Evaluating the performance of airborne and spaceborne lidar for mapping biomass in the United States' largest dry woodland ecosystem

The ability of remote sensing to accurately quantify live aboveground biomass (AGB) varies by ecosystem. Given their important role in global carbon dynamics, deriving accurate, spatially and temporally explicit AGB estimates in dryland ecosystems is uniquely valuable. However, the shorter stature and sparser cover of vegetation in dryland ecosystems makes remote sensing of AGB particularly challenging. The United States' largest dry woodland ecosystem is that of piñon-juniper (PJ) woodlands, a diverse and widespread vegetation type whose AGB has not been mapped in a comprehensive and targeted manner using lidar. In this study, we investigated airborne and spaceborne lidar for their respective AGB estimation abilities in PJ woodlands. Using data from 177 field plots distributed over 18 sites capturing the spatial and ecological variability within the full range of PJ in the US, we compared three different modeling approaches: (1) using field-measured AGB to train and validate models built from airborne laser scanning (ALS) data (Field→ALS); (2) using field-measured AGB to train and validate models built from simulated Global Ecosystem Dynamics Investigation (GEDI) waveforms (Field→GEDIsim); and (3) using ALS-modeled AGB to train and validate models built from real GEDI waveforms (ALS→GEDIreal). In doing so, we also compared three different ensemble decision tree-based machine learning algorithms: (1) cubist; (2) random forests; and (3) extreme gradient boosting (XGBoost). The Field→ALS models performed very well, with a mean R2 of 0.69 and nRMSE of 36.91% across the three machine learning algorithms. The Field→GEDIsim models saw decreased performance (R2mean = 0.50; nRMSEmean = 47.47%), likely due to the simulated waveforms' inability to sufficiently capture vegetation structure in the short, sparse woodlands. The ALS→GEDIreal had the lowest mean R2 (0.36), but relatively constrained predictions yielded similar mean nRMSE to Field→GEDIsim (46.19%), though that is without accounting for the propagation of error resulting from being trained and validated on modeled predictions rather than measured values. Cubist's ability to extrapolate proved helpful in the presence of stronger predictors (i.e., Field→ALS), enhancing prediction of extreme AGB values not represented in the reference data. Conversely, when predictive capacity was comparably low (i.e., Field→GEDIsim and ALS→GEDIreal), random forests and XGBoost's inability to extrapolate yielded lower predictive error. We compared our results to the GEDI Level 4A (L4A) footprint-level AGB product, which revealed that L4A tends to significantly underestimate AGB in PJ woodlands and fails to capture variability on the low end of the AGB spectrum (0–100 Mg/ha). These results demonstrate promise for broad-scale, lidar-driven PJ and other dry woodland ecosystem AGB mapping, and suggest that with more ecosystem-tailored models, near-global products such as L4A could be improved.