Aplicação de um algoritmo de detecção de ervas daninhas baseado em redes neurais na pulverização pontual de herbicidas utilizados na produção agrícola

Abstract

The use of pesticides, both in Brazil and globally, is the subject of research related to intoxication, soil and water contamination, and cancer-related trends. In this context, new studies and projects aimed at increasing the efficiency of these compounds are being disseminated in academic circles. Currently, artiĄcial intelligence is increasingly used in agriculture to seek gains and improvements in agronomic processes. Thus, the main purpose of this work was to develop a weed detection model with the intention of applying this model in spread spray systems, a promising technique that reduces the amount of herbicides applied through variable-rate application. The YoLo algorithm, which utilizes a single neural network in the Darknet, was employed for class detections. Other studies with the same scope were examined to provide support and scientiĄc basis for the work, resulting in detections of four weed species: caruru, juá-de-capote, vassoura, and trapoeraba. At the end of the neural network training, an accuracy of 51% was achieved for the juá-de-capote weed class, 31% for the caruru class, 29% for the trapoeraba class, and 17% for the vassoura class.