Machine Learning Approach to Simulate Soil CO\u3csub\u3e2\u3c/sub\u3e Fluxes under Cropping Systems

With the growing number of datasets to describe greenhouse gas (GHG) emissions, there is an opportunity to develop novel predictive models that require neither the expense nor time required to make direct field measurements. This study evaluates the potential for machine learning (ML) approaches to predict soil GHG emissions without the biogeochemical expertise that is required to use many current models for simulating soil GHGs. There are ample data from field measurements now publicly available to test new modeling approaches. The objective of this paper was to develop and evaluate machine learning (ML) models using field data (soil temperature, soil moisture, soil classification, crop type, fertilization type, and air temperature) available in the Greenhouse gas Reduction through Agricultural Carbon Enhancement network (GRACEnet) database to simulate soil CO2 fluxes with different fertilization methods. Four machine learning algorithms—K nearest neighbor regression (KNN), support vector regression (SVR), random forest (RF) regression, and gradient boosted (GB) regression—were used to develop the models. The GB regression model outperformed all the other models on the training dataset with R2 = 0.88, MAE = 2177.89 g C ha−1 day−1, and RMSE 4405.43 g C ha−1 day−1. However, the RF and GB regression models both performed optimally on the unseen test dataset with R2 = 0.82. Machine learning tools were useful for developing predictors based on soil classification, soil temperature and air temperature when a large database like GRACEnet is available, but these were not highly predictive variables in correlation analysis. This study demonstrates the suitability of using tree-based ML algorithms for predictive modeling of CO2 fluxes, but no biogeochemical processes can be described with such models