Accurate estimation of gross primary production (GPP) is fundamental for quantifying the terrestrial carbon cycle. However, coarse-resolution products often fail to capture fine-scale spatial variations in carbon uptake across heterogeneous landscapes. While recent studies have begun to employ 10m Sentinel-1 and Sentinel-2 imagery, they typically reduce these data to pixel-wise spectral indices, discarding the two-dimensional spatial structure (canopy architecture, land-cover transitions, within-stand heterogeneity) that the imagery encodes. This study investigates whether explicitly exploiting this spatial context via convolutional neural networks yields robust, transferable gains over tabular machine-learning baselines. We curate a quality-controlled dataset of 23,528 eight-day multi-sensor composites from 222 AmeriFlux sites (2015&ndash;2025), evaluated under site-wise cross-validation, temporal generalisation, and geographic transfer to an 18-site upper Midwest forest holdout. Under temporal transfer to unseen years (2023&ndash;2025), the best convolutional model achieves <em>R²</em> = 0.77 and RMSE = 1.95 gCm^&minus;2 d^&minus;1, an 18.6% RMSE reduction over ridge regression (<em>R²</em> = 0.65, RMSE = 2.40 gCm^&minus;2 d^&minus;1). Although this advantage narrows under geographic transfer to structurally novel regions (<em>R² </em>= 0.59 vs. 0.54), the convolutional models still outperform all tabular baselines. Spatial structure at 10m therefore supports more robust temporal generalisation than spectral aggregates alone.