Prediction of Greenhouse Gas Fluxes in Coastal Salt Marshes with Machine Learning

Abstract:

We explored the suitability of machine learning models to predict the major greenhouse gas (GHG) fluxes (CO2 and CH4) in salt marshes. Eight artificial neural network (ANN) models were employed in this research: linear layer neural network (LLNN), feed forward neural network (FFNN), multi-layer perceptron neural network (MLNN), cascade forward neural network (CFNN), layer-recurrent neural network (LRNN), nonlinear autoregressive neural network with exogenous inputs (NARX), radial basis function neural network (RBNN), and generalized regression neural network (GRNN). The models were used to predict the GHG fluxes in salt marshes of Waquoit Bay, MA, USA as a case study. First, we developed ANN models to simulate the GHG fluxes with individual environmental variables. The ANN models indicated soil temperature and salinity as the major predictors of both GHG fluxes. Additionally, photosynthetically active radiation appeared as an important predictor for the daytime net uptake fluxes of CO2. The dominated predictors were then used to develop multivariate ANN models. RBNN predicted both GHG fluxes with the most success (R2 = 0.90-0.99). LLNN was the least successful model in prediction (R2 = 0.61-0.81), except for the predictions of net emission fluxes of CO2 (R2 = 0.90-0.92). Other ANN models provided similar performance (R2 = 0.88-0.95), although LRNN and GRNN exhibited better accuracy. The results suggest a high potential of machine learning for accurately predicting GHG fluxes and carbon storage in coastal wetlands.