This study evaluates the synergy of NASA’s Global Ecosystem Dynamics Investigation (GEDI) spaceborne LiDAR, Sentinel-2 multispectral imagery, and L-band Argentine Satellite System for Emergency Management (SAOCOM) 1A SAR data for aboveground biomass (AGB) estimation in the Belgrade Forest, Istanbul. Utilizing 1,356 GEDI L4A footprints as reference data, the research incorporates ten Sentinel-2 bands, five optical indices (NDVI, NDVIred, EVI, LSWI, CIre), SAR backscattering coefficients (σ°HH and σ°HV), polarimetric H/A/α polarimetric decomposition parameters and dual polarimetric radar vegetation indices, namely the Dual-Pol Radar Vegetation Index (DpRVI). High-dimensional feature spaces were optimized through ensemble-based, correlation-based, and hybrid RFECV selection strategies before evaluating four machine learning architectures: Multi-layer Perceptron (MLP), Kernel Ridge, Lasso, and Elastic Net. The MLP model achieved the highest predictive accuracy (R² = 0.20, RMSE = 62.93 Mg/ha, MAE = 51.31 Mg/ha), outperforming linear regularization models, which exhibited R²values between 0.15 and 0.16. Sensitivity analysis identified red-edge and SWIR bands, alongside indices such as NDVIred, LSWI, and CIre, as the most robust predictors, while the contribution of SAR-derived features remained comparatively limited. These findings underscore the efficacy of non-linear deep learning architectures and multi-source data fusion in resolving complex biophysical interactions within heterogeneous forest environments.