Fungal diseases such as Purple Blotch, Downy Mildew, Stemphylium Leaf Blight, and Botrytis Leaf Blight pose serious threats to onion production, particularly in developing agricultural regions where access to expert diagnosis is limited. This study proposes and evaluates a mobile-deployable, lightweight convolutional neural network (CNN) based system for automated classification of fungal diseases in onion leaf images captured under real field conditions. A dataset comprising 9,300 annotated onion leaf images across five classes (four fungal diseases and healthy leaves) was utilized. Three lightweight CNN architectures, MobileNet, EfficientNet, and a custom-designed lightweight CNN, were trained, optimized, and comparatively evaluated using accuracy, precision, recall, F1-score, confusion matrix analysis, and inference latency. Statistical validation using oneway Analysis of Variance (ANOVA) revealed a significant effect of model architecture on classification performance (p < 0.001), with a large effect size (η² = 0.62, corresponding to Cohen’s f = 1.27). Subsequent Tukey’s Honestly Significant Difference (HSD) post-hoc tests confirmed that the MobileNet-based model achieved statistically superior performance compared to both EfficientNet and the custom lightweight CNN. Post-training optimization techniques, including transfer learning and INT8 quantization, were applied to improve mobile readiness. The best-performing model was successfully deployed on Android devices using TensorFlow Lite, enabling real-time, offline inference. Field validation with onion farmers further demonstrated high usability and practical effectiveness, confirming the suitability of lightweight CNNs as efficient and accessible decisionsupport tools for real-world agricultural disease detection. The proposed system offers a practical, accessible, and efficient decision-support tool for onion farmers, contributing to improved disease management and agricultural productivity. Future work may explore additional disease classes, multi-modal data integration, and broader device-level evaluations.

Keywords: Lightweight CNN, EfficientNet, onion leaf diseases, MobileNet, TensorFlow Lite, Android, automated disease detection