Abstract
Gross Primary Productivity (GPP) is a key indicator of ecosystem functioning and carbon sequestration. GPP has been estimated using several tools, including remote sensing, one of which is the Moderate Resolution Imaging Spectroradiometer (MODIS). Despite its coarse spatial resolution limiting effectiveness in capturing finer landscape-scaled ecosystem and carbon dynamics, this tool is widely used. This study aims to develop the first high-resolution GPP product by downscaling MODIS GPP to Sentinel-2 spatial resolution using machine-learning algorithms and multi-source predictors in the Golden Gate Highlands National Park (GGHNP) in the eastern Free State of South Africa.
The objectives of the study were to (a) determine the relationship between GPP and selected predictor variables (viz.., Sentinel-2 spectral bands, vegetation indices, and topographical variables); (b) evaluate the performance of Random Forest (RF) and Support Vector Machine (SVM) algorithms in downscaling the MODIS GPP product to Sentinel-2 resolution under the selected predictor combinations; and (c) identify important predictors using RF variable importance analysis. The results showed strong positive relationships between GPP and the selected vegetation indices (r >0.6), including the Chlorophyll Red-edge Index (CI-Red-Edge), the Modified Soil-adjusted Vegetation Index (MSAVI), the Normalised Difference Red-edge Index (NDRE), and the Green NDVI (GNDVI). In contrast, the blue and green spectral bands were negatively correlated with GPP (r similar to-0.6). Moreover, the modelling results showed that the SVM model, comprising bands and indices (R-2 = 0.68, RMSE = 10.07 g C m(-2)), was superior to the RF model (R-2 = 0.60, RMSE = 10.2 g C m-2), but inferior when topographical variables were combined with spectral bands and vegetation indices. The key predictors identified through variable importance analysis included the blue and green bands, the Enhanced Vegetation Index (EVI), the CI-Red-edge, MSAVI, and NDRE indices. These results highlight the value of the high-resolution bands for Sentinel 2, as well as the red-edge and near-infrared bands, in modelling vegetation productivity. This study demonstrates the potential of combining high-resolution remote sensing data with machine-learning algorithms to estimate GPP in data-scarce mountainous regions and provides the first localised model for generating vegetation productivity estimates in nature reserves such as the GGHNP. Overall, the approaches offer a scalable tool for improving ecological monitoring and understanding carbon dynamics in mountainous environments.