A Novel Brazilian Dataset for Automatic Detection of Psyllid Attacksin Guava

• Iago Magalhães de Mesquita UFC http://orcid.org/0000-0003-4347-7979
• Larissa Ferreira Rodrigues Moreira UFV https://orcid.org/0000-0001-8947-9182
• Rhuan da Silva Nunes UFC https://orcid.org/0009-0001-2625-8937
• Iális Cavalcante de Paula Júnior UFC https://orcid.org/0000-0002-2374-4817

Resumo

Guava, a fruit native to America, plays a crucial role in the Brazilian economy, with production reaching 564,000 tons by 2022. It is widely used in culinary dishes and herbal remedies, and is exported to numerous countries. However, guava is also susceptible to pests and diseases. Among these, Psyllid Triozoda Limbata is particularly problematic, attacking young leaves and leading to reduced fruit yields and impaired plant development. To address this issue and assist producers in detecting pest infestations, this study introduces a novel dataset comprising images from Brazil. By leveraging Artificial Intelligence and Computer Vision techniques, this dataset provides a foundation for developing pest detection systems. We evaluated various object detection architectures including different versions of YOLO and Vision Transformer (ViT). The experiments indicated that YOLO V5 achieved promising results, with a precision of 93.14%. Moreover, the availability of this dataset opens new opportunities for enhancing pest management strategies, improving crop protection, and increasing the overall productivity in the guava industry.

Palavras-chave: Psyllids, Deep Learning, Guava Trees, Object Detection