Comparação de modelos de deep learning na detecção de esporos de fungos em lâminas de microscópio

Abstract

One of Brazil’s principal crops is soybeans, which plays an important role in the Brazilian agribusiness chain. Negative effects on this crop have a major impact on the economy, from the initial process to harvesting and distribution. One of the key problems is the existence of pests in the cultivars, which are of great relevance to production. Monitoring and decision-making for the prevention and application of fungicides at planting have shown sustainability, productivity, and healthier production for consumption. Computational approaches have been applied to monitoring pests, but they are being applied to the phenotype of diseases, so that cultivars already show signs of infection and require fungicide application. However, fungi use living hosts for their reproduction and the wind as a means of migrating their spores to a new host, as is the case with the principal cause of production losses, Asian soybean rust. Spore collectors and microscope slides are used for early detection of the pathology, but this requires a specialized professional and is a time-consuming and tiring process that is not scalable. Deep learning approaches can provide greater accuracy in detecting spores in less time and in a standardized way. Therefore, this study presents an evaluation of different deep learning approaches for the automatic recognition of spores causing soybean diseases: rust, downy mildew, and powdery mildew from microscopy slides. The results presented indicate the suitability of the methods and point to directions for further research.