Estimating ecosystem respiration using eddy covariance and neural networks : A comparison with traditional models for an intermediate-aged Douglas-fir stand

Because forest ecosystems assimilate roughly one-third of anthropogenic carbon dioxide (CO₂) emissions (N. L. Harris et al., 2021), understanding terrestrial carbon (C) cycles is essential for forest management. Eddy covariance measures net CO₂ fluxes between terrestrial ecosystems and the atmosphere, providing an opportunity to monitor these critical exchanges. However, partitioning net fluxes into ecosystem respiration (Rₑ) and gross primary productivity (GPP) remains challenging. Existing models often rely on nighttime or low-light data to build temperature-dependent Rₑ relationships and extrapolate to daytime conditions. Still, a knowledge gap remains around including additional meteorological and soil factors. This study used multilayer neural network (NN) models to test the hypothesis that incorporating additional factors would provide more accurate Rₑ estimates. Data were collected from a temperate rainforest stand on the east coast of Vancouver Island, BC, Canada, from Sep. 2002 until Dec. 2024. Two NN models, based on nighttime and low-light fluxes, predicted Rₑ from 23 and 22 soil and meteorological variables, respectively. Outputs were compared to analogous temperature-dependent models on half-hourly, daily, and monthly time scales. NN models did not significantly lower root mean square error relative to traditional models due to systematic overestimation. NN models also overestimated analogous traditional model annual Rₑ totals by a percent difference of 11-31%. These biases likely reflect differences between training and testing data due to interannual Rₑ trends. The results imply that further research is needed to improve model performance and provide novel approaches to understand forest C dynamics, such as through long short-term memory networks or physics-informed NNs.