A comparative assessment of multi-temporal Landsat 8 and machine learning algorithms for estimating aboveground carbon stock in coppice oak forests

Remote sensing of low biomass forests has challenges related to the contribution of soil and understory reflectance recorded by sensors, hampering accurate forest aboveground carbon (AGC) quantification. To improve Landsat-based AGC estimates in forests with low biomass, this study explored the use of multi-temporal Landsat 8 Operational Land Imager (OLI) derived spectral information in Zagros forests by testing four machine learning algorithms: support vector machine (SVM), boosted regression trees (BRT), random forest (RF) and multivariate adaptive regression splines (MARS). We selected two forest areas with different levels of human activity for AGC reference plots: un-degraded forest (UD) and highly-degraded forest (HD). The results of the study showed that the Landsat image acquired in the peak of the growing season (10 August) provided the best AGC estimates for the UD site, but that for the HD site, AGC estimates were not affected by the timing of the imagery. The comparison of different modelling methods demonstrated lower accuracies from BRT, considerably biased estimates from SVM, and generally robust results from the RF algorithm. Overall, the study demonstrated the utility of applying the free Landsat 8 OLI dataset to AGC estimation, in particular non-commercial forests in developing countries where little budget is allocated for management.