Monitoring ecosystem dynamics in semi-arid environments using multi-sensor Earth-observation

Climate change and a growing human population are instigating major changes on the Earth's surface. Monitoring and understanding these changes as they unfold is critical for society and the environment. Satellite remote sensing provides the only means of achieving this over large spatial and temporal scales, and major progress in the application of Earth-observation imagery has been made since the beginning of the space age in the mid-20th century. However, savannahs - dynamic systems comprised of shrubs, trees, and grass species - have proved challenging for EO-based monitoring. Yet, these ecosystems cover almost 25% of the Earth's surface and are home to some of the poorest people on the planet. This thesis investigates the use of EO for monitoring ecosystem dynamics in African savannahs, focusing specially on woody cover and biomass provision. One of the most common Earth-observation (EO) based tools for monitoring vegetation is the Normalised Difference Vegetation Index (NDVI). A detailed review of the application of NDVI for monitoring land degradation was undertaken. This covered the historical context and ongoing debates around NDVI analyses, and highlighted key research gaps. NDVI was then used to map grass biomass for the Kruger National Park in South Africa, by combining in situ data with a downscaled NDVI dataset in a machine-learning framework. These predictions highlighted that the NDVI-biomass relationship is vulnerable to overfi�tting in space and time, due to spatial autocorrelation and a variable species composition, respectively. The NDVI was further explored at the continental scale using multiple time-series analyses. These revealed that a majority of African savannahs have only experienced vegetation greening in the 1982-2016 period. Areas of declining vegetation, or changes in the trend direction, were associated with phenological changes (i.e. a shrinking growth season), woodland degradation, or population increases. Finally, fractional woody vegetation cover was mapped for the Limpopo province of South Africa using Landsat spectral metrics and ALOS PALSAR radar imagery and a series of Random Forest regression models. The most accurate models combined multi-seasonal Landsat data and the radar layers. However, this was only marginally more accurate than just using dry and wet season metrics alone. When using a single season of imagery, the dry season preformed best. These results were reaffirmed for categorical savannah land-cover classifications, highlighting the importance of multi-sensor and multi-temporal data. The thesis contributes new insights for monitoring savannahs using EO imagery. By combining EO data with modern statistics and machine-learning methods novel insights to ecological and environmental issues can be gained. In the coming years, the increasing number of operational sensors and the volume of data collected will be of great benefit for environmental monitoring, especially in savannahs.