A synergistic approach to sugarcane yield forecasting using machine learning, remote sensing, and process-based modeling

• Ano de defesa: 2023
• Autor(a) principal: Grubert, Daniel Alves da Veiga
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Agricultural yield prediction; Algoritmos de aprendizado de máquina; APSIM-Sugar; Hybrid modeling; Machine learning algorithms; Modelagem híbrida; Modelagem preditiva; Predição de produtividade agrícola; Predictive modeling; Simulação de cana-de-açúcar; Sugarcane simulation
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/11/11152/tde-03112023-103445/

Accurate and precise crop yield forecasts are essential for farmers and decision-makers. This study aims to assess a hybrid approach involving remote sensing data, crop modeling with process-based models, and machine learning algorithms to improve sugarcane yield predictions. To achieve this, a hybrid yield forecasting approach was developed, combining various data sources, including simulated soil and plant variables from the APSIM model (a process-based crop model), meteorological data, and vegetation indices. These data were used as inputs in machine learning models to forecast end-season sugarcane yield. In this study, 16 regression models were evaluated to forecast sugarcane yield at the municipal level in the state of São Paulo, Brazil, during the period 2010-2020. The results indicated that the hybrid approach developed using the K-Neighbors Regressor algorithm showed the best statistical performance, resulting in the lowest Mean Absolute Error (MAE) of 3.26 t ha-1, with a Mean Absolute Percentage Error (MAPE) of 4.54%. Sugarcane yield predictions were most accurate 1-2 months before harvesting. Furthermore, the study determined which variables had the greatest influence on sugarcane productivity prediction by partially excluding some variables from the prediction model. The results showed that adding variables simulated by the process-based model (APSIM) as input variables for machine learning models could reduce the Root Mean Square Error (RMSE) of yield prediction, ranging from 7.7% to 26.9%, while vegetation indices had the least impact on predictions. The analysis revealed that meteorological data had a greater impact on yield prediction when provided to the process-based model than when directly used in machine learning algorithms. This result suggests that the simulated variables provided by APSIM offer a more comprehensive biophysical description of the interaction between soil, plant, and atmosphere.