Deep learning for early detection of guava fruit anthracnose and fruit fly infestation

Anthracnose and fruit fly infestation are major biotic constraints on guava (Psidium guajava) production in South Asia, causing substantial yield losses and post-harvest degradation. Early-stage symptoms (lesions < 2 mm, subtle puncture marks) are often imperceptible to the naked eye, making timely intervention difficult. This study develops a deep learning-based classification system for early detection of both conditions using a dataset of 473 annotated guava fruit images from Bangladesh, verified by a plant pathologist, and pre-processed with unsharp masking and Contrast Limited Adaptive Histogram Equalization (CLAHE). After data augmentation, four architectures ResNet50, EfficientNetB3, DenseNet121, and Vision Transformer were evaluated using transfer learning with ImageNet pretrained weights. EfficientNetB3 achieved the highest performance: 98.5% accuracy, 98.3% precision, 98.2% recall, and 98.2% F1-score, with minimal cross-class confusion between anthracnose and fruit fly (ROC-AUC = 0.993). Ablation studies confirmed that CLAHE and unsharp masking together improved accuracy by 4.3% over baseline. Computational efficiency analysis showed that EfficientNetB3 (12.3 million parameters, 13 MB quantized size) achieves 214 ms inference on a Raspberry Pi 4B, confirming edge-deployment feasibility.