Ethiopian Soybeans Seed Type Classification Using Deep Learning

Abstract

Soybean, a crucial oil crop worldwide, is mainly grown for its protein-rich seeds containing highlevels of both protein and oil. It ranks second in global oil consumption after palm oil, holdingsignificant economic value as both a food commodity and a source of foreign exchange. InEthiopia, the current method of identifying soybean seed varieties relies on manual classificationby domain experts, a process that is time-consuming, labour-intensive, and prone to errors. Thus,there is a need for an efficient classification model to streamline the selection of soybean seeds.This study aims to experimentally evaluate a proposed system for classifying soybean seedvarieties using deep learning models. The dataset comprises 7,200 images representing eightdifferent soybean varieties: Cheri, Jalale, Dhidhessa, Korme, Boshe, Katta, Ethio-Yugoslavia,and Gute. The images were divided into 80% for training and 20% for testing. Pre-processingtechniques such as Median and Gaussian filtering were applied to remove noise introducedduring image acquisition, and image augmentation was used to expand the dataset. Three deeplearning models, Custom CNN, InceptionV3, and VGG16, were trained on both original and pre-processed images. The performance of each model was evaluated using metrics such asaccuracy, recall, f1-score, precision, and confusion matrix, under various conditions. CustomCNN, InceptionV3 and VGG16 achieved accuracies rates of 100%, 98.25%and 99.39%respectively, across the different image sets. The overall better performance was achieved by theCustom CNN model with Median filtering, attaining an accuracy of 100%. This model is poisedto significantly aid agricultural research institutes and the Ethiopian Commodity Exchange inswiftly and accurately identifying soybean varieties.