FOR‑species20K is the largest public LiDAR benchmark for tree‑species classification, aggregating 25 independently collected terrestrial, mobile and airborne laser‑scanning datasets covering 33 species. We introduce a multi‑resolution, multi‑view ConvNeXtV2 model that lifts the official hidden‑test accuracy from 79.5 % to 85.0 %, ranking first at submission time.
The hidden test is species‑balanced, so its overall accuracy behaves more like a macro metric than the long‑tailed public split. The random split places every source dataset in both training and evaluation. A logistic probe can recover source identity from out‑of‑fold model probabilities with 69 % accuracy, while geometric features identify the source above 90 % for two widely shared species; 18 of the 33 species appear in only one source.
In a source‑species‑cell holdout evaluated on 5 566 trees from eight eligible species that retain the same LiDAR type in other sources, the unweighted species mean drops from 82 % to 36 % and the tree‑weighted accuracy falls from 95 % to 60 %. Count‑matched retraining and experiments with two alternative backbones show that reduced training support alone does not explain the gap, and an external PureForest experiment reveals a smaller yet consistent grouped‑split effect. These findings confirm that the 85 % leaderboard score represents strong in‑distribution performance but not source‑disjoint generalization, motivating balanced metrics and source‑aware evaluation.
Review