Search

Some invasive tree species threaten biodiversity and cause irreversible damage to global ecosystems. The key to controlling and monitoring the propagation of invasive tree species is to detect their occurrence as early as possible. In this regard, one-class classification (OCC) shows potential in forest areas with abundant species richness since it only requires a few positive samples of the invasive tree species to be mapped, instead of all the species. However, the classical OCC method in remote sensing is heavily dependent on manually designed features, which have a limited ability in areas with complex species distributions. Deep learning based tree species classification methods mostly focus on multi-class classification, and there have been few studies of the deep OCC of tree species. In this paper, a deep positive and unlabeled learning based OCC framework— ITreeDet —is proposed for identifying the invasive tree species of Eucalyptus spp. (eucalyptus) and Acacia mearnsii (black wattle) in the Taita Hills of southern Kenya. In the ITreeDet framework, an absNegative risk estimator is designed to train a robust deep OCC model by fully using the massive unlabeled data. Compared with the state-of-the-art OCC methods, ITreeDet represents a great improvement in detection accuracy, and the F1-score was 0.86 and 0.70 for eucalyptus and black wattle, respectively. The study area covers 100 km2 of the Taita Hills, where, according to our findings, the total area of eucalyptus and black wattle is 1.61 km2 and 3.24 km2, respectively, which represent 6.78% and 13.65% of the area covered by trees and forest. In addition, both invasive tree species are located in the higher elevations, and the extensive spread of black wattle around the study area confirms its invasive tendency. The maps generated by the use of the proposed algorithm will help local government to develop management strategies for these two invasive species. [ABSTRACT FROM AUTHOR]