Researchers have released FOR-species20K, described as the largest public LiDAR benchmark for tree-species classification. The benchmark pools 25 independently collected datasets spanning 33 species, gathered with terrestrial, mobile, and airborne laser-scanning systems. This variety is meant to test how well models trained on one set of scanning conditions can recognize species in another.
The paper's central concern is cross-source generalization: LiDAR data differ by sensor, scan angle, and density, so a model that works on one dataset often fails on another. To tackle this, the authors introduce a multi-resolution approach, though the abstract does not detail the method. The benchmark itself is the main contribution, providing a shared testbed for future work on robust tree-species recognition.