Global GPP from microwave remote sensing–driven dynamics: feasibility, strengths, and limitations

Abstract. Long-term monitoring of gross primary production (GPP) is essential for quantifying terrestrial carbon uptake and ecosystem responses to climate variability and extremes. VODCA2GPP is a data-driven upscaling framework that links eddy-covariance (EC) GPP observations to passive-microwave vegetation optical depth (VOD), together with reanalysis air temperature, providing a multi-decadal microwave-based complement to optical GPP products. However, its reliance on reanalysis temperature reduces the methodological independence of the microwave-based estimate, and its sensitivity to water stress is limited in some ecosystems. Here, we develop and evaluate an updated microwave-driven GPP framework in which the time-varying environmental drivers are derived from microwave observations. At the same time, static land-cover information provides biome context. The framework produces 8-day GPP at 0.25° spatial resolution over 2002–2020. Using a sequence of Random Forest (RF) models trained to upscale EC GPP, we test the incremental contribution of different predictor groups. We replace ERA5-Land 2-m air temperature (T2M) with land-surface temperature (LST) derived from harmonised Ka-band brightness temperatures from AMSR-E, AMSR2, and SSMIS F17, and then evaluate land-cover (LC) fractions as biome context, a spatially varying multi-year mean of L-band VOD as persistent canopy-structure information, and surface soil moisture (SSM) and root-zone soil moisture (RZSM) as shallow and deeper water-availability constraints. The models are compared with external GPP products, sun-induced chlorophyll fluorescence (SIF), and drought-related datasets.

Predictive skill is assessed using site-grouped 10-fold cross-validation, keeping all observations from a site together in either the training or validation set within each fold. Replacing reanalysis T2M with microwave-derived LST yielded broadly comparable site-level predictive performance. Adding LC fractions produces the largest individual improvement in predictive skill (median r: 0.690 to 0.762; RMSE: 2.646 to 2.345 g C m⁻² d⁻¹), while the spatially varying multi-year L-band VOD mean provides a smaller additional improvement by adding persistent canopy-structure information. Structural and hydrological constraints improve agreement with external GPP products, SIF, and drought indicators, with the clearest gains in short-vegetation and water-limited regions; among the hydrological predictors, RZSM provides the strongest improvement. The best-performing configuration achieves a median site-level correlation of r = 0.798 and an RMSE of 2.294 g C m⁻² d⁻¹. Despite these improvements, the GPP magnitude of our models remains systematically higher than in the comparison products.

Overall, the results show that the updated framework can reproduce long-term GPP dynamics without reanalysis-based temperature forcing, while static LC fractions provide complementary biome context. Among the added predictors, LC fractions provide the largest overall gain in predictive skill and help distinguish productivity among vegetation types, the multi-year L-band VOD mean adds persistent canopy-structure information, and RZSM improves the representation of water-limited GPP variability, especially at short-vegetation sites. Remaining limitations include the mismatch between EC footprints and the much coarser satellite grid, uneven environmental and geographic coverage of the tower network, and environmental controls on GPP that are not explicitly represented by the current predictor set.