Comparison of CNN Architectures via Transfer Learning for Avocado Fruit Disease Classification

Abstract

Anthracnose and scab reduce the commercial quality of avocado fruit, and their visual diagnosis can be difficult when lesions present similar features. This study compares VGG16, ResNet50, EfficientNetB2, and InceptionV3 using transfer learning to classify healthy, anthracnose-infected, and scab-infected fruits. The protocol employed laboratory images: 569 for five-fold stratified cross-validation and 100 independent images for the final test. All four architectures were trained under the same two-phase tuning scheme, online data augmentation, class weighting, and evaluation using macro metrics; the selected model was further examined using a confusion matrix, ROC/AUC curves, and Grad-CAM. VGG16 achieved the best performance in cross-validation (accuracy = 0.817 ± 0.054; macro F1 = 0.808 ± 0.060), whereas InceptionV3 ranked fourth (macro F1 = 0.731 ± 0.038). In the independent test, VGG16 reached 74% accuracy and 0.72 macro F1. The main source of error was the classification of anthracnose as scab; likewise, anthracnose presented the lowest AUC (0.767). Grad-CAM maps concentrated attention on fruit lesions, although they constitute qualitative interpretive evidence. The results are limited to controlled conditions: no field images from Apurímac were used. Therefore, regional validation and deployment on lightweight devices are proposed as future work.