Early vs. Late Fusion of Sentinel-1 and Sentinel-2 for GEDI-Supervised Above-Ground Biomass Regression

Combining Sentinel-1 synthetic aperture radar (SAR) with Sentinel-2 optical imagery is a common strategy for mapping above-ground biomass (AGB), yet the question of how the two sensors should be fused is seldom evaluated in a controlled setting. This letter compares two fusion strategies—early fusion by channel concatenation and late fusion with independent per-sensor encoders—against single-sensor baselines for GEDI L4A-supervised AGB regression over a study region in Galicia, northwest Iberia. All four configurations share an identical convolutional backbone, training procedure, and spatially blocked cross-validation, so that only the fusion strategy varies. On a held-out test set, late fusion achieves the lowest overall error (RMSE 50.33 Mg/ha, R2 0.44) and significantly outperforms early fusion, whereas early concatenation does not improve on the optical-only baseline—indicating that the fusion architecture, not the mere inclusion of SAR, drives performance. Stratified analysis localizes the advantage: it is concentrated in high-biomass tree cover, where optical reflectance saturates and SAR retains sensitivity, and is negligible at low biomass. An ablation confirms the effect is not an artifact of terrain-feature duplication. A wall-to-wall map from the best model reproduces the spatial structure of an independent biomass product while exhibiting the dynamic-range compression characteristic of GEDI-supervised regression; we show this compression is not remedied by target log-transformation, pointing instead to tail-sensitive training as the key direction.