaf area distributions with an error of only 10–14%. Except for strongly overlapping tree crowns, the overall accuracy of the single tree delineation in interlinked tree crowns is higher than 80%. Considering these error margins, we apply the modelling strategy to two forest plots and derive architectural models for three dates during the growing season. Using DHPs as reference data, it can be shown, that the estimated gap fraction values derived from the generated models show an error of only 10–15%.

中文翻译：

---

使用地面激光雷达和半球照片对落叶松林地块进行多时相精细尺度建模

摘要 精细尺度的建筑树模型是三维植物材料分布的有效表示。它们可以帮助量化木材体积和生物量，在详细尺度上估计叶面积分布，并可用于基于物理的建模方法。如果可以为多个采集日期导出建筑树模型，则它们允许对物候效应进行详细调查。虽然从陆地 LiDAR 数据生成建筑树/森林模型的有前途的方法是可用的，但它们通常是重要的，并且它们在森林地块中的应用通常很困难。这限制了这些重建方法的灵活性，尤其是对于多时态分析。在本文中，利用地面激光雷达数据和数字半球照片 (DHP) 重建了两个落叶松森林地块的森林模型。最近的建模策略得到了进一步增强和发展，以提高建筑树模型重建过程的鲁棒性和可用性。原始点云数据直接用作输入，在单个处理流水线中解决树的勾画和树的重建。这包括地形过滤、强度过滤和主干提取。这些步骤之后是分层和迭代的多树分支和树枝重建。基于多时态 DHP，记录了各种树叶状态。这些 DHP 和重建的分支架构用于灵活地生成和更新树叶的多时相 3D 模型。为了量化关于各种森林特征的建模性能，建立了基于模拟森林和采集几何的测试设置。可以证明，树重建过程中的典型误差源通过所提出的方法被最小化。可以估计木材体积分布、树干锥度和叶面积分布，误差仅为 10-14%。除了强烈重叠的树冠外，在相互连接的树冠中，单一树的勾画整体准确率高于80%。考虑到这些误差幅度，我们将建模策略应用于两个森林地块，并在生长季节推导出三个日期的建筑模型。使用DHPs作为参考数据，可以显示，