Aboveground biomass estimation in forests with random forest and Monte Carlo-based uncertainty analysis

•Estimation of biomass by random forest combined with Monte Carlo simulations.•Using Monte Carlo method to reduce the error caused by residual variability.•Quantifying the effect of different variable combinations on model uncertainty. Estimates of aboveground biomass (AGB) in forests and the associated uncertainty are important pieces of information outlined in the United Nations Framework Convention on Climate Change. Traditional integrated surveying methods require high inputs of manpower and material resources, whereas remote sensing technology can be used to improve the efficiency of forest resources surveying at a lower measurement cost. Previous studies have primarily utilized spectral information and textural features as the main modeling factors for AGB prediction models, however, few studies have examined how different modeling factors affect prediction accuracy in detail. At the same time, uncertainties in AGB are usually determine by a data set producing a single prediction result, ignoring the model uncertainty due by residual variability in the prediction results. To solve the above problems, this study focuses on Milin County of the Tibet Autonomous Region of China as the research area and uses data from the Landsat Operational Land Imager, digital elevation models (DEM), and the national forest inventory (NFI) to extract and analyze spectral information, textural features, and terrain factors. Four regression models based on different combinations of variables were established using a combination of random forest and Monte Carlo simulation methods (RF-MC), and the uncertainties of each model were determined. The purpose of this paper is to study the influence of different types of variable factors on AGB inversion and to measure the model uncertainty resulting from changes in the variable factors. The coefficient of determination (R2), mean absolute error (MAE), and root mean square error (RMSE) were used to evaluate the fit and accuracy of the model, while the relative root mean square error (rRMSE) was used to quantify the uncertainty. The results show that the degree of fitting and accuracy are greatest for the RF-MC model based on three kinds of variables (spectrum and plot data + DEM + texture characteristics) the best (R2 = 0.85, RMSE = 11.76 t/ha, MAE = 8.49 t/ha, rRMSE = 5.97 %), which performed better than the RF-MC model based on two variables (spectrum and plot data + texture characteristics, R2 = 0.78, RMSE = 13.93 t/ha, MAE = 1