Machine learning for predicting greenhouse gas emissions from agricultural soils

Area of Science:

• Environmental Science
• Agricultural Science
• Data Science

Background:

• Machine learning (ML) models are increasingly utilized for environmental studies due to their ability to handle complex, variable data.
• Predicting soil greenhouse gas (GHG) emissions is crucial for understanding agricultural impacts on climate change.

Purpose of the Study:

• To evaluate and compare the performance of different ML regression models for predicting soil CO2 and N2O fluxes.
• To identify the most effective ML approach for simulating GHG emissions from agricultural fields.

Main Methods:

• Collected five years of environmental, agronomic, and soil data alongside CO2 and N2O flux measurements from an agricultural field in Quebec, Canada.
• Applied and cross-validated classical regression (RF, SVM, LASSO), shallow learning, and deep learning (LSTM) models.
• Statistically compared model performance using R coefficient and RMSE values.

Main Results:

• The LSTM model demonstrated superior performance in predicting both CO2 (R=0.87) and N2O (R=0.86) fluxes, outperforming a biophysical model (RZWQM2).
• Classical regression models showed moderate success for CO2 flux but failed to accurately predict peak N2O fluxes.
• Shallow ML models were less effective and highly sensitive to hyperparameter tuning.

Conclusions:

• The LSTM model is a highly effective tool for accurately simulating soil GHG emissions, offering a significant advancement over traditional methods.
• This study highlights the potential of deep learning, specifically LSTM, for environmental modeling and predicting agricultural GHG emissions.