Random Forests in the Supervised Classification of Multidimensional Images of the Tetrazolium Test

The quality of the soybean seed can be influenced by several factors that may occur at any stage of production. Mechanical damage, deterioration by humidity and the damage caused by bed bugs are among such problems. The tetrazolium test is adopted by the seed industry, especially for testing soybeans, due to its accuracy, fast result, and the large amount of information it provides. Digital processing and image analysis can be used to aid the extraction and classification of standards for minimizing the subjectivity implicit in the test, thus allowing more credibility to the information. The aim of this work is testing the effectiveness of Random Forests in the supervised classification of soybean embryos images submitted to the tetrazolium test. In order to do so, we used the Trainable Weka Segmentation plugin to perform the segmentation process, and the WEKA software to evaluate the quality of the classifier model obtained. During the process, 222,646 instances among 230,388 instances were correctly classified (96.7%), with Kappa index of 0.95, showing the classifier excellent performance regarding the proposed dataset. The supervised classification, combined with pixel-based segmentation, proved to be efficient in extracting more coherent visual information on seed damage. Also, we conclude that the choice of image attributes, along with the algorithm used in the work, showed to be competent in the classification process of high dimensionality samples.