Pushing the ceiling of species-level mapping in a hyperdiverse tropical forest with multi-temporal segmentation and airborne hyperspectral classification
Journal:
bioRxiv
Published Date:
May 5, 2026
Abstract
Species-level maps of tropical forest canopies are needed for biodiversity monitoring, conservation planning, and carbon accounting, yet the structural complexity and species richness of these forests make remote classification challenging. Here we evaluate the limits and possibilities of a two-step mapping approach applied to hyperdiverse moist forest at the Paracou Field Station, French Guiana. First, we delineate individual tree crowns from ten repeat UAV RGB surveys using a CNN (Mask R-CNN) and combine predictions across dates via a temporal consensus-fusion method, improving mean segmentation F1 from 0.68 (single date) to 0.78 (ten dates) and covering approximately 86% of canopy area. Second, we classify the species of each crown from a single airborne hyperspectral acquisition (416--2500 nm, 1 m resolution) using several machine learning classifiers trained and tested on 3,186 field-verified crowns spanning 169 species (drawn from a labelled pool of 3,256 crowns across 239 species; see \cref{sec:representative_tree_species_classificati}). Linear Discriminant Analysis (LDA) achieved the highest accuracy (weighted F1 = 0.75), though performance was uneven: repeated cross-validation (20 x 5-fold) showed that on average 50 species (95% CI: 41--63) attained F1 >= 0.7 in any given fold, with 38 maintaining this level on average and 15 doing so reliably (>= 80% of folds), while many rare species with few training examples remained unclassifiable (macro-average F1 = 0.48). Combining segmentation and classification, we estimate that approximately 70% of the landscape's canopy area was correctly mapped to species. Band-importance and ablation analyses identified the far-red edge (748--775 nm) as the most informative spectral region, with secondary contributions from the red, green, and SWIR. While these results represent a substantial advance over previous studies limited to fewer than 20 species, we caution that accuracy is strongly conditioned by training data availability, site-specific spectral conditions, and the single-acquisition design, and that generalization to other sites and sensors remains to be demonstrated.
