Modelling thicket above-ground biomass using LiDAR and SAR data, in the Addo Elephant National Park

Abstract

Due to the ecological effects arising from climate change, it is necessary to quantify terrestrial carbon stocks at various spatial and temporal scales for mitigation purposes. Whereas deforestation is known to increase carbon dioxide emissions and negatively impact the fight against global climate change, studies have also shown that increased carbon results in woody thickening and the transformation from open savanna to bush encroached thickets. Hence, this study sought to quantitatively model the Above Ground Biomass (AGB) within the Addo Elephant National Park in South Africa. The dominant vegetation type of the park is the unique and threatened Albany thicket. Field-based AGB measurements and LiDAR metrics were used to derive a local AGB model that represented the Addo thicket vegetation type, and could then be upscaled using regional-scale satellite-based Synthetic Aperture Radar (SAR) backscatter values. Furthermore, this study sought to determine the performance of Sentinel-1 C band SAR data versus ALOS PALSAR-2 L band SAR data in retrieving the thicket AGB. Specific objectives to the study were to use field-collected AGB plots and airborne LiDAR data to develop and apply a LIDAR-based thicket AGB model to the LiDAR dataset; to use the LiDAR-based AGB product as calibration and validation data in exploring different modelling techniques for estimating regional scale SAR-based AGB and to investigate the effects of SAR wavelength in retrieving thicket AGB of the Albany thicket biome by comparing Sentinel-1 multi-temporal model against the Global mosaic ALOS PALSAR-2 L band model. The coefficient of determination (R2 ), the Root Mean Square Error (RMSE) and the relative Root Mean Square Error (rRMSE) were used to evaluate and compare the performance of the models. Field-based AGB measurements were successfully upscaled using LiDAR metric and the best Simple Linear Regression (SLR) model established was AGB= 0.258 (CC * MH) with an R2 of 89%, RMSE of 3.779t/ha and rRMSE of 35.6%. Then Simple linear regression and Random Forest modelling techniques were used to derive a regional AGB model. The Random Forest model yielded respectable results (R2 of 67% and rRMSE of 38.4%) and outperformed the Simple linear regression model. Furthermore, the time series of Sentinel-1 C band data showed higher retrieval results (give best result, R2 & rRMSE) than the one-off ALOS PALSAR-2 L band data (give best result, R2 , rRMSE) in retrieving AGB using both Random Forest and Simple Linear Regression techniques. Importantly, the study demonstrated the effective use of machine learning predictive mapping techniques for thicket AGB using freely available Sentinel-1 C band remote sensing data by taking advantage of its relatively high spatial and temporal resolution.