<![CDATA[Manual visual inspection to assess banana ripeness is subjective and unable to meet the scale of industrial needs, while previous studies have used heavy CNN architectures and have not systematically explored the depth of fine-tuning. This study proposes a two-phase transfer learning framework using EfficientNetB0 on a pre-augmented dataset of 13.478 images across four ripeness classes: unripe, ripe, overripe, and rotten. Class imbalance is addressed through class weighting during training. In Phase 1 (Feature Extraction), all base layers are frozen, and the classification head is trained until it achieves a best validation accuracy of 98.58%. In Phase 2 (Fine-Tuning), the optimal unfrozen layer depth was determined through systematic ablation across five configurations (5, 10, 15, 20, and 25 layers), with the 25-layer configuration yielding the highest validation accuracy of 98.75%. Evaluation on 562 test images yielded an accuracy of 99.46% and a test loss of 0.0538, with an F1-score of 1.00 for the overripe and ripe classes, and 0.99 for the rotten and unripe classes. The ROC curve confirmed high discriminative capability with an AUC of 1.000 for the overripe and ripe classes, and 0.999 and 0.998 for the unripe and rotten classes. These results demonstrate that the combination of a two-phase strategy, depth ablation, and fine-tuning with class weights yields a robust classification system with potential for application in automated banana sorting using edge devices.]]>
