Aprendizado profundo aplicado na pulverização seletiva em tempo real para controle de Ipomoea spp.

• Ano de defesa: 2020
• Autor(a) principal: Sabóia, Hederson de Souza
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Mato Grosso; Brasil; Instituto de Ciências Agrárias e Tecnológicas (ICAT) – Rondonópolis; UFMT CUR - Rondonopólis; Programa de Pós-Graduação em Engenharia Agrícola
• Palavras-chave em Português: Redes neurais convolucionais; Agricultura de precisão; Plantas daninhas; CNPQ::CIENCIAS AGRARIAS::ENGENHARIA AGRICOLA; Convolutional neural networks; Precision agriculture; Weeds
• Link de acesso: http://ri.ufmt.br/handle/1/3843

The culture of soybean and cotton have great importance in the Brazilian economic scenario, both are commodities that move billions of reais per year in exports. The importance is demonstrated in the increased in planted areas and production year after year, keeping the country between the world ́s largest producers of crops. The weed management are of paramount importance, to achieve greater productivity year after year. However, due to the incorrect use of controls, mainly of herbicides, it has been causing resistance of some biotypes to the most popular active ingredients. Among the plants that have been representing resistance/ tolerance are those of the genus Ipomoea spp., most popularly known as Morning Glory. These plants affect soybean and cotton crops throughout their cycle, affecting their productivity. In this context, the object of this work was to evaluate the implementation of two object detection algorithms in real time (Faster R-CNN and YOLOv3), and to develop an embedded system for selective spraying of herbicides. Morning Glory plants in crops soybean and cotton, in the Cerrado Matogrossense. The project was developed at the Agricultural machinery laboratory of the Federal University of Mato Grosso, campus of Rondonopolis. The algorithms were trained to detect three classes (Soybean, Morning Glory and cotton) and evaluated in terms of precision and recall in the laboratory and field. The laboratory results of the Faster R-CNN algorithm showed results with an average accuracy of 87.20% and recall 77.20%, while the YOLOv3 tiny showed 81.16% accuracy and recall 68.00%. In the field tests, Faster R-CNN showed better results in comparison to YOLOv3 tiny in both modules analyzed, showing weed control average of 81.70% in cotton and 77.00% in soybean. The YOLOv3 tiny did not present satisfactory results in the field, presenting results less than 21.00% in the control of Morning Glory, present in the modules. The spray precision results of the Faster R-CNN demonstrate that object detection algorithms in real time for the selective control of post-emergent Morning Glory weeds in soybean and cotton crops.