In-Situ Validation Reveals Poor Performance of Extrapolated GEDI Aboveground Biomass Estimates Across Miombo Landscapes

The Global Ecosystem Dynamics Investigation (GEDI) collected a near-global sample of forest structural metrics between 2019 to 2023, which were used to estimate Aboveground Biomass Density (AGBD). GEDI AGBD estimates were derived from predictive models broadly stratified by plant functional type and continental region. However, the miombo region, encompassing 2.7 million km2 of Southern Africa and characterised by heterogenous and complex vegetation, was critically under-represented in the calibration of GEDI AGBD estimates. Furthermore, sparse GEDI sampling prevents direct in-situ validation of GEDI AGBD estimates with field AGBD estimates. In lieu of airborne LiDAR data to simulate GEDI metrics, this study employed Random Forest machine learning to extrapolate GEDI AGBD estimates across two miombo landscapes and produce a time series from 2017 to 2023 to coincide with 48 field AGBD estimates from permanent sample plots. This study found moderate predictive relationships between GEDI AGBD estimates and EO predictor variables including phenology variables from Landsat and L-band backscatter from PALSAR-2, which yielded R2 = 0.29, RMSE = 23.12 Mg/ha, Bias = +2.65 Mg/ha. However, spatial cross-validation revealed drastic differences between site-specific models, highlighting the importance of exercising caution to avoid misrepresenting the heterogeneity and complexity of mixed woodland-savanna ecosystems in large-scale models. Ultimately, in-situ validation of extrapolated GEDI AGBD estimates across both sites and census years yielded quasi-null predictive power with R2 = 0.09, RMSE = 33.42 Mg/ha, and Bias = +19.75 Mg/ha. This study concludes that reliable GEDI AGBD estimates in the miombo region require locally calibrated models and highlights the critical importance of integrating in-situ field AGBD estimates across space and time for calibration and validation of large scale AGBD estimates.