Aboveground biomass modeling using simulated Global Ecosystem Dynamics Investigation (GEDI) waveform LiDAR and forest inventories in Amazonian rainforests

Abstract

NASA's Global Ecosystem Dynamics Investigation (GEDI) mission one of the objectives is to estimate global forest aboveground biomass (AGB) using full waveform (WF) LiDAR data. GEDI's relative height (RH) metrics, derived from vertical energy distributions, serve as key predictors in AGB modeling, with energy quantiles ranging from 0 % to 100 %. Despite extensive studies on RH metrics, the selection of optimal RH metrics for AGB estimation remains inconsistent, and using fewer metrics can result in a loss of vertical structural complexity. This study explores the potential of dense sampling of RH metrics (RH5 to RH100, in 5 % increments) to retain forest structural complexity, even across diverse forest regimes. Using noise-free simulated GEDI WF data, we developed machine learning models (Cubist, Random Forest, and XGBoost) to estimate AGB across 174 forest plots in the Brazilian Amazon. Results showed that dense RH sampling outperformed models using fewer recommended RH metrics. Our proposed suite of mean RH (mRH) metrics (R² = 0.71, RMSE = 65.88 Mg/ha, nRMSE = 0.36) – derived at plot level from an extensive suite of RH metrics (RH5 to RH100, in 5 % increments) at sub-plot level, and vertical mean RH (vmRH) RH metrics within the 20 % waveform vertical energy distribution (vmRH20, vmRH40, vmRH60, vmRH80, and vmRH100) approach showed similar performance, at the plot level of an average size of 50 m by 50 m. The single vmRH metrics versus plot-level AGB estimates – vmRH80 consistently gives the best results for all ML models and Ordinary Least Square (OLS) regression with R² ranges from (0.65–0.68), RMSE (53.18 – 70.51) Mg/ha – highest RMSE reported for OLS regression. All model’s performances were comparable giving similar RMSE, nRMSE, and coefficient of determination (R²) for derivative RH metrics – mRH and vmRH – compared with the traditional approach of selective RH metrics at GEDI footprint level estimates. The trained model provided AGB estimates at 30 m resolution for entire ALS survey areas of sites (n = 174) in the Brazilian Legal Amazon (BLA) region. Overall, this approach retains GEDI waveform information effectively and offers a scalable solution for regional and potentially global AGB modeling.