Hannah Victoria Herrero, Jane Southworth, Erin Bunting, Romer Ryan Kohlhaas, and Brian Child
Article first published online: 07 AUG 2019 Sensors
ABSTRACT: Southern African savannas are an important dryland ecosystem, as they account for up to 54% of the landscape, support a rich variety of biodiversity, and are areas of key landscape change. This paper aims to address the challenges of studying this highly gradient landscape with a grass–shrub–tree continuum. This study takes place in South Luangwa National Park (SLNP) in eastern Zambia. Discretely classifying land cover in savannas is notoriously difficult because vegetation species and structural groups may be very similar, giving off nearly indistinguishable spectral signatures. A support vector machine classification was tested and it produced an accuracy of only 34.48%. Therefore, we took a novel continuous approach in evaluating this change by coupling in situ data with Landsat-level normalized difference vegetation index data (NDVI, as a proxy for vegetation abundance) and blackbody surface temperature (BBST) data into a rule-based classification for November 2015 (wet season) that was 79.31% accurate. The resultant rule-based classification was used to extract mean Moderate Resolution Imaging Spectroradiometer (MODIS) NDVI values by season over time from 2000 to 2016. This showed a distinct separation between each of the classes consistently over time, with woodland having the highest NDVI, followed by shrubland and then grassland, but an overall decrease in NDVI over time in all three classes. These changes may be due to a combination of precipitation, herbivory, fire, and humans. This study highlights the usefulness of a continuous time-series-based approach, which specifically integrates surface temperature and vegetation abundance-based NDVI data into a study of land cover and vegetation health for savanna landscapes, which will be useful for park managers and conservationists globally.
Read the full publication at Sensors