Vegetation cover is a key parameter in analyzing the state and dynamics of ecosystems. Africa's semi-arid savanna's are particularly prone to degradation, due to increasing population pressure as well as ongoing climatic
changes. In most global land cover classifications inhomogeneous areas are aggregated into few discrete classes,
delivering unsatisfying results in highly variable biomes, especially savanna's with their small scale patches of
woody and herbaceous vegetation and bare soil. Fractional cover(FC) classifications, which provide an estimate
of sub-pixel continuous cover percentages of underlying land cover classes, and are therefore an improved thematic representation, can deliver additional information for monitoring and decision making. Prior research
demonstrated that multi-scale approaches are suitable for transferring en-detail information from a small subset
to a larger study area via statistical up-scaling (e.g. Random Forest). In this case study the robustness of this
up-scaling approach and the limits of the spatial and temporal transferability at the very high and intermediate
resolution were analysed in the Caprivi Strip in Namibia and the adjacent Western Province of Zambia. The key
research questions were to quantify i) the robustness of the upscaling, ii) the loss of accuracy depending on the
lag in image acquisitions, iii) the loss of accuracy dependent on the time of image acquisition in the phenological
cycle. To this end 12 Worldview(WV) and all usable Landsat TM and ETM+ images, covering all phases of the
vegetation cycle were obtained. The analysis showed that continuous FC mapping is a highly suitable concept
for semi-arid ecosystems with gradual transitions. The optimal time for WV acquisition was at the beginning
of the dry season. The RMSE was unusable for LS images recorded in the rainy season between November and
March, but otherwise it was usable even for larger lags up to a month, with deviations below 15%. As long as the
spatial training subset(s) cover the whole occurring range of vegetation densities, comparably small WV scenes
are sufficient to reliably scale to regional results.
|