Identificação de cultivares do gênero Brachiaria e Panicum utilizando aprendizado profundo

• Ano de defesa: 2024
• Autor(a) principal: Leonardo Lazarino Crivellaro
• Orientador(a): Edson Takashi Matsubara
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Fundação Universidade Federal de Mato Grosso do Sul
• País: Brasil
• Palavras-chave em Português: Inteligência Artificial, Forrageiras, Aprendizado Profundo
• Link de acesso: https://repositorio.ufms.br/handle/123456789/10011

Tropical forages are plants that cover pasturelands and are the main source of nutrition for cattle raised on grazing systems, which accounted for 81.8% of Brazilian beef production in 2023. Each forage variety, known as a cultivar, has distinct nutritional qualities and specific climate, soil, and management requirements. Ensuring that pastures are covered with forages of adequate nutritive value, productive potential, and appropriate usage suited to regional conditions requires the acknowledgment of the plant's genus, species, and cultivar. However, the automated and precise identification of forage cultivars remains an unsolved challenge, although neural network models have already shown good results in recognizing genera and species. This work presents an approach based on deep learning models for identifying Brachiaria and Panicum maximum cultivars using images collected at different stages of plant development to leverage distinct morphological characteristics across developmental phases. The research was conducted in an experimental area at Embrapa Beef Cattle, where 18 cultivars were planted, and images were captured during their first 8 months of growth. The neural networks analyzed and tested were MobileNet v3 and MobileVIT, as low computational cost for execution was a prerequisite to integrate this technology into the Pasto Certo® mobile application. The networks were tested with different combinations of datasets by plant growth stages. The best results were achieved in the reproductive stage due to differences in inflorescence, where the MobileNet v3 architecture achieved 82% accuracy, and MobileVIT achieved 87%, with the latter showing better generalization capabilities. The constructed datasets contribute to research in the field, while the trained models have the potential to become important tools for assisting producers and technicians in pasture management.