Evaluating the impact of field-measured tree height errors correction on aboveground biomass modeling using airborne laser scanning and GEDI datasets in Brazilian Amazonia

Abstract

Forest field inventory (FFI) data provide valuable reference estimates of aboveground biomass (AGB) at the plot level, forming a basis for developing AGB prediction models that can be scaled to larger extents using predictor variables derived from remote sensing datasets e.g., LiDAR. Historical FFI datasets typically include tree diameter at breast height (DBH) and, in some cases, tree height ($H_{tree})$. Allometric equations yield more accurate AGB estimates when $H_{tree}$ is incorporated; however, while DBH is commonly recorded, $H_{tree}$ is often partially available or entirely missing from forest field plots. An alternative approach uses DBH as a predictor variable to estimate $H_{tree}$ through $H_{tree}$ – DBH allometric model. In this study, we present a framework to harmonize and incorporate existing yet inconsistent FFI datasets in AGB modeling at the regional scale. We optimized $H_{tree}$ – DBH allometric model based on the previously developed pantropical model of the Western Amazon using existing FFIs data. For this study, we used data from 174 forest field plots each measuring 50 m by 50 m, and coincident with airborne LiDAR data in the Brazilian Legal Amazon (BLA) region, South America. Using existing field-measured $H_{tree}$, we calibrated the H-DBH model to reflect regional conditions, resulting in an RMSE of a maximum of 6 m for trees with unknown $H_{tree}$. We then assessed tree height over- and under-estimations by using a 1-m canopy height model (CHM) originating from airborne laser scanning (ALS) as an explicit concurrent unbiased proxy dataset. The results indicate that under tropical forest conditions – BLA region, field measured $H_{tree}$ is generally underestimated when exceeding 30 m, particularly in dense forest canopies. Under-estimation is rarely observed in degraded forests, where over-estimation may occur if forest conditions have changed post-FFI (e.g., due to burning or logging). Following height correction, we applied allometric equations to estimate AGB using simulated GEDI waveform metrics—specifically relative height metrics such as RH5, RH10, RH15, through RH100—as predictor variables, validated against field-measured AGB from FFI data. We evaluated AGB estimates before and after tree height correction, using three machine learning models—Cubist, Random Forest, and XGBoost—to compare performance. Random Forest produced the most accurate AGB estimates in both harmonized and non-harmonized scenarios. This article makes three primary contributions: (a) optimizing the H-DBH allometry model with existing datasets, (b) estimating and harmonizing tree height to address over- and under-estimation issues in FFI data, and (c) evaluating the impact of $H_{tree}$ discrepancies on AGB modeling. The proposed framework provides a baseline for the quantitative use of FFI datasets in AGB modeling, highlighting biases in field datasets and their implications for AGB estimation. For this study, we used data from 174 forest field plots in the BLA region, South America, each measuring 50 m by 50 m. Our findings offer valuable insights for other tropical regions where tree height estimates are challenging, contributing to more reliable AGB quantification.