Classification of Mango Varieties from Leaf Images Using ResNet-50 CNN Architecture

Abstract

Early identification of mango (Mangifera indica L.) varieties is crucial for optimizing cultivation, as each variety possesses distinct characteristics and regional adaptability. However, the morphological similarities in leaf shape and texture especially among Harumanis, Erwin, Cokanan, Gedong Gincu, and Mahatir pose challenges for novice farmers and hobbyists. This study proposes a classification system using the Convolutional Neural Network (CNN) method with the ResNet-50 architecture to classify mango leaf varieties based on image data. A total of 5,000 images were collected and augmented from 1,250 original samples using a high-resolution camera under controlled indoor conditions. The dataset was split into training (80%), validation (10%), and testing (10%). Sixteen experimental configurations were evaluated using combinations of image resolutions (160×160 and 320×320 pixels), learning rates (0.01, 0.001), batch sizes (16, 32), and training epochs (50, 100). The best results were achieved using a 320×320 image size, learning rate of 0.001, batch size of 32, and 100 epochs, yielding a validation accuracy of 89.9%, precision of 89.87%, recall of 89.9%, and F1-score of 89.83%. These results confirm that high-resolution images and fine-tuned hyperparameters significantly enhance classification performance. The findings demonstrate the effectiveness of the ResNet-50 model for fine-grained classification in agriculture and support its future deployment in real-world environments for cultivar identification, quality control, and intelligent crop management.