<![CDATA[Eggplant (Solanum melongena) is a vital agricultural commodity, but its yield is highly vulnerable to foliar diseases. Early and accurate detection using deep learning is essential for effective crop management. However, deploying automated detection in real-world agricultural settings faces two primary challenges: (1) severe classification bias caused by complex background noise and data imbalance, and (2) extreme computational constraints that hinder the deployment of conventional deep learning models on farmers' edge devices. This study presents a robust methodology for classifying four conditions of eggplant leaves (Healthy, Leaf Spot, Mosaic Virus, and Insect Pest) by implementing an automated background removal technique and targeted data augmentation, resulting in a perfectly balanced dataset of 1,400 images. Furthermore, this research conducts a comparative analysis between two distinct categories of lightweight Convolutional Neural Networks (CNNs): MobileNetV3-Large (representing ultra-lightweight architectures with 224x224 input resolution) and EfficientNetV2-B2 (representing medium-lightweight architectures with 260x260 input resolution). The models were evaluated based on their accuracy, loss convergence, and computational efficiency using an 80:20 data split and early stopping callbacks to prevent overfitting. Experimental results demonstrate that both models achieved exceptional performance. EfficientNetV2-B2 exhibited superior stability and precision, achieving a peak validation accuracy of 97.50% and a validation loss of 0.073. Meanwhile, MobileNetV3-Large reached a validation accuracy of 96.07% with significantly faster training iterations. These findings indicate that while EfficientNetV2-B2 is highly recommended for precision-critical agricultural diagnostics, MobileNetV3-Large remains a formidable alternative for deployment on edge devices with extreme computational constraints]]>
