Abstract:
Grass quality as measured by leaf nitrogen (P) and phosphorus (N) plays a major role in understanding the distribution, densities and population dynamics of herbivores including livestock and wild herbivores. The aim of this study was to estimate and monitor grass N and P using Sentinel-2 derived spectral information and vegetation indices during the wet season for the selected period between 2017 and 2022 spanning a savanna ecosystem in the Kruger National Park area. Sentinel-2 satellite images were used as they provide images with high spatial, spectral and temporal resolutions. Field data were used where grass samples were collected, and spectral reflectance measurements (400 and 2350 nm) were undertaken. Three analysis scenarios were employed to estimate grass N and P in conjunction with classical regression and machine learning techniques. The scenarios included: (i) specific spectral bands, (ii) conventional and red edge-based vegetation indices (VIs) and (iii) a combination of VIs and spectral bands using the Stepwise Multiple Linear Regression (SMLR), Random Forest (RF) and Support Vector Machine (SVM) statistical models. Results showed that SMLR yielded the highest estimation accuracy based on a combination of bands and VIs for leaf N (i.e. Coefficient of determination (R2) = 0.69 and Root Mean Square Error (RMSE) = 0.14%, Relative Root Mean Square Error (RRMSE) = 6.73%, Mean Absolute Error (MAE) = 0.11) and for P based on a combination of VIs only (i.e. R2= 0.40 and RMSE= 0.04%, RRMSE = 34.322% and MAE = 0.04). This study confirms that the combination of red edge bands and VIs of Sentinel-2 data are crucial for accurately estimating biochemical concentrations in a savanna ecosystem. This study has significant implications for mapping and monitoring grass quality over large spatial extents.
