A Preliminary Random Forest Baseline for Above-Ground Carbon Stock Estimation in Abuja Municipal Area Council Using Free Multi-Sensor Satellite Data: A GeoAI Approach

Accurate quantification of above-ground carbon (AGC) stock is essential for climate mitigation planning, yet fine-scale carbon mapping remains scarce for rapidly urbanizing West African savanna cities. This study presents a preliminary geospatial artificial intelligence (GeoAI) baseline for AGC estimation in the Abuja Municipal Area Council (AMAC), Federal Capital Territory, Nigeria, using entirely free, cloud-hosted satellite data. A Random Forest regression model was trained on Google Earth Engine using a dry-season Sentinel-2 optical composite, a Sentinel-1 SAR composite, Copernicus DEM elevation, and NASA GEDI L4A above-ground biomass density (AGBD) footprints as the reference label. A total of 10,978 quality-filtered GEDI footprints were retrieved across AMAC's approximately 172,993-hectare extent, of which 3,309 were held out for testing. On the held-out set, the model achieved a root-mean-square error (RMSE) of 47.02 Mg/ha, a mean absolute error (MAE) of 16.52 Mg/ha, and a coefficient of determination (R²) of 0.345. Wall-to-wall prediction yielded a mean AGC density of 10.12 Mg C/ha and an estimated total AGC stock of approximately 1.75 million Mg C for AMAC. A small fraction (<1%) of anomalously high GEDI AGBD values was found to disproportionately influence RMSE relative to MAE, consistent with known GEDI performance limitations in structurally heterogeneous, non-forest-dominated landscapes. These results are presented explicitly as a reproducible starting baseline rather than a validated final product: the model uses GEDI as both the training label and the accuracy reference, and the train/test split was not spatially blocked. Field-inventory calibration, spatially blocked validation, and fine-tuning of multimodal geospatial foundation models are identified as concrete future research directions building directly on this baseline.