Comparison of deep and shallow one-class classifiers for detecting invasive Prosopis trees in Kenya from airborne hyperspectral data

Prosopis spp. are globally widespread woody invasive plants that negatively impact biodiversity and rural livelihoods. Accurate distribution information is essential for understanding their ecological consequences and for effective conservation and land management. In this study, we evaluated the suitability of airborne hyperspectral data in the visible and near-infrared (VNIR) range for pixel-level mapping of Prosopis trees at the beginning of the dry season in Kenya. For the mapping, we compared the performance of five one-class classifiers, a type of weakly supervised method trained solely using labels for the target class. The classifiers compared were Biased Support Vector Machine (BSVM), Maximum Entropy (Maxent), Boosted Regression Tree (BRT), ITreeDet (with experiments using both 2D and 3D convolutional neural networks [CNNs]), and HOneCls (a 2D fully convolutional neural network). The models were trained using observations of 120 Prosopis trees across the study area, corresponding to 2561 pixels, along with 10,000 unlabeled tree pixels. Testing was done using data from five 1 ha study plots, totalling 14,637 tree pixels. To analyse feature importance we used a model-agnostic permutation approach. F1-score was highest for ITreeDet with 2D CNN (0.81), followed by MaxEnt (0.73), HOneCls (0.71), BSVM (0.68), and BRT (0.65). The critical hyperspectral features were consistent across the models and primarily in the blue and green region, with some importance in the red and red-edge region, while near-infrared showed low importance. The results highlight the effectiveness of VNIR hyperspectral data and one-class classification for mapping Prosopis trees in Kenya. The demonstrated methodology is a promising tool for high-resolution mapping of Prosopis trees, supporting ecosystem restoration and conservation efforts.