Lightweight Deep Convolutional Models for Jackfruit Leaf Disease Identification

Abstract

This work evaluates three lightweight convolutional neural network (CNN) architectures MobileNetV2, EfficientNet, and NASNetMobile for automatic recognition of six jackfruit leaf conditions: Healthy, Algal Spot, Anthracnose, Leaf Spot, Phyllosticta, and Sooty Mold. A transfer-learning approach was applied, where pretrained ImageNet weights were retained and only the classifier layers were trained. Among the three architectures, MobileNetV2 achieved the highest accuracy at 97.11%, followed by EfficientNet at 96.15%. NASNetMobile yielded a lower result of 77.00%, indicating limitations in its ability to distinguish certain disease categories under the given configuration. The evaluation further showed strong per-class scores for MobileNetV2, including perfect F1-scores for the Healthy and Sooty Mold classes. The findings confirm that compact CNN models, especially MobileNetV2, are well-suited for agricultural disease monitoring and can be integrated into mobile or edge-based decision-support systems for field use.

Access to Document

Other files and links

• Link to publication in Scopus

Fingerprint