Efficient Net B3 For The Classification of Citrus Leaf Diseases

Abstract

Global citrus production is significantly impacted by diseases, leading to economic losses and requiring accurate diagnostic tools. The system proposed here presents a deep learning system leveraging the EfficientNetB3 architecture for automated detection of Citrus Greening (HLB), Black Spot, Canker, and healthy instances. Trained on approximately 7,500 augmented images primarily sourced from the PlantVillage repository, the pre-trained EfficientNetB3 model was meticulously fine-tuned. This involved integrating a custom classification head comprising Global Average Pooling, Dense layers with ReLU activation, and Dropout regularization and employing a two-stage training strategy to adapt powerful learned features to specific citrus disease symptoms. Rigorously evaluated, the system achieved a target test accuracy of 97.5%, with robust per-class performance. This research demonstrates fine-tuned EfficientNetB3 as a powerful, scalable tool for precision agriculture, offering farmers rapid, image-based citrus disease identification to enhance crop management and mitigate losses.