UAV-based multi-spectral analysis and classification for detecting individual endangered Peirson’s milkvetch in a dune landscape

The use of unmanned aerial vehicles (UAVs) to monitor critically threatened and endangered species is increasingly important for agencies seeking efficient and repeatable methods to assess population status and trends. This 3 yr study evaluated imagery classification techniques using drones and machine learning methods to assess the feasibility of modeling individual Peirson’s milkvetch Astragalus magdalenae var. peirsonii (PMV) within the Imperial Sand Dunes Recreation Area (ISDRA), California, USA. The species is both state and federally listed due to its extremely limited distribution and impacts to all life stages by off-highway vehicle use in the heavily trafficked ISDRA. We used the random forest classification method by isolating the best performing variable combinations, including multispectral imagery bands, vegetation indices, and texture measures, to specifically identify individual PMV within the survey imagery. The vegetation maps produced an average classification accuracy of 93.25% for PMV identified and were validated with ground truth data within the southern range of survey plots collected. Comparisons of Bureau of Land Management ground survey results with our UAV results identified significant challenges associated with GPS accuracy between surveys, which may impact plant count totals from year to year. The inclusion of vegetation maps produced from UAV imagery identifying individual plants within surveyed areas would greatly enhance current methods used to quantify plant totals and aid in predicting trends in species distribution.