Avaliação da eficácia de detecção de pragas em culturas de soja utilizando visão computacional

Abstract

Early identification of diseases and pests in soybean crops is essential to avoid significant production losses, since these pathogens can partially or completely compromise the work. Traditional methods of manual monitoring and inspection make the process time-consuming, subject to human error and unfeasible for large areas of plantation. In this context, the use of computer vision and deep learning techniques emerges as a promising alternative, allowing the automated detection of anomalies and enabling the adoption of preventive or corrective strategies more efficiently.

This work proposes the development of a solution based on deep neural networks for the classification of diseases in soybean plants. Initially, the model was trained with images captured on the ground, aiming to understand and improve the detection capacity before transitioning to aerial images obtained by drones. The main intention of this approach is to provide a tool that assists producers in the identification of diseases on a large scale, reducing costs and optimizing decision-making in work management.

The experiments conducted found that the difference between the terrestrial images used for training and the aerial images applied in the inference significantly impacted the model performance. The prediction in a real environment, using the DenseNet201 architecture, presented limitations due to the discrepancy between the training and inference domains, highlighting the need for adjustments in the image processing pipeline and in the model generalization strategy. These results indicate the importance of domain adaptation and dataset refinement techniques to improve the applicability of the solution in real scenarios.