An improved machine learning approach to estimate GPP using vegetation optical depth and other microwave remote sensing observations

Long-term monitoring of gross primary production (GPP) is essential for quantifying terrestrial carbon uptake, understanding ecosystem responses to climate variability and extremes, and evaluating Earth system models. Yet, long-term global GPP estimates derived from optical remote sensing, eddy covariance upscaling, and process-based models still diverge in magnitude and trends, motivating the development of complementary products based on independent observations. Passive microwave vegetation optical depth (VOD) provides an all-weather, largely illumination-independent signal linked to vegetation water content and biomass and is used in the microwave-driven VODCA2GPP product. However, the current VODCA2GPP implementation uses reanalysis 2 m air temperature (T2M) and shows reduced performance in water-limited regions. 

Here, we assess a microwave-driven, model-independent GPP framework using random forest models trained on FLUXNET GPP and subsets of primarily microwave predictors. We replace T2M with daytime land-surface temperature (LSTday) retrieved from Ka-band brightness temperatures (AMSR-E, AMSR2, SSM/I). To better represent hydraulic and structural constraints, we additionally test land cover (LC), an L-band VOD biomass composite (LVOD), and surface and root-zone soil moisture (RZSM), alongside VOD (VODCA v2). 

Replacing T2M with LSTday preserves or slightly improves skill at FLUXNET sites and against independent GPP references, while producing near-identical global trend patterns, supporting LSTday as an observation-based thermal constraint consistent with large-scale controls on photosynthesis. Adding physiologically plausible predictors yields robust gains, with the most significant improvement from LC, which reduces cross-biome mixing and curbs unrealistically high GPP in open vegetation. The best performance is achieved when using VOD, LSTday, LC, LVOD, and RZSM together as predictors, highlighting the complementary constraints from plant-available water and biomass/long-term vegetation state. These results motivate an updated VODCA2GPP release, using LSTday instead of T2M and incorporating LC, LVOD, and RZSM, to better capture the structural and hydrologic limitations on carbon uptake.