Researchers at QUT and the Australian Antarctic Division employ UAV-mounted hyperspectral imaging combined with gradient boosting and convolutional neural network models to distinguish healthy, stressed, and moribund moss alongside lichen, rock, and ice in Antarctica. Their workflow integrates ground-based scans, GNSS RTK georeferencing, and custom spectral indices to achieve up to 99.8% accuracy in vegetation mapping under extreme polar conditions.

Key points

• UAV-mounted Headwall Nano-Hyperspec camera captures 400–1000 nm imagery over ASPA 135 with 4.8 cm/pixel GSD.
• Custom spectral indices (NDMLI, HSMI, MTHI) and PCA features feed XGBoost, CatBoost, and SE-UNet models, reaching weighted F1-scores up to 99.7%.
• Light-model variants using eight wavelengths (404–920 nm) achieve >95.5% accuracy, enabling rapid preliminary moss and lichen assessments.

Why it matters: This approach establishes a high-precision, scalable method for non-invasive vegetation monitoring in extreme environments, advancing conservation and climate research.