Soybean Seed Quality Assessment Classification Using EfficientNet B0 Algorithm
Isaac Angelo M. Dioses, Jesusimo L. Dioses, Russell N. Aquino, Catleen Glo Feliciano, James Bryan Tababa, Loida F. Hermosura
Abstract
This study employs a convolutional neural network (CNN) Efficientnet B0 model to classify soybean seeds into five groups: broken, immature, intact, skin-damaged, and spotted. To enhance model performance, more than 5,000 photos were preprocessed and enhanced with noise, flipping, and contrast tweaks. An SGDM optimizer with a learning rate of 0.01 and mini batch size was set to 32 and 30 epochs was used for training after the dataset was split into training and testing in a 75:25 ratio. The model's validation accuracy was 93.83 % and its training accuracy was 100%. High classification accuracy was demonstrated by the confusion matrix, especially for undamaged soybeans, which had the highest F1 score of 97.35%. Although the model worked well for all classes, there were a few misclassifications because to similarities with other categories, especially for broken soybeans. The outcomes show how well the CNN model can classify soybeans, which may have applications in evaluating the quality of agriculture.