COFFEE LEAF DISEASE CLASSIFICATION USING THREE LIGHTWEIGHT DEEP LEARNING MODELS: ROBUSTNESS EVALUATION AGAINST IMAGE DEGRADATION AND GRAD-CAM VISUALIZATION

Tóm tắt

This study evaluates three lightweight deep learning models, namely MobileNetV3-Small, MobileNetV3-Large, and EfficientNet-B0, for coffee leaf disease classification using the Ethiopian Coffee Leaf Disease dataset. The dataset includes four classes: Cerscospora, Healthy, Leaf rust, and Phoma, with 9,180 training images, 1,620 validation images, and 1,200 test images. All models were trained using transfer learning with ImageNet-pretrained weights, and the input images were resized to 224 × 224 pixels. Model performance was first assessed on the original test set and then further evaluated under two image degradation conditions, including blur and low-light. Grad-CAM was applied to the model with the highest overall performance to visualize image regions contributing to its predictions. The experimental results show that MobileNetV3-Large achieved the highest weighted-average F1-score on the original test images, reaching approximately 0.9992. Under degraded image conditions, the same model also obtained the highest weighted-average F1-scores, with 0.9114 under blur and 0.9464 under low-light. EfficientNet-B0 showed comparable performance on the original and low-light images but was more affected by blur. MobileNetV3-Small achieved high performance on the original images but showed a larger performance decrease under low-light conditions. Grad-CAM visualizations indicate that MobileNetV3-Large mainly focused on diseased leaf regions, while several blurred samples showed less distinct activation patterns, corresponding to some misclassification cases. The results indicate that, for lightweight deep learning models in coffee leaf disease classification, evaluation should consider not only classification performance on original images but also robustness under degraded image conditions and model interpretability.