Lightweight Deep Learning Approach for Early Detection of Lemon Leaf Diseases Using a Modified MobileNetV2 Architecture

Abstract

Lemon (Citrus limon L.) cultivation is vital to agricultural economies, but its productivity is threatened by diseases such as Anthracnose, Bacterial Blight, Black Spot, Citrus Canker, Citrus Leaf Miner, Curl Leaf, Curl Virus, Deficiency Leaf, Dry Leaf, Greening Disease (HLB), Healthy Leaves, Melanose, Sooty Mould, Spider Mites. This work presents a lightweight deep learning model based on Enhanced MobileNetV2 with targeted architectural modifications for early lemon leaf disease detection. Using a custom dataset of 3,285 images across fourteen classes, the model achieved79.42% accuracy, against baseline CNN models including ResNet50, InceptionV3, Xception, EfficientNetB0. Experiments were conducted with 32 epochs, batch size 32, and Adam optimizer (learning rate 0.0001). Comparative analysis demonstrates that the proposed approach delivers superior performance with lower computational overhead, making it suitable for mobile and IoT-based agricultural applications.