Modeling temporal patterns of methane effluxes using multiple regression and random forest in Poyang Lake, China

Abstract

In this study, we used two statistical models to predict daily CH4 effluxes and compared the prediction accuracy of two models in Poyang Lake. Statistical models included linear model and Random forest model (RF) which can handle high dimensional non-linear relationships, categorical and continuous predictors, and highly collinear predictor variables. Seven climatic factors and water level data, together with the field CH4 efflux at monthly intervals from 2011 to 2014 were used for model development and cross-validation. We found that the RF model provided the best prediction accuracy for daily CH4 effluxes, whereas the linear model gave low prediction accuracy for CH4 effluxes. The coefficient of determination was 0.93 and 0.63 for the “best” RF and linear models with the same climatic variables, respectively. The “best” linear model had the highest model-performance errors including the mean absolute error, root mean-square error, and the normalized root-mean-square error, followed by the “best” RF models. In addition, cross-validation results for the two “best” models also showed that the RF model was the best model for estimating CH4 effluxes. We applied the optimum RF model to simulate daily CH4 effluxes from 1 January 2011 to 31 December 2014, and then estimated the seasonal and annual CH4 emissions in Poyang Lake. The mean CH4 efflux in the summer was notably higher than that in the other seasons, with values of 0.097, 0.28, 0.11, and 0.045 mmol m−2 day−1 in the spring, in the summer, in the autumn, and in the winter over a 4-year period, respectively. The mean annual emission was 3.13 g m−2 year−1, which was considerately lower than the mean global annual emission in lakes and that in the other subtropical lakes of the world. We found that the RF model may be used to estimate CH4 effluxes and emissions in other lakes in the world.