Abstract
Vegetation cover is a key environmental variable often mapped from satellite and aerial imagery. The derivation of ultra-high spatial resolution fractional vegetation cover (FVC) based on multispectral imagery acquired from an Unmanned Aerial System (UAS) has several applications, including the potential to revolutionise the collection of field data for calibration/validation of satellite products. In this study, abundance maps were derived using three methods, applied to data collected in a typical Australian rangeland environment. The first method used downscaling between Landsat FVC maps and UAS images with Random Forest regression to predict bare ground, photosynthetic vegetation and non-photosynthetic vegetation cover. The second method used spectral unmixing based on endmembers identified in the multispectral imagery. The third method used an object-based classification approach to label image segments. The accuracy of all UAS FVC and Landsat FVC products were assessed using 20 field plots (100 m diameter star transects), as well as from 138 ground photo plots. The classification method performed best for all cover fractions at the 100 m plot scale (12-13% RMSE), with the downscaling approach only able to accurately predict photosynthetic cover. The downscaling and unmixing generally over-predicted non-photosynthetic vegetation associated with Chenopod shrubs. When compared with the high-resolution photo plot data, the classification method performed the worst, while the downscaling and unmixing methods achieved reasonable accuracy for the photosynthetic component only (12-13% RMSE). Multispectral UAS imagery has great potential for mapping photosynthetic vegetation cover in rangelands at ultra-high resolution, though accurately separating non-photosynthetic vegetation and bare ground was only possible when the data was scaled-up to coarser resolutions.
NSTL
Elsevier