<![CDATA[Durian is a high-value horticultural commodity in Indonesia, yet variety identification based on leaf morphology remains largely manual and prone to subjective error. This study develops a deep learning-based classification system for durian leaf varieties using MobileNetV3Small, designed for practical deployment on mobile and edge computing devices in agricultural field settings to support real-time variety identification by farmers and agricultural practitioners. A dataset of 1,680 images across four durian varieties Bawor, Musang King, Duri Hitam, and Super Tembaga was constructed through direct field collection and public dataset acquisition, followed by augmentation via rotation and flipping techniques. Preprocessing incorporated background removal, cropping, resizing to 224×224 pixels, and Sobel edge extraction to enhance morphological features such as leaf veins and contours. Transfer learning from ImageNet weights was applied, with training conducted in two phases: a 30-epoch warm-up with frozen base layers, followed by 90-epoch fine-tuning of the 20 deepest layers selected to adapt high-level semantic features while preserving general low-level representations and avoiding catastrophic forgetting using cosine annealing and early stopping. The model achieved a validation accuracy of 98.42% and a macro average F1-score of 0.9842, with only 4 misclassifications out of 253 test images. All misclassifications occurred exclusively between the morphologically similar Bawor and Duri Hitam classes. Comparative analysis against eight prior studies using similar durian leaf classification tasks suggests that the proposed approach achieves competitive performance relative to studies employing larger architectures such as ResNet and ConvNeXt, though direct comparison on an identical dataset was not conducted and remains a direction for future work.]]>
Copyrights © 2026
