Predicting Methane Emission from Paddy Fields with Limited Soil Data by Artificial Neural Networks

Abstract

Paddy fields with flooded water are commonly known as the main source of methane (CH 4 ) emission. Measurement of methane gas is complicated and costly due to utilizing advanced instrumentation such as Gas Chromatography (GC). In addition, some soil parameter data that is effected on methane emissions were sometimes limited. The current study proposed the artificial neural network (ANN) models to predict methane emission from paddy fields using soil parameters of soil moisture (SM), soil temperature (Ts) and soil electrical conductivity (EC). The models were performed based on the experimental data in one planting season during 20 January to 13 May 2018. The ANN models were developed by backpropagation learning algorithm. The sigmoid function was used as an activation function with three layer, i.e. input, hidden and output layers. There were six ANN models with different input parameters. Model 1 with three input parameters (SM, Ts, EC) was the best model with highest R2 (0.97) and lowest RMSE (32.9 mg/m2/d). However, the model needs more input parameters. The Model 2 can be used as an alternative to predict methane emission if there were only two measured parameter, i.e., SM and Ts. This model was fairly satisfied as indicated by R2 of 0.64 and moderate RMSE of 122.8 mg/m2/d