Abstract
Woody canopy cover (CC) is important for characterising terrestrial ecosystems and understanding vegetation dynamics. The lack of accurate calibration and validation datasets for reliable modelling of CC in the indigenous forests in South Africa contributes to uncertainties in carbon stock estimates and limits our understanding of how they might influence long-term climate change. The aim of this study was to develop a method for monitoring CC in the Dukuduku indigenous forest in South Africa. Advanced Land Observing Satellite (ALOS) Phased Arrayed L-band Synthetic Aperture Radar (PALSAR) global mosaics of 2008, 2015, and 2018, polarimetric features, and Grey Level Co-occurrence Matrix (GLCMs) were used. Machine learning models Random Forest (RF) vs Support Vector Machines (SVM) were developed and calibrated using Collect Earth Online (CEO) data, a free and open-access land monitoring tool developed by the Food and Agriculture Organisation (FAO). The addition of GLCMs produced the highest accuracy in 2008, R 2 (RMSE) = 0.39 (36.04%), and in 2015, R 2 (RMSE) = 0.51 (27.82%), and in 2018, only SAR variables gave the highest accuracy R 2 (RMSE) = 0.55 (29.50). The best-performing models for 2008, 2015, and 2018 were based on RF. During the ten-year study period, shrubland and wooded grassland had the highest transition, at 6% and 13%, respectively. The observed changes in the different canopies provide valuable insights into the vegetation dynamics of the Dukuduku indigenous forest. The modelling results suggest that the CEO calibration data can be improved by integrating airborne LiDAR data.