Comparison of the Performance of Statistical Models that Predict Soil Respiration from Forests

Soil respiration (Rs) is an important source of CO2 to the atmosphere, yet understanding the processes controlling the combined autotrophic and heterotrophic components has proven challenging. Numerous statistical models have been developed to explain Rs as a function of physical and chemical conditions, but there has been little effort to systematically evaluate and compare the performance of these different models. Soils in a sugar maple (Acer saccharum Marsh.) forest were monitored for Rs and its physical and chemical drivers, and the monitored data were used to fit the most common of these statistical models. Each model was fit using a repeated trials approach and the performance of the model was evaluated with a range of goodness-of-fit measures (RMSE, mean absolute error, r2, index of agreement, bias, and Akaike information criterion). An exponential model in which the exponent is a polynomial expression that is linear with respect to temperature (°C) and quadratic with respect to soil moisture (% by volume) explained 57% of the variance in Rs and performed well among the goodness-of-fit measures. Inclusion of C quantity and substrate quality (as measured by the C/N ratio) in the soils, however, increased the explanation of variance in Rs to 71%. This study shows that combining soil temperature and moisture drivers with soil C quantity and substrate quality significantly improves statistical models predicting soil respiration. Our findings support the comparison of the performance of different statistical models using a repeated trials approach coupled with a range of goodness-of-fit measures to identify controls on Rs within an ecosystem and to assess the generality of these controls on Rs across ecosystems.