CNN-Based Identification of Longan Varieties Using Leaf Vein Patterns

Abstract

Visual classification of longan seedlings remains challenging due to the similarity of characteristics among varieties, particularly in young leaves. This study applies the Convolutional Neural Network (CNN) method to classify five types of longan seedlings—Diamond River, Matalada, Merah, Itoh, and Pingpong—based on leaf vein patterns, which serve as distinctive features. The dataset consists of 1,000 high-resolution images, divided into 900 for training and 100 for testing. The training process includes preprocessing steps such as cropping to focus on vein patterns, resizing to standardize input dimensions, augmentation to enhance data variety, normalization to scale pixel values, and splitting into training and validation sets. Hyperparameter tuning was performed using a grid search, evaluating combinations of learning rate, batch size, and epochs. The best configuration was achieved at the 80th epoch, with a learning rate of 0,0001 and a batch size of 8. The model achieved a validation accuracy of 0,8444 and a loss of 0,3865. During testing, it reached an accuracy of 0,8000, with an average precision of 0,8266, recall of 0,8000, and f1-score of 0,7843. The best performance was observed in the Merah and Matalada classes, while the Diamond class remained challenging due to visual similarities. CNN proved effective for this task, though further improvement is needed for visually similar classes to enhance classification accuracy.