Machine learning for prediction of soil carbon stock changes in sugarcane crop due to straw removal

• Ano de defesa: 2021
• Autor(a) principal: Araujo, Ralf Vieira de
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Atributos do solo; Carbono orgânico do solo; Machine learning; Manejo de resíduos; Soil attributes; Soil organic carbon; Waste management
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/11/11140/tde-11102021-103725/

Brazil, as other countries, has established energy and climate policies that foster the use of biofuels as sugarcane ethanol, in which a growing practice is to use harvesting residues, the straw, for cogeneration of electricity or to produce second-generation ethanol. In this study, it was aimed to create machine learning (ML) models capable of predict short-term changes in the soil organic carbon stocks according to the mass of sugarcane straw leftover the soil during harvest. Considerations were also made on the current state of the art regarding the application of ML in soil science, with an emphasis on tropical soils and soil carbon stocks. The initial data was generated between 2015 and 2018 in five experimental sites under commercial cultivation of sugarcane in Brazilian south-central region and the available variables were related to climate, soil physical and chemical attributes, organic matter and crop variety. The variable to be predicted (y) was the rate of carbon stock change per area per year (Mg C ha-1 yr-1) in relation to the total dry mass of straw. The initial dataset was divided into training (80%) and test (20%) and eight ML models were trained using the algorithms Random Forest (RF) and Support Vector Machine (SVM) associated to four feature selection methods. Results were evaluated using 10-fold cross-validation of the root mean squared error (RMSE) in the training set and prediction RMSE in the test set. The trained models were statistically compared among them and to the use of mean y stratified by straw mass deposited and soil layer. All the ML models surpassed the simple generalization of previously known mean values of y. The model SVM associated with RF feature selection performed better with a considerable reduction in the number of attributes, which could reduce the costs and effort of data acquisition and processing in future applications. The achievements indicate that ML models are good tools to predict short-term changes in carbon stocks due to total or partial straw remotion from the field. The obtained results and applied methodology have the potential to help producers and decision-makers interested in identifying cause-effect relationships between in situ crop conditions, straw management and expected soil carbon variations.