Fusion of Sentinel-2 Data with High Resolution Open Access Planet Basemaps for Grazing Lawn Detection in Southern African Savannahs

Abstract

Short grass grazing lawn patches are significant components of habitat heterogeneity in southern African savannah ecosystems. Accurate maps of grazing lawn distribution is essential to enhance understanding of important ecosystem processes such as mega-herbivore population dynamics, nutrient cycling and plant community composition. The inherent heterogeneity of savannah landscapes however creates significant challenges for accurate discrimination of vegetation components and thus grazing lawn detection. Recent studies favour very high spatial resolution (VHR) multi-spectral imagery for dealing with this challenge. However, such data are costly for use in operational management. Planet Labs, through Norway s International Climate and Forests Initiative (NICFI), now grant free access to high-resolution, analysis-ready mosaics over the tropics, with great potential for fine-scale vegetation mapping. However, the spectral characteristics of these data are limited and fail to resolve the spectral similarity of different savannah vegetation components. We address these issues using Gram-Schmidt transformation to fuse Planet Basemaps and Sentinel-2A images for grazing lawn detection within the Lower Sabie region of Kruger National Park, South Africa. The original and fused images were classified using a random forest approach. Overall, the fused image achieved the best grazing lawn detection accuracy (0.85) and general map accuracy (0.72) results compared to Sentinel-2 (0.67 and 0.62) and Planet basemap (0.64 and 0.62 respectively). Our findings provide a foundation for cost-effective and accurate high spatial resolution vegetation mapping in heterogenous savannah landscapes. Further studies will investigate the potential of multi-temporal fused data and object-based approaches for enhanced savannah vegetation mapping