Multispectral UAV Fusion for Native Tree Species Recognition in the Cerrado

The Brazilian Cerrado, one of the world’s most biodiverse and threatened savannas, urgently requires scalable tools for identifying native tree species at the individual level. This work presents the first deep learning benchmark for this task, introducing a curated dataset of 1,620 UAV multispectral images covering 16 native Cerrado species, fully annotated by field biologists at an Ecological Reserve. We propose a Gram-Schmidt fusion pipeline that condenses RGB, Red Edge, and NIR bands into a pseudo-RGB representation, enabling unmodified object detectors to exploit near-infrared information. Eight models were evaluated across four modalities (fused, RGB, NDRE, NDVI). The fused modality consistently outperformed all others, with RT-DETR-L and YOLOv5s achieving 85.2% mAP@50. Per-species analysis revealed that near-infrared reflectance is critical for rare and morphologically ambiguous taxa, yielding gains exceeding 70 percentage points over RGB-only inputs for several species. These results demonstrate that multispectral fusion is not merely beneficial but essential for fine-grained Cerrado species discrimination, establishing a practical foundation for automated biodiversity monitoring in this ecologically critical biome. Open code and the dataset will be made available here after reviewing.

This paper is currently under review for CVBMC 2026