Towards comprehensive individual tree species mapping in diverse tropical forests by harnessing temporal and spectral dimensions

Monitoring tropical forest diversity and resilience requires taxonomically resolved, tree-scale data, but the structural complexity and species richness of tropical canopies make remote classification challenging. We present a scalable two-step approach to tree crown species mapping, combining CNN-based crown segmentation from multi-date UAV RGB imagery with species identification from high-resolution airborne hyperspectral data (416-2500 nm). Working in moist tropical forests in French Guiana, we improved crown delineation accuracy by evaluating consensus among ten segmented RGB surveys, increasing mean F1 from 0.69 (single date) to 0.78 (ten dates). Using 3,500+ hand-delineated, field-verified crowns, we classified 169 species with a frequency weighted F1 of 0.75, achieving F1 > 0.7 for 65 species. Together [~]70% of crown area at landscape-scale was accurately delineated and labelled. Rare species remained difficult to identify, underscoring the need for larger, targeted training datasets. Band-importance analyses isolated a narrow far-red-edge window (748-775 nm) as the most informative region for species discrimination, consistent with spectral convergence in biochemical, anatomical, and hydric leaf and structural traits. Species were easiest to classify when crowns produced a tight, well-separated spectral cluster at acquisition -- promoted by pronounced, synchronised leaf-flush cycles, minimal liana load, and few close relatives -- indicating signal degradation, rather than specific trait identity, governs separability. Spatial dispersion of conspecifics improved accuracy, suggesting learned features generalised, whereas clustering amplified local artefacts and reduced separability. Our results demonstrate that combining multi-temporal crown segmentation with hyperspectral imaging can map species-level canopy composition in complex tropical forests with unprecedented scope, and highlight spectral-phenological features critical for scaling biodiversity monitoring from airborne to satellite platforms.