Estimating aboveground forest carbon density using Landsat 8 and field-based data: a comparison of modelling approaches

Abstract

ABSTRACTAccurately estimating and mapping forest aboveground carbon density (AGCD) is important for evaluating forest carbon sequestration efficiency and dynamics. This study compares different modelling approaches for estimating AGCD using Landsat 8 Operational Land Imager (OLI) imagery and field forest inventory data. We tested it in Yongshun County, Hunan Province, China. Linear regression (LR) and Random Forest (RF) were used to map crown density (CD). We fitted models with and without CD dummy variables for AGCD estimations. The fitting results and performances of these two types of AGCD models were analysed and compared. There was a difference in the estimation of remote-sensing-based CD models and AGCD models; the coefficients of determination (R2) were 0.66 for LR and 0.84 for RF, respectively, in CD estimations; R2 values were 0.46 for LR and 0.65 for RF, respectively, in AGCD estimations as indicated by basic AGCD models. RF was the best algorithm for CD and AGCD estimations. All of the dummy variable AGCD models provided more accurate AGCD estimates than their basic AGCD counterparts. The dummy variable models (DLR and DRF), which took the observed CD as a dummy variable, were the most accurate for the R2 were 0.61 and 0.85, respectively. The AGCD estimations were significantly overestimated and underestimated by basic AGCD models in thin and dense CD classes, respectively. Nevertheless, none of the AGCD estimates produced by any of the dummy variable models showed this problem. The models that used predicted CD as a dummy variable were more consistent with those that considered observed CD as a dummy variable. The accuracies of AGCD estimations were significantly improved by the dummy variable AGCD models, especially in thin and dense plots; in addition, the CD estimations produced by CD models could be used as dummy variables in AGCD remote-sensing-based models. This work provides that a method considering CD dummy variable in remote sensing-based models was reliable for mapping AGCD at a regional scale.