Técnicas de aprendizado ativo para avaliação do vigor de sementes de soja

Abstract

Growing seed companies increasingly seek excellence in production quality through rigorous processes such as the tetrazolium test and the definition of vigor. However, these are extremely laborious processes, since the experience of a specialist is necessary, as well as visual analysis of a considerable quantity of seeds as sampling for determining the vigor of seed lot. Moreover, although the tetrazolium test has a defined protocol, this analysis may vary from analyst to analyst because it is a subjective human process. In this context, several efforts highlight the relevance of the topic and have been carried out in an attempt to automate the analysis process, in order to reduce the problems intrinsic to it. Thus, this work presents methodologies for processing soybean seeds from the tetrazolium test, as well as for learning and classification of vigor. In addition, a new approach to active learning is proposed to select more informative samples for learning. To validate the proposals, an extensive experimental evaluation is carried out, using different sets of seeds and state-of-the-art techniques for description and learning. From the results obtained, it is possible to observe that the proposed approach allows obtaining more robust classifiers, which reach higher accuracy faster (in less learning iterations) in relation to traditional supervised learning approaches. Therefore, it is expected to minimize the effort and the time of the specialist’s involvement in his laboratory routine, during the process of visual analysis and manual annotation. The faster and more precise process makes it possible to increase the market competitiveness between the producers and the beneficiation of consumers with higher quality products and more appropriate and specific prices according to the quality of the seed lot analyzed. Loss of productivity in the processes of planting, multiplication and/or commercialization of low quality seeds can also be avoided.