Globally, semi-arid savannas are being affected by degradation processes. This causes a significant decline in ecosystem functioning and thus the use thereof to communities depending on savannas for their livelihoods. As a step towards a fast and cost-effective assessment of this degradation at the local to regional scale, we developed a change detection approach that uses Landsat TM time series to detect bush encroachment and vegetation loss. This approach uses spectral change characteristics as well as spatial change features derived from spectral change derivatives. The resulting decision tree classifier is sensitive to gradual increases or decreases and total loss of vegetation cover related to natural and land use related processes, while being independent of vegetation type. Results show that rainfall related vegetation phenology can bias the detection of gradual vegetation change (bush encroachment), while detecting vegetation loss processes reliably.