Comparison of satellite-based estimates of aboveground biomass in coppice oak forests using parametric, semiparametric, and nonparametric modeling methods

Accurate estimates of forest biomass are essential for several purposes, ranging from carbon accounting and ecological applications to sustainable forest management. There are, however, critical steps for mapping aboveground forest biomass (AGB) based on optical satellite data with an acceptable degree of accuracy, such as selecting the proper statistical modeling method and deriving spectral information from imagery, at known field locations. We compare nine modeling techniques including parametric, semiparametric, and nonparametric methods for remotely estimating AGB based on various spectral variables derived from Landsat 8 Operational Land Imager (OLI). We conduct this research in Zagros oak forests on two sites under different human disturbance levels: an undegraded (UD) forest site and a highly degraded (HD) forest site. Based on cross-validation statistics, the UD site exhibited better results than the HD site. Support vector machine (SVM) and Cubist regression (CR) were more precise in terms of coefficient of determination ( R 2 ), root-mean-square error (RMSE), and mean absolute error (MAE), though these approaches also result in more biased estimates compared to the other methods. Our findings reveal that if the degree of under- or over-estimation is not problematic, then SVM and CR are good modeling options ( R 2   =  0.73; RMSE  =  31.5  %   of the mean, and MAE  =  3.93  ton  /  ha), otherwise, the other modeling methods such as linear model, k -nearest neighbor, boosted regression trees, generalized additive model, and random forest may be better choices. Overall, our work indicates that the use of freely available Landsat 8 OLI and proper statistical modeling methods is a time- and cost-effective approach for accurate AGB estimates in Zagros oak forests.