Drone-based multispectral differentiation of subalpine vegetation at the treeline in the Southern Alps of New Zealand

Abstract. Subalpine ecosystems are highly dynamic environments that are particularly vulnerable to environmental change, yet their remote and rugged nature poses challenges for long-term monitoring. Unoccupied aerial vehicles (UAVs) equipped with multispectral sensors offer a scalable solution for high-resolution vegetation mapping in these landscapes. In this study, we integrated UAV-derived spectral data with machine learning (ML) classifiers to assess the effectiveness of different vegetation indices (VIs) in distinguishing subalpine plant communities. Principal component analysis (PCA) revealed that NDVI, SIPI2, MCARI, and CHL were highly correlated and strongly influenced the primary variance in the dataset, while NDRE and LCI contributed more evenly across principal components, and GNDVI was largely independent. Among the ML classifiers tested, extreme gradient boosting (XGBoost) achieved the greatest overall accuracy (81.3 %) and Kappa (0.75), outperforming support vector machines (SVM) and random forest (RF). Classification confidence was highest for Chionochloa tussock (64.6–69.7 %) and Dracophyllum scrub (70.6 %), suggesting moderate reliability for these dominant vegetation types. Scrub and prostrate mat-forming communities exhibited lower classification accuracy, likely due to their heterogeneous canopy structure and greater spectral variability. The ecological boundaries of the subalpine zone, formed by Fuscospora forest and scree, were classified with high confidence, but the vegetation is dominated by tussock and shrubland. Feature importance analysis ranked NDVI, SIPI2, CHL, and MCARI highly in SVM and RF models, whereas LCI prevailed in XGBoost, underscoring how different algorithms leverage spectral information in classification tasks. These results emphasize the role of vegetation structure in classification accuracy, with dense, spectrally homogeneous vegetation types more reliably distinguished than mixed-species communities. Our study highlights UAV-based classification as a valuable tool for landscape-scale monitoring of subalpine vegetation. As UAV applications and ML workflows continue to evolve, optimizing classification approaches will enhance our ability to track ecological changes in subalpine and alpine regions worldwide.