Modelling carbon dioxide emissions under a maize-soy rotation using machine learning

Climatic parameters influence CO2 emissions and the complexity of the relationship is not fully captured in biophysical models. Machine learning (ML) is now being applied to environmental problems, and it is, therefore, opportune to investigate ML models in CO2 predictions from agricultural soils. In this study, six ML models were compared for their predictive performance by comparing field measurements of CO2 emissions from two fertiliser treatments: inorganic fertiliser (IF) and solid cattle manure supplemented with inorganic fertiliser (SCM) applied to a maize-soy rotation. The study also included a generalised scenario where all the data from IF and SCM were included in one dataset. The ML models include support vector machine (SVM), random forest (RF), least absolute shrinkage and selection operator (LASSO), feed-forward neural network (FNN), radial basis function neural network (RBFNN), and extreme learning machine (ELM). The input parameters were soil moisture, soil temperature, soil organic matter, soil total carbon, soil total nitrogen, air temperature, solar radiation and pan evaporation, while the output parameter was field measured CO2 emissions. The results of this study demonstrated that RF was the best at predicting CO2 emissions from IF [coefficient of determination (R-2) = 0.92 and root mean square error (RMSE) = 2.27], SCM (R-2 = 0.94 and RMSE = 2.86) and generalised scenarios (R-2 = 0.86 and RMSE = 3.05). We conclude that ML models provide an innovative, robust and time-efficient alternative to biophysical models. (C) 2021 IAgrE. Published by Elsevier Ltd. All rights reserved.

Automated summary

This study compares six machine learning models in predicting CO2 emissions from agricultural soils, and finds that random forest (RF) is the best model for predicting CO2 emissions under different fertilizer treatments.