Soil CO2 emission and its related with atmospheric CO2 in agricultural areas of Mato Grosso do Sul: a machine learning approach

• Grant number: 21/02487-0
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date: October 01, 2021
• End date: September 30, 2022
• Field of knowledge: Agronomical Sciences - Agronomy - Soil Science
• Principal Investigator: Alan Rodrigo Panosso
• Grantee: Letícia Roberta de Lima
• Host Institution: Faculdade de Ciências Agrárias e Veterinárias (FCAV). Universidade Estadual Paulista (UNESP). Campus de Jaboticabal. Jaboticabal , SP, Brazil

Abstract

Projections indicate a continuous growth of Brazilian agribusiness and a consequent increase in the bases of Greenhouse Gases (GHG) arising from this sector. Carbon dioxide (CO2) represents about 66% of the planet's total GHG, with soil organic carbon being one of the main terrestrial reservoirs for the storage and exchange of atmospheric carbon (C), since it depends on use and agricultural soil management, can act as sources or sinks of this carbon. Modeling the carbon dynamics in agricultural areas is a strategic action to reduce the uncertainties associated with GHG mitigation processes and improve the analysis capabilities to build more accurate scenarios. In recent decades, artificial intelligence and data mining techniques have been successfully applied in soil science attribute modeling. Thus, the objective of the proposal will be to evaluate the predictive performance of machine learning algorithms for the study of the relation between carbon dioxide emission in the soil (FCO2) and atmospheric CO2 in agricultural areas in the region of the state of Mato Grosso do Sul (MS), from a time series from 2015 to 2017. The techniques that will be used: support vector machine (SVM) and decision trees (Random Forest and Gradient Boosting Machines). Overall 70-80% of the choices used for learning (training process) of the models and 30-20% for validation. New experiments are being carried out in the field to validate the results in agricultural areas. The accuracy of the models will be provided by Pearson's correlation (r), coefficient of determination (R²), mean square error (RMSE), mean error (ME), agreement index (d), confidence coefficient (c), and lowest mean absolute percentage error (MAPE). It is expected that this approach will contribute to improving the understanding of the dynamics of atmospheric FCO2 and CO2 in different regions, land uses and management in central Brazil, producing scenarios with less uncertainty that can serve as a support in decision-making focused on mitigation of CO2 in agricultural areas. (AU)