Processamento e reconhecimento de imagens para aplicação de herbicidas com alta precisão

Abstract

Over the last century, the population has quadrupled, leading to a considerable increase in the demand for food and the necessary raw material for large-scale production. With this growing demand, innovation in mechanisms that make agricultural processes more efficient, cost-effective, and less harmful to the environment is increasingly necessary. The early development stages, including the emergence stage (when the plant begins to emerge from the soil) and the tillering stage (characterized by vertical growth and the generation of new leaves), are critical for crop development. These initial stages are often hindered by weed competition, as invasive plants compete for the same resources. To address this issue, the application of herbicides is an essential control method for large-scale plantations. Currently, herbicide application methods are complex, involving manual and inflexible planning and rate calculations, which are prone to errors and result in unnecessary time and financial costs. This can lead to low-quality production, excessive raw material usage, and environmentally unsustainable practices. To optimize these crucial processes, this study focuses on the research and development of crop and weed identification, plant counting, and growth estimation using a YOLOv8 object detection neural network model. The model is trained with an image dataset of sugar beet crops, employing preprocessing techniques and data augmentation. The results are also presented, along with performance comparisons against another model.