Mapping melliferous tree species in Kenya via one-class classification with hyperspectral unsupervised domain adaptation

Mapping the spatial distribution of specific nectar-producing (melliferous) tree species is critical for sustainable beekeeping in Kenya's arid and semi-arid regions. However, traditional field surveys are labor-intensive, and existing remote sensing classification models often fail when applied to new, unseen landscapes due to domain shift—where environmental or phenological differences alter the spectral signatures of the same tree species.

The researchers developed the Hyperspectral Unsupervised Domain Adaptation One-Class Classification (HyUDA-One) framework. This method uses a positive-unlabeled (PU) learning approach, which allows the model to identify a target species without needing extensive labeled data for every other species in the landscape. To handle domain shift, the framework incorporates a spatial-spectral regularized pseudo-positive learning (SSPPL) strategy. This allows the model to iteratively adapt its decision boundaries to the unlabeled target domain, effectively aligning the feature distributions between the source and target sites without requiring new ground-truth labels.

The framework was tested on three key melliferous species—Senegalia mellifera, Vachellia tortilis, and Commiphora africana—across two distinct savanna sites in southern Kenya. In the trained domain, the model achieved high F1-scores (0.788, 0.845, and 0.768, respectively). Crucially, when applied to an entirely untrained domain, the model maintained strong performance, achieving F1-scores of 0.756 for S. mellifera and 0.884 for V. tortilis. These results demonstrate that the HyUDA-One framework significantly improves generalization in label-scarce environments, providing reliable distribution maps that can guide local beekeeping development.

By enabling accurate species-level mapping without the need for extensive, costly field data collection in every new area, this framework offers a scalable tool for environmental management. It supports sustainable apiculture by identifying nectar source availability and has broader potential applications in monitoring invasive species or managing biodiversity in complex, data-poor ecosystems.