Classificação multirrótulo de sementes de soja utilizando técnicas de aprendizado de máquina: uma abordagem com rótulos suaves e fusão de características

Abstract

This work investigated the effectiveness of machine learning techniques for multi-label classification of soybean seeds using the dataset developed in TCC I. the experiments demonstrated the viability and advantages of the multi-label approach for automated seed quality assessment, significantly contributing to precision agriculture advancement. The research achieved its proposed objectives, generating relevant contributions for both the scientific community and practical agricultural applications. The results empirically confirmed that approximately 65% of seed images present multiple characteristics simultaneously, validating the relevance of multi-label classification in contrast to traditional single-label methods. The Soft Labels (%) strategy demonstrated consistent superiority, achieving Mean Average Precision (MAP) of 93.74% with Support Vector Regression (SVR). Feature fusion by averaging extracted from Vision Transformer (ViT-Base) proved to be the most effective strategy, combining multiple orientations of each seed for a more complete evaluation. This technique reduced the Mean Squared Error (MSE) by 54.1% compared to individual image analysis. SRV demonstrated consistent superiority over k-Nearest Neighbors (KNN), with 7.4% higher MAP and 25.4% lower MSE, evidencing better capability for modeling complex relationships between visual characteristics and multi-label classifications.