MACHINE LEARNING MODELS FOR EARLY FORECAST OF SUGARCANE YIELD FOR THE MAIN PRODUCING REGIONS IN BRAZIL

• Grant number: 22/11597-7
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date: October 01, 2022
• End date: September 30, 2023
• Field of knowledge: Physical Sciences and Mathematics - Probability and Statistics - Applied Probability and Statistics
• Principal Investigator: Glauco de Souza Rolim
• Grantee: Marina Moreira Santos
• Host Institution: Faculdade de Ciências Agrárias e Veterinárias (FCAV). Universidade Estadual Paulista (UNESP). Campus de Jaboticabal. Jaboticabal , SP, Brazil

Abstract

Brazil is the world's largest producer of sugarcane, in addition to having record sugar production and exports, it is a major producer of ethanol and its bagasse is used to generate electricity. Climatic conditions have a great influence on the development and productivity of sugarcane cultivation. It has a long cycle, of a year or a year and a half, being very exposed to the conditions, therefore, studies that quantify the climatic influence on development and productivity are necessary.Estimation is the quantification of a current event, while prediction is the quantification of a future event (Aparecido, 2016). Thus, many mechanistic models have been developed to estimate sugarcane productivity, such as Pereira (1987), Inman-Bamber and Thompson (1989), Barbieri (1993), APSIM (McCown et al, 1995), QCANE (Liu and Kingston , 1995) and DSSAT-CANEGRO (Jones, 2003), Mosicas (Martiné and Todoroff, 2004), among others.Although there are many models for estimating yields in sugarcane, there are few works on yield prediction models.Forecasting yields using agrometeorological models are robust techniques for planning crop areas, prior knowledge of agricultural yields facilitates decision-making by managers (Aparecido, 2016).Several works indicate that machine learning models are accurate enough to predict agricultural quality and yields, for example for soybean (Yoosefzaden-Najafabadi, et al., 2021), sugarcane (Xu, et al. 2020), peanut (Radyab and Watson, 2022), coffee (Aparecido et al. 2022). In addition, machine learning models have the advantage of directly evaluating which environmental variables and at which times are important for predicting productivity without any assumptions.Thus, the objective of this project is to use the Multiple Linear Regression, Random Forest, Support Vector Machine and Multilayer Perceptron models to forecast sugarcane productivity, with a period as early as possible, based on climatic variables and vegetation indices for the main sugarcane producing regions in Brazil.