Improved random forest algorithms for increasing the accuracy of forest aboveground biomass estimation using Sentinel-2 imagery

A simpler, unbiased, and comprehensive random forest (RF) model is needed to improve the accuracy of aboveground biomass (AGB) estimation. In this study, data were obtained from 128 sample plots of Pinus yunnanensis forest located in Chuxiong prefecture, Yunnan province, China. Sentinel-2 imagery data were applied to extract the important predictors of forest AGB, which were screened using the Boruta algorithm. We compared the fitting performance of two modified random forest models − regularized random forest (RRF) and quantile random forest (QRF) − with the random forest model. Moreover, we combined the smallest mean error of each quantile model as the best QRF (QRFb). The result showed: (1) Window sizes of 3 × 3 pixels and 5 × 5 pixels demonstrated greater sensitivity and suitability for estimating AGB than the 7 × 7 pixels window size. Enhanced vegetation indices derived from Red Edge 1 (B5) and Near-Infrared bands (B8A) were strongly correlated with AGB, indicating the heightened sensitivity of B5 and B8A bands to biomass and their potential in AGB estimation. (2) The RRF model outperformed both the standard RF and QRF in fitting performance, with an R2 of 0.56 and RMSE 57.14 Mg/ha. (3) The QRFb model exhibited the highest R2 of 0.88 and lowest RMSE of 29.56 Mg/ha, significantly reducing overestimation and underestimation issues. The modified RF regression supplies new insights into improving forest AGB estimation, which will be helpful for future research addressing carbon cycling.