In this paper, a stacked Bidirectional Long Short-Term Memory (BDLSTM) is proposed to predict the emission of Nitrous Oxide ($N_2O$) from agricultural soils. The model is trained using the data collected by LI-COR soil-gas measurement equipment in Area X.O. Ottawa, Canada. With MSE as the loss function and Adam as the optimizer, the model is evaluated against mean absolute error(MAE), mean absolute percentage error (MAPE), and root mean square error(RMSE). In comparison with the MLP model, it is observed that the stacked BDLSTM model has superior performance.
The model is trained on two thousand data points with an early stopping technique, which, in general, is used to avoid overfitting in a highly complex model. Although the overfitting is depressed with the early stopping technique, it is necessary to consider the trade-off between the computation efficiency and the prediction accuracy. In this study, two BDLSTM layers (each with 100 hidden nodes) would have achieved the optimal balance. Also, the experiment shows that if an unrelated feature is added to the input features, it will degrade the performance of the model, and the simulation also indicated that it is necessary to choose an appropriate number of time steps (12 time steps) to obtain the best prediction accuracy.
Article ID: 2022L6
Venue: Canadian Conference on Artificial Intelligence
Publisher: Canadian Artificial Intelligence Association