Estimation of Above-ground Biomass in a Semi-arid savannah Grassland Applying Machine Learning Algorithms with SAR and MSI Sentinel datasets

Grasslands and savannahs are primarily used to feed livestock and recognised for their essential ecosystem services. Aboveground biomass (AGB) is an important indicator for grassland productivity. Yet, AGB does not guarantee forage quality for livestock. To redress grassland degradation and design appropriate strategies for sustainable grazing management, data-driven decision support tools are required. Earth observation systems and machine learning approaches have become indispensable for mapping and estimating the accurate spatial distribution of AGB. This study examined the synergistic use of Sentinel-1 Synthetic Aperture Radar (SAR) polarization bands and Sentinel-2 Multi-spectral Instrument (MSI)-derived spectral vegetation indices to predict and map AGB spatial patterns at the selected protected livestock grazing area in South Africa. A total of 90 ground-truth AGB samples were collected and partitioned into 80% training and 20% testing input datasets for machine learning models. The cross-validation (k = 10) method was used for model performance. Three machine learning algorithms were applied to establish their predictive performances. The findings indicate that random forest regression (RFR) (R² = 0.347, RMSE = 9.47 Mg) yielded slightly better predictive performance than gradient boosting regression (GBR) (R² = 0.277, RMSE = 9.97 Mg) and extreme gradient boosting regression (XGBR) (R² = 0.173, RMSE = 10.66 Mg) models in predicting AGB. The low predictive capacity of these models may be linked to the sensitivity of grassland senescence phenological stage during the experiment. Furthermore, the Enhanced Vegetation Index (EVI) and Difference Vegetation Index (DVI) were constant higher important predictor variables across RFR, BGR, and XGBR machine learning models. These findings may be useful for rotational grazing management and monitoring seasonal productivity in semi-arid grassland.