Municipal tree registers mainly document trees managed by local authorities. Trees on private land and some other woody vegetation may be absent from these records.
The original study, Victoria Luise Pissulla’s master’s thesis at Ruhr University Bochum, used open RGB orthophotos and airborne LiDAR data to map tree-like vegetation in Troisdorf’s settlement areas and estimate crown height and geometric volume. CATCH4D Upscaling then tested whether LiDAR-derived height could also improve detection of individual crowns, especially where neighbouring crowns touch.
A small part of Troisdorf, seen from above. Trees appear between buildings, beside roads and in groups. This is one of the study's image patches.
The starting point
Looking beyond the tree register
A register is built for managing particular trees. A wider vegetation inventory asks a different question: what is growing across the settlement, including on private land? Even the boundary between a tree and a shrub can affect what a register includes.
The original study explored this question in Troisdorf, near Cologne and Bonn in Germany. It focused on about 29 km² of settlement areas. Open aerial data made it possible to examine those areas together and connect vegetation across individual property boundaries.
Where the story takes place. Troisdorf is highlighted in red within the Rhein-Sieg district.
Locator map based on “Troisdorf in SU.svg”
by TUBS, Wikimedia Commons,
licensed under CC BY-SA 3.0.Where the analysis took place. Yellow areas mark the settlement study areas; orange marks the municipal boundary. The study did not cover every tree in the municipality.
The first step
Giving each crown a place on the map
In an aerial image, we can see patches of vegetation. To describe individual trees, we also need to identify which part of the image belongs to each crown—the leafy canopy visible from above.
The original work trained a deep-learning model, Mask R-CNN, using hand-drawn examples. It traced separate crown outlines across the study areas. Those outlines provided the link between a tree's appearance in the image and its position in the landscape.
One landscape, different challenges. Magenta lines show the original detector's crown outlines. Individual trees are easier to distinguish where there is space between them (left). In a continuous canopy (right), deciding where one crown ends and the next begins is harder.
The resulting inventory contained 70,438 detected tree-like vegetation structures. These were automated detections, so the inventory could still include missed, combined or incorrectly identified crowns. It provided a starting point for exploring vegetation structure across the study areas.
Adding the third dimension
What an outline cannot tell us
A crown outline tells us how much space a tree occupies when viewed from above. It cannot tell us the crown's height or three-dimensional form.
Airborne laser scanning, or LiDAR, supplies measurements at different heights. Using the mapped crown to select the corresponding points allowed the original workflow to estimate height and the space enclosed by the crown. The image and laser data could now be read together.
Seen from above. The crown outline connects an object in the image to the laser measurements around it.Seen in three dimensions. The same example reveals its vertical form. Each dot is a laser measurement; the original colours distinguish point groups rather than height.
This opened a way to explore differences in crown size across Troisdorf. The volume estimates describe geometric space around measured points; they are exploratory measures of structure, rather than direct measurements of wood or carbon.
The next question
Could height help us find the crowns in the first place?
The difficulty visible in the dense canopy remained: neighbouring trees can blend together in an overhead image. If a crown is missed or combined with another, the later analysis cannot describe it separately.
That challenge became the starting point for CATCH4D Upscaling. In the original workflow, height helped describe crowns after they had been detected. We asked whether it could also help the model during detection, when it decides which pixels belong to each crown.
Studies elsewhere had reported different outcomes from combining imagery and height (Lucena et al., 2022; Ruschhaupt et al., 2025). An improvement was plausible, but needed to be tested here. We compared an image-only model with a model that also received a layer highlighting height above nearby surroundings, called relative height.
The two parts of the work now offer complementary views: a way to explore the structure of detected vegetation, and a test of whether height can improve the detection that comes first.
Explore the work
Follow the experiment, or explore the trees and their three-dimensional structure.
The detection experiment
Finding individual crowns
Compare the models, inspect crowns in the aerial images and see how the study measured success.
Read the notebook HTML files without installation, or choose Run All in Jupyter to recalculate the analyses. See the README for files and running instructions.
Contributions
Person
Contribution
Victoria Luise Pissulla master's thesis, Ruhr University Bochum
Reference crowns, the colour-only (RGB) tree detector, the town-wide detection, the
per-tree LiDAR point clouds, the volume method and all crown-volume results
Zahra Chehreghan CATCH4D Upscaling, ILS
The height study, its experiments and GIS package, and this repository and
website, including the presentation of the crown-volume work
Dr. Shaojuan Xu ILS
Supervision of CATCH4D Upscaling
References
Pissulla, V. L. (2026). AI-based detection of tree-like vegetation structures from orthophotos using Mask R-CNN with subsequent volume estimation from LiDAR data. Master's thesis, Institute of Geography, Ruhr University Bochum.
Lucena, F., Breunig, F. M. and Kux, H. (2022). The Combined Use of UAV-Based RGB and DEM Images for the Detection and Delineation of Orange Tree Crowns with Mask R-CNN: An Approach of Labeling and Unified Framework. Future Internet, 14(10), 275. doi:10.3390/fi14100275.
Ruschhaupt, S., Troles, J. and Schmid, U. (2025). Comparing Mask R-CNN and Mask2Former architectures for individual tree crown delineation. 45. GIL-Jahrestagung, pp. 167–178. doi:10.18420/giljt2025_13.
