Machine learning in the prediction of sugarcane production environments

• Published in: Computers and electronics in agriculture 2021-11, Vol.190, p.106452, Article 106452
• Main Authors: Almeida, Gabriela Mourão de, Pereira, Gener Tadeu, Bahia, Angélica Santos Rabelo de Souza, Fernandes, Kathleen, Marques Júnior, José
• Format: Article
• Language: English
• Subjects: Decision tree; Decision trees; Machine learning; Precision agriculture; Raw materials; Site-specific management; Soils; Spatial variability; Sugarcane
• ISSN: 0168-1699; 1872-7107
• DOI: 10.1016/j.compag.2021.106452

•Sugarcane has become the most prominent crop in Brazil.•The use of production environments is an alternative for specific soil management.•Machine learning can be used to define production environments.•The decision tree and geostatistics allow the creation of production environment maps. Sugarcane is one of the most important crops in the Brazilian agricultural market. Techniques that aim to increase the productivity and quality of raw materials, such as localized management, have been applied manually for many years by farmers and have great potential. This study aimed to determine sugarcane production environments using a reduced number of low-cost variables through the machine learning technique. The experiment was conducted in Guatapará, São Paulo State, Brazil. Initially, the database consisted of thirty variables, and six agronomic criteria were selected, three related to soil management and three to pedogenetic processes. The descriptive statistics was performed to understand the behavior of the data, followed by the stepwise regression to determine which variables would be useful to the model. Subsequently, a multicollinearity test and a decision tree were applied. A confusion matrix was prepared to assess the efficiency of the model. The variables related to soil formation factors, in particular sand, were chosen to determine the production environments. The stepwise regression was efficient in selecting the variables, while the decision tree was effective in determining the environments, with a satisfactory accuracy of 75% and the generation of more continuous management environments in the cultivation area.