Simulating Carbon Flux and Evapotranspiration of Two Forage Production Systems in Middle Tennessee Using Machine Learning.

Forage production systems play important roles in supplying high-quality animal feed and altering carbon flux and hydrological cycle on an ecosystem scale. Accurately quantifying CO₂ flux and evapotranspiration of different agroecosystems requires advanced instrumentation methods such as the Eddy Covariance (EC) Systems. In this study, we have established tower-based EC systems in two quintessential forage production systems in Middle Tennessee, including one warm-season forage production field dominated by bermudagrass [Cynodon dactylon (L.) Pers.] and one organic transitioning cool-season field dominated by tall fescue (Festuca arundinacea L.), kentucky bluegrass (Poa pratensis L.), and orchardgrass (Dactylis glomerata L.). Field-scale fluxes of CO₂ and water vapor (latent heat or evapotranspiration) have been continuously measured since 2016. A suite of machine learning algorithms (linear vs. non-linear, parametric vs non-parametric) were adopted to help identify the optimal modeling approach that could be used for upscale the results on a much larger temporal and spatial scale. Particularly, 11 easy-to-measure environmental variables (features) were included in the model construction process, including net radiation, irradiance, photon flux density, soil heat flux, vapor pressure deficit, relative humidity, ambient air and soil temperature, wind speed, soil water content, and precipitation. Across both forage fields, our best models yielded the coefficient of determination values (R²) of 0.84 and 0.87 for predicting 30-min average CO₂ flux and evapotranspiration, respectively. The top three informative features include net radiation (P < 0.001), soil water content (P < 0.001), and wind speed (P < 0.001) for predicting CO₂ flux; and net radiation (P < 0.001), relative humidity (P < 0.001), and ambient air temperature (P < 0.001) for predicting evapotranspiration.