Accurate estimation of forest aboveground biomass (AGB) is important for carbon accounting. Forest AGB estimation has been conducted with a variety of data sources and prediction methods, but many uncertainties still exist. In this study, six prediction methods, including Gaussian processes, stepwise linear regression, nonlinear regression using a logistic model, partial least squares regression, random forest, and support vector machines were used to estimate forest AGB in Jiangxi Province, China, by combining Geoscience Laser Altimeter System (GLAS) data, Moderate Resolution Imaging Spectroradiometer (MODIS) data, and field measurements. We compared the effect of three factors (prediction methods, sample sizes of field measurements, and cross-validation settings) on the predictive quality of the methods. The results showed that the prediction methods had the most considerable effect on the prediction quality. In most cases, random forest produced more accurate estimates than the other methods. The sample sizes had an obvious effect on accuracy, especially for the random forest model. The accuracy increased with increasing sample sizes. The random forest algorithm with a large number of field measurements, was the most precise (coefficient of determination (R2) = 0.73, root mean square error (RMSE) = 23.58 Mg/ha). Increasing the number of folds within the cross-validation settings improved the R2 values. However, no apparent change occurred in RMSE for different numbers of folds. Finally, the wall-to-wall forest AGB map over the study area was generated using the random forest model.
A Random Forest Model Based on Lidar and Field Measurements for Parameterizing Surface Roughness in Coastal Modeling
