Assessing spatial dynamics of GEDI biomass prediction in managed versus unmanaged tropical forest ecosystems in Kenya

Above ground biomass (AGB) estimation is vital for monitoring carbon storage and ecosystem fluxes, especially in tropical forests for climate mitigation in the Global South. The Global Ecosystem Dynamics Investigation (GEDI) instrument offers high resolution forest monitoring; however, field measurements are crucial to enhance spatial accuracy. This study assessed the limitations of using machine learning models trained on GEDI data to estimate AGB for two distinct forest ecosystems in Kenya: Kakamega National Forest Reserve (KNFR) and Karura Forest Reserve (KFR). Our specific objectives were (i) developing Random Forest (RF) machine learning model using GEDI data for training, (ii) comparing GEDI estimates with field measurements, and (iii) quantifying the limitations of using integrated GEDI-RF model. Plots were coincided with GEDI beams for comparison to assess variability and bias in AGB estimates. The p-value of 0.005 for heteroskedasticity in KNFR indicated high variability and bias of the GEDI-RF model relative to the field measured model. In contrast, p-value of 0.195 for heteroskedasticity in KFR indicated low variability and bias of the GEDI-RF model relative to the field measured model. 55% of plots which coincided with GEDI had less than 10% relative difference in AGB estimates between models. Plots outside of GEDI had a relative difference in AGB estimates between models greater than 10%. Relative to the field measured model, the GEDI-RF model overestimated AGB values less than 100 Mg ha⁻¹ and underestimated greater than 200 Mg ha⁻¹. This study contributes to effective forest monitoring, carbon accounting, and conservation in heterogeneous forests.