Modelo de predição da ocorrência de ferrugem asiática na cultura da soja a partir de variáveis climáticas e clusterização

• Ano de defesa: 2022
• Autor(a) principal: Côrtes, Filipe da Silva
• Orientador(a): Abdalla, Klaus de Oliveira
• Banca de defesa: Abdalla, Klaus de Oliveira, Wander, Alcido Elenor, Heinemann, Alexandre Bryan
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Goiás
• Programa de Pós-Graduação: Programa de Pós-graduação em Agronegócio (EA)
• Departamento: Escola de Agronomia - EA (RG)
• País: Brasil
• Palavras-chave em Português: Aprendizado de máquina; Modelos preditivos; Agricultura de precisão; Agrometeorologia; Ferrugem asiática da soja
• Palavras-chave em Inglês: Machine learning; Predictive models; Precision agriculture; Agrometeorology; Asian soybean rust
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA
• Link de acesso: http://repositorio.bc.ufg.br/tede/handle/tede/11988

Asian rust is a disease with a significant impact on soybean in Brazil. Despite the great economic relevance of soybeans for Brazilian agribusiness, there are few studies on the conditions that cause the disease. This work aimed to create a predictive model considering the influence of climatic variables (temperature, precipitation, humidity and solar radiation), based on a dataset of rust occurrence, using the decision tree induction technique and logistic regression . The model was created with data on the occurrence of the disease in the cities of Cristalina, Jataí and Rio Verde - GO in the harvests from 2004/05 to 2016/17. For each occurrence record (detection), a corresponding “non-occurrence” was generated, this being the thirtieth day prior to the day of detection, assuming that on this date there would be no inoculum present in the field. The training set for the modeling has 10 variables totaling 393 records. The predictive model was proposed from the comparison of the best performance between the decision tree and logistic regression algorithms. After the accuracy results obtained (decision tree 77.88%, against 56.53% of the logistic regression algorithm), we used the clustering algorithm to group the data in the data preparation phase, again comparing the result between decision tree and logistic regression. With the support of clustering, we obtained the average accuracy in the range between 99 and 100% for decision tree and 66.75 and 100% for logistic regression.