Maisha Maliha, Megan E. Cattau, Stella M. Copeland, Megan Dolman, Valorie Marie, Peter J. Olsoy, Richard Rachman, Amethyst Tagney, Sandra Velazco, Ryan Wickersham, Andrii Zaiats, T. Trevor Caughlin
Using Structure-from-Motion photogrammetry and a stacked ensemble learning approach to classify plant species in sagebrush steppe landscapes. Achieving a mean classification accuracy of 92.1% and a weighted F1 score of 91.5% despite a highly imbalanced dataset. Indicating strong performance of UAV-based machine learning models for plant species classification with limited training data.
Monitoring plant species composition at management-relevant scales remains a persistent challenge in rangeland ecosystems. High-resolution imagery from unoccupied aerial vehicles (UAVs) offers a promising solution, but high sensor costs, the need for extensive field training data, and the difficulty of covering large areas have hindered widespread adoption. We evaluated the performance of UAV-based machine learning models for plant species classification and transferability to sites without training data across 14 big sagebrush ( Artemisia tridentata L.) steppe landscapes in the northern Great Basin, USA. We tested models for 18 common overstory plant species in different functional groups including shrubs, Artemisia arbuscula Nutt. (low sagebrush) and Ericameria nauseosa (Pall. ex Pursh) G.L. Nesom & Baird (rubber rabbitbrush), and the perennial bunchgrass Pseudoroegneria spicata (Pursh) Á. Löve (bluebunch wheatgrass). Using Structure-from-Motion photogrammetry and a stacked ensemble learning approach, we achieved a mean classification accuracy of 92.1% (95% CI, 90.9–93.2%) and a weighted F1 score of 91.5% (95% CI, 90.2–92.8%), indicating strong performance despite a highly imbalanced dataset dominated by a few common species. Models trained with low-cost red, green, and blue imagery performed nearly as well as those using multispectral data, with only about a 1% difference in F1 score. However, model transferability was limited: classification accuracy declined sharply at sites where species composition differed from training data, with F1 scores ranging from <0.09 to >0.90 across test sites. These results suggest that although low-cost UAVs can produce accurate and scalable species maps, reliable application across diverse rangelands will require strategic field sampling or shared training datasets. Our findings provide practical guidance for researchers and land managers seeking to incorporate UAV technology into biodiversity monitoring, restoration planning, and invasive species management.