Speaker
Description
Approximately more than 50% of Africa’s land surface is covered by the savanna biome. Estimation of savanna vegetation are essential to provide detailed land cover information. Prediction of savanna vegetation at pixel level is important for conservation purposes, understanding bush encroachment, herbivory and fire dynamics in savannas. However, estimating vegetation cover in the savannas can be challenging due to the dynamic factors such as climate, fires and herbivory which interact with the savannas at different scales contributing to high cover heterogeneity and the mixed pixel problem.
Sentinel-1 C-band is a radar remote sensing which can ideally complement the optical data which are affected by weather. SAR has especially been used for forest mapping, deforestation and forest degradation. Radar backscatter sensitivity to forest structural change offers new methods especially where optical time series are not available. Fewer studies which combine radar and optical for mapping savanna vegetation fractions are reported. This is despite that increased accuracies are reported when radar and optical data are combined for savanna vegetation mapping. Therefor combining radar and optical remote sensing data presents an opportunity to explore methods for estimating African savanna vegetation fractions.
Mixed pixel analysis quantifies the proportions of vegetation within a single pixel. This poster presentation presents results on the use of multi-temporal Sentinel-1 VV, VH and VV Coherence combined with Sentinel-2 data multi-temporal statistics to estimate woody vegetation fractions in Benfontein Nature Reserve project in South African savanna environment using machine learning regressions, with fieldwork supported by South African Environmental Observation Network (SAEON) and in conjunction with South African Land Degradation Monitor (SALDI). Training data is collected by combining fieldwork and data collected from Very High Resolution (VHR) images, along with Analysis Ready Data (ARD) Sentinel-1 and Sentinel-2 multi-temporal statistics which are used as predictors and applied to machine learning regressions to estimate woody vegetation and grass combined with baresoil fractions. The analysis is performed at a pixel level to estimate the percent of vegetation within each each pixel. A combination of Sentinel-1 and Sentinel-2 performs slightly better than Sentinel-1 and Sentinel-2 alone. The lowest accuracies are observed for Sentinel-1 models alone.
