A Comprehensive Evaluation of Machine Learning Algorithms for Harvested Arecanut Variety Classification and Detection

Traditional image processing techniques for nut classification often struggle due to the subtle visual differences between nut varieties, natural variations within varieties, and complexities in isolating nuts from cluttered backgrounds. This study explores the proficiency of machine learning and deep learning methods for accurate and efficient classification and detection of arecanut varieties. The Roboflow tool was used to meticulously label and annotate a dataset of 9,000 images, categorizing them into Isolated Green Arecanut, Isolated Yellow Arecanut, Clustered Green Arecanut, and Clustered Yellow Arecanut. The dataset also includes varied real-life scenarios to ensure robust validations of the model. Our classification process used classifiers like Support Vector Machines, K-nearest neighbor, and Random Forest. Object detection was performed using YOLOv8 and YOLO-NAS architectures. The study involved preprocessing, feature extraction (for classification), model training, and rigorous evaluation using metrics like accuracy, precision, recall, and mean average precision (mAP). This comparative analysis provides insights into the strengths and weaknesses of each algorithm, paving the way for robust automated systems in the agriculture industry. Among the four different classifiers, SVM yields the highest accuracy of 88%, highlighting its effectiveness in the classification task. Among the two detection algorithms, YOLOv8 has achieved 79.5% mean average precision and YOLO-NAS has achieved 78.5% mean average precision, while the best precision was produced by YOLO-NAS (82.5%).