Retrieval of aboveground grass carbon stocks in savannah ecosystems using SAR data products.

Abstract

Savannah grassland is one of the world’s most extensive biomes, covering approximately 25% of the Earth’s surface. The proliferation of natural and human-induced activities, such as grass sward removal, fire occurrence and climate change, deprives the ecosystems’ ability to render services such as climate regulation. Degraded savannah grasslands tend to sequester and store carbon at a slower rate than usual. Consequently, high levels of greenhouse gases such as carbon dioxide (CO2) accumulate in the atmosphere at a faster rate than normal. Savannah grasslands sequester and store carbon in their aboveground plant matter, aboveground litter as well as in belowground matter. Fifty percent (50%) of the aboveground grass biomass (AGGB) comprises carbon. As such, concerns about global warming have sparked studies on ways to limit the accumulation of CO2 in the atmosphere. The decline in capacity to store carbon contributes to elevated atmospheric CO₂ levels, thereby accelerating global warming. However, limited understanding exists regarding how specific degradation factors affect the carbon storage potential of AGGB. This gap hinders the development of effective mitigation strategies aimed at enhancing the carbon sequestration function of savannah grasslands. Therefore, the development and implementation of aboveground grass carbon stock (AGGCS) quantification methods and ultimately monitoring operations in savannahs are necessary to determine if the ecosystem is acting as a carbon source or sink. As such, remote sensing tools coupled with scalable machine learning and artificial intelligence offer potential for effective quantification and monitoring of AGGCS. The purpose of this study was to determine if the Synthetic Aperture Radar (SAR) data products can quantify aboveground grass carbon stocks in savannah ecosystems with considerable predictive accuracy. The scope of this study was limited to (1) reviewing the progress and challenges in the retrieval of AGGCS in Savannah ecosystems using remote sensing, (2) determining the optimal SAR parameters suitable for AGGCS in Savannah ecosystems and further determining if incorporating ancillary variables can boost the predictive performance of SAR-based models in quantifying AGGCS in Savannah ecosystems, (4) evaluating intraseasonal variations of AGGCS in savannah ecosystems using SAR data (5) assessing the impacts of seasons on the predictive performance of SAR data in quantifying the seasonal variation of AGGCS in Savannah ecosystems. The Kruger National Park (KNP) in the Mpumalanga Province of South Africa served as an experimental site for extracting grass samples required to achieve the above-mentioned objectives. Ground observation and grass biomass measurements were taken across the wet and dry seasons of 2020 and 2021. To achieve this, Sentinel-1 was used to derive radar indices and texture matrices to quantify AGGCS. Next topographic variables were extracted from the SRTM (Shuttle Radar Topography Mission) Digital Elevation Model (DEM) and added into the model. The XGBoost algorithm was utilised to predict the aboveground grass carbon stocks and computed the regression coefficient (R²) and error matrices such as MAE (Mean Absolute Error), RMSE (Root Mean Square Error) and RMSE% (Percent Root Mean Square Error) to assess the model performance. The study found that SAR-derived parameters alone do not fully capture the variability in aboveground grass carbon stock, particularly in the complexly configured savannah ecosystems. The study highlighted the Radar Vegetation Index (RVI) as the most sensitive and consistent parameter for quantifying AGGCS within and across seasons. The study also found that the performance of SAR-based models can be improved with the incorporation of texture and topographic variables. However, when testing the models within seasons and across the years, the study found that the contribution of SAR, texture and topographic parameters varied between the wet and dry seasons in both years. Notably, the predictive capabilities of the models also differed within the seasons and across the years. Overall, RVI, Slope, Terrain Ruggedness Index (TRI), Aspect, Vertical Transmit/Horizontal Receive (VHcor), Topographic Position Index (TPI), Topographic Wetness Index (TWI), Lat, Vertical Transmit/Vertical Receive (VVcor) and Hill shade optimally predicted AGGCS. The XGBoost models that integrated all the variables (topographic, texture and SAR indices) showed the highest predictive power in both the wet and dry seasons across both years. Notably, the model developed for the dry season exhibited superior predictive power (R2 = 0.86, RMSE% = 13.30 and MAE = 2.81) compared to that of the wet season. The findings underscore the dynamic and variable characteristics of ecosystem processes within savannah environments and their impact on the predictive power of remote sensing-based models. Models that are calibrated for a single season may effectively capture specific conditions, which could hinder their transferability across other temporal scales with different environmental conditions. Overall, the findings obtained in this study underscore the importance of SAR sensors such as Sentinel-1 as proxies for quantifying aboveground grass carbon stocks in complex savannah ecosystems. These findings have implications for informing ecosystem conservation strategies and climate change modelling, and therefore, support Goal 13 of the Sustainable Development Goals (SDG), which calls for immediate action to combat climate change and mitigate its impacts.