Towards Automatic Soybean Cultivar Identification: SoyCult Dataset and Deep Learning Baselines
Resumo
Soybean, the most economically important crop in Brazil, is projected to reach a production of 154 million metric tons in 2023, solidifying its position as the world's largest soybean producer and exporter. In this context, automatic soybean cultivar identification can play a crucial role in numerous applications, including ensuring seed homogeneity and preventing the presence of invasive seeds in crops. To contribute towards these advancements, this paper presents SoyCult, a publicly available dataset comprising seed images for accurate soybean cultivar identification. The dataset is created using a contour-based segmentation method proposed to extract individual seed images from captured input images, encompassing seeds from various cultivars commonly grown in Brazil. The performed experiments first assess a range of baseline systems, which combine a pre-trained convolutional network with a top-level classifier, representing a widely utilized deep learning-based strategy for classification tasks. Subsequently, selected baseline systems are subjected to additional evaluation using the Friedman test, followed by the Nemenyi post-hoc test. The SoyCult dataset, along with the source code developed to support this study, is publicly available online at https://github.com/eliezersflores/SoyCult.
