Proposta de um sistema computacional para monitoramento do ambiente de produção agrícola

• Ano de defesa: 2020
• Autor(a) principal: Santos, Marcio Aurélio Soares
• Orientador(a): Omar, Nizam
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Presbiteriana Mackenzie
• Palavras-chave em Português: aprendizado de maquina; sensoriamento remoto; agricultura; uso do solo; acurácia
• Área do conhecimento CNPq: CNPQ::CIENCIAS AGRARIAS::ENGENHARIA AGRICOLA
• Link de acesso: https://dspace.mackenzie.br/handle/10899/28592

The identification and geospatial monitoring of areas dedicated to agribusiness is fundamental information in the elaboration of strategies and management of economic and environmental resources, both of interest to the wider society. In this work, the objective is to highlight agricultural areas dedicated to seasonal crops, that is, apply a robust computational method, Machine Learning, and enable the generation of information about the agricultural areas with the required accuracy and timely. The challenge requires an approach capable of interacting with different data sources to timely generate accurate field information. Therefore, it is about dealing with the complexity of the environment through the lens of sensors and proper modeling. There are several contributions that aim to meet this demand and in particular, those that address the extraction of information in large volumes of data, as is the case in this research which makes use of temporary series extracted from images. Likewise, other related works share their achievements and improvements using machine learning to classify agriculture areas and the specifics from each of the studied environments. The preliminary results are promising, a layer of knowledge that allows the application of the current techniques and methods to improve information at the culture level has been generated, a two levels classification process. A comparison of the results with the combination of similarity metrics to the dataset as additional attributes was made using the following algorithms: Naive Bayes, Generalized Linear Model, Logistic, Deep Learning, Decision Tree, Randon Forest, Gradient Boosted tree e Support Vector Machine. The overall accuracy achieved was between 93.8% e 99.6%, the highest performance using Boosted Decision Tree algorithms. This information is useful for future research and also to support the private and public sectors in the monitoring and spatial planning of food crops in Brazil.