Scaled biomass estimation in woodland ecosystems: Testing the individual and combined capacities of satellite multispectral and lidar data

Airborne laser scanning (ALS) data enable accurate modeling and mapping of aboveground biomass (AGB), but the limited spatial and temporal extents of ALS data collection limit the capacity for broad-scale carbon accounting. Conversely, while space-based remote sensing instruments provide increased spatial and temporal coverage, it can be difficult to directly link field-level vegetation biometrics to satellite data due to coarser spatial resolution and positional uncertainty. The combined use of ALS and satellite remote sensing data may offer a solution to efficient, accurate, and consistent AGB mapping across time and space. Such airborne-spaceborne data fusion has been demonstrated successfully in high-biomass settings; however, the unique structural conditions of dryland woodland ecosystems, with open canopies and low leaf area indices, pose mapping challenges that require further study. These challenges are particularly acute with large footprint spaceborne lidar, where short, widely-spaced trees may limit the capacity for accurate AGB estimation. In this study, we present a scaled methodological framework for linking field-measured woodland AGB to ALS data and, in turn, linking ALS-modeled AGB to satellite data, using piñon-juniper woodlands in southeastern Utah as a case study. We compare the effectiveness of this scaling approach using two satellite sensors, Landsat 8 OLI (multispectral) and GEDI (lidar). Since the predicted outputs of our local-scale model are being used as inputs to our regional-scale model, we also demonstrate an approach for propagating uncertainty throughout this nested, multiscale analytical framework, leveraging the inherent variability within a random forest's decision trees. Given the positional uncertainty of GEDI footprints, we test a range of different footprint sizes for their relative effects on ALS-GEDI AGB model accuracy. Our local-scale (field-ALS) predictive model was able to account for 74% of variance in AGB, and estimate AGB with a root mean squared error (RMSE) of 14 Mg/ha, a mean absolute error (MAE) of 11.09 Mg/ha. Our regional-scale (ALS-Landsat/GEDI) analysis with propagated uncertainty revealed that the combined use of Landsat and GEDI metrics produced the best predictive model (R2 = 0.68; RMSE = 12.71 Mg/ha; MAE = 9.40 Mg/ha), followed by Landsat-only metrics (R2 = 0.66, RMSE = 13.08 Mg/ha; MAE = 9.71 Mg/ha), and GEDI-only metrics (R2 = 0.49, RMSE = 16.01 Mg/ha; MAE = 12.14 Mg/ha). These results suggest that Landsat may be better-suited than GEDI for estimating AGB in woodland environments where low canopy covers and short trees limit the capacity for precisely characterizing vegetation structure within large-footprint, waveform lidar data. The footprint size analysis revealed that larger simulated footprints (e.g., 30 m radius and greater) produced higher GEDI model accuracies; however, increasing footprint radii beyond 30 m does not significantly increase model accuracy. This research represents an important step forward in improving our capacity for reliably mapping woodland AGB, and provides an early test case for the application of GEDI data to woodland AGB mapping.