Evaluation of deep learning approaches in the identification and classification of fungal spores from microscopy slides

Soybean cultivation is one of the principal crops in agricultural production in Brazil, moving a relevant agribusiness production chain. However, there are problems during the production process, involving care from planting, development to harvesting. In particular, the control of pests such as fungi is of great relevance to production. Monitoring and decision-making for the prevention and application of fungicides in the sector has become a major focus in terms of sustainability, productivity, and healthier production for human and animal consumption. Computational approaches have been applied to pest monitoring, however they focus on detecting the phenotype of diseases, when cultivars already show signs of infection and require fungicide application. However, fungi reproduces through living hosts and uses the wind to migrate its spores to a new host, such as Asian soybean rust. One possibility for early identification, before the manifestation of the pathology, is to detect the presence of spores by spore collectors and microscopy slides. But this leads to manual analysis performed by a specialist, being a time-consuming and tiring process. Deep Learning approaches can assign greater accuracy when counting spores with less time, finding patterns and leading to classification. Therefore, this study presents an evaluation of different deep learning approaches for the automatic recognition of spores that cause soybean diseases: rust, downy mildew, and powdery mildew from microscopy slides.