Mapping forest environments as a frontier for 3D vision
High-resolution LiDAR from terrestrial scanners and drones can produce dense 3D maps of forests. Turning these maps into useful information is hard – forests have a complex, multi-layered structure, full of fractal-like shapes and heavy occlusion. Classical geometric methods struggle with such data, but deep learning is showing very promising results.
At IDEAS Research Institute I work on an AI pipeline that segments individual trees from point clouds and classifies their species. A critical bottleneck is training data. Labelled 3D forest data is scarce, so we created TreeScanPL10K, a hand-labelled dataset of over ten thousand trees with species and locations. Building on it, I started working on species classification leveraging foundation models and location context. We designed a fusion model that combines point cloud features with AlphaEarth satellite embeddings describing the location. In my latest project I render each tree as a set of 2D images and classify them with the pretrained vision model DINOv3, which outperforms networks designed specifically for 3D point clouds and only requires training a lighweight head. Together, these steps bring us closer to operational, automated tree-level forest inventory.
Bio
Maksymilian Kulicki is a researcher at IDEAS Research Institute and a PhD candidate at IPPT PAN. He got his Master’s degree in Artificial Intelligence at the University of Amsterdam. Today his research sits at the boundary of AI and forest science. He develops methods that turn LiDAR scans, imagery and satellite data into information about individual trees, treating forests as a demanding test bed for computer vision models.