Modeling net ecosystem exchange of CO2 with gated recurrent unit neural networks

The complex forcing mechanisms of biophysical drivers on ecosystem productivity, as well as some unknown relationships, stipulates the application of artificial intelligence techniques in modeling the ecosystem carbon cycle. In particular, various deep learning algorithms have been successfully applied to comprehend temporal relationships, though some issues regarding capturing temporal features for modeling CO2 fluxes across time scales remain unsolved. In this study, a gated recurrent unit (GRU) model was trained for simulating half-hourly net ecosystem exchange of CO2 (NEE) using 12-year continuous flux data (2009–2020) from seven experimental bioenergy crops in southwest Michigan, USA. The fields were historically managed either as corn-soybean rotation agricultural (AGR) or Conservation Reserve Program (CRP) lands and planted to no-till corn (AGR-C and CRP-C), restored-prairie (AGR-Pr and CPR-Pr), switchgrass (AGR-Sw and CRP-Sw) and smooth brome grass (CRP-Ref). We compared NEE simulations for the entire year, growing season, and non-growing season datasets, and analyzed relative importance of biophysical variables on variations of NEE. GRU performed well in simulating NEE at single site, with R2 of 0.89–0.93, 0.85–0.90, and 0.61–0.85 for the annual, growing season, and non-growing season datasets, respectively, which provided comparable accuracy and 6 % of run-time shorter than those of long short-term memory. R2 reached 0.70 and 0.80 across all sites and unique land-cover types. Radiation parameters contributed most to the variability of NEE. Contributions of individual variables appear more complex during the non-growing season and include incoming shortwave radiation, day of year, temperature, Monin-Obukhov stability, wind direction, and soil water content. The six most important forcing variables are consistent with our current understanding, although their ranked importance varied by ecosystem type and modeling scale. These results bring renewed insights in modeling observations at seasonal scales and provide guidance for variable selection to scale up measurements from one site to multiple land-cover types across the landscape.