AI helps find trees in a forest: 3D forest reconstruction from remote sensing data
January 20, 2026 - West Lafayette, Ind.
Existing algorithms can sketch a single, cleanly scanned tree. Forests are messier. A team from Purdue University's Department of Computer Science and Institute for Digital Forestry, working with Kiel University in Germany, has introduced TreeStructor - an AI method that isolates and reconstructs entire forests from lidar point clouds. The work appears in IEEE Transactions on Geoscience and Remote Sensing.
Why forests are hard to digitize
Built objects have symmetry. Trees don't. Vegetation is irregular and occluded: canopies overlap, branches intertwine, and laser pulses miss what sits behind leaves and stems. That's why methods tuned for buildings or cars break down in forest data.
TreeStructor leans on a different signal: repetition at multiple scales. Twigs echo branches. Branches echo larger limbs. By learning these repeated patterns, the model can piece together full tree structures even when points are sparse or partially hidden.
How TreeStructor works
The team first generated thousands of synthetic trees, scanned them with a virtual lidar, and chopped the results into identifiable parts. That became a "dictionary" of point-cloud snippets and matching geometric meshes (trunks, limbs, and branching units).
Given an unseen forest point cloud, TreeStructor splits it into small chunks, searches the dictionary for the best match, swaps each chunk with its corresponding geometric part, and then links those parts into connected, pipe-like structures. The result: a forest-scale mesh that separates individual trees and captures their branching geometry. Once trained, the system processes hundreds of trees in minutes on standard hardware.
- Collect multi-angle lidar (terrestrial, backpack, and/or drone) to reduce occlusion.
- Partition the point cloud into patches for matching.
- Match each patch to the learned part library (the dictionary).
- Replace points with geometric primitives; connect segments into tree graphs.
What this enables
Isolating and reconstructing trees at scale improves structure analysis, inventory workflows, growth modeling, and hazard assessments. It also sets the stage for species identification by expanding the dictionary with species-specific patterns, as noted by collaborators at the Institute for Digital Forestry.
Performance and limitations
The team validated TreeStructor by reconstructing the same trees from backpack, terrestrial, and drone-based scans, and comparing results to single-tree methods. Performance held up across sensors and setups.
Current limits: dead trees, shrubs, and understory debris are difficult to detect; lidar resolution and occlusion still cap what's recoverable. Better sensors help, but physics and canopy structure still matter.
Who built it
Key contributors include Bedrich Benes (Purdue), Sören Pirk (Kiel University), first author Xiaochen Zhou (Purdue), and Purdue collaborators Ayman Habib, Bosheng Li, and Jinyuan Shao. Support came from the National Science Foundation, the National Institute of Food and Agriculture, the Natural Resources Conservation Service, and the European Research Council. The effort aligns with Purdue Computes.
Practical notes for researchers and practitioners
- Mix perspectives: combine terrestrial and aerial scans to reduce blind spots in the canopy and mid-story.
- Preserve detail: avoid aggressive denoising that strips thin branches; TreeStructor benefits from fine features.
- Plan ground truth: tag species in sample plots to evaluate dictionary-based species cues.
- Compute needs: inference runs on a standard workstation; training is heavier but infrequent.
- Downstream use: export meshes/graphs to your existing inventory, biomass, or risk-analysis pipeline.
Paper and resources
The method is reported in IEEE Transactions on Geoscience and Remote Sensing. See the journal for related work and technical details: IEEE Xplore: TGRS.
Interested in applied AI upskilling for data-heavy field research? Explore curated learning paths here: Complete AI Training - Courses by job.
Your membership also unlocks:
