A Robust Machine-Learning Classification Framework for Early Detection of Charcoal and Bleeding Canker in Cork Oak Bark

The cork oak (Quercus suber) is an ecologically and economically significant species, supporting forest biodiversity, carbon sequestration, and the global cork industry. Climate-driven pathogens, specifically Biscogniauxia mediterranea and Phytophthora cinnamomi, increasingly threaten cork oak forests with charcoal and bleeding cankers, respectively, leading to rapid forest decline. Early, reliable identification of these pathogens in the field is critical for effective management, but distinguishing visually similar diseases remains challenging for non-specialist personnel. This study investigates whether a small set of representative bark images can provide an interpretable, field-oriented approach to cork oak disease classification. Using a dataset of 1,198 bark images including healthy, B. mediterranea-infected, and P. cinnamomi-infected samples, we extracted features with MobileNet-v3 Small and applied K-Means clustering to identify pseudo-centroids, or real, representative images closest to each cluster’s center. These pseudo-centroids serve as interpretable reference points that field workers can visually compare with new samples, rather than relying solely on "black box" model outputs. A k-Nearest Neighbor (KNN) classifier evaluated against these pseudo-centroids achieved 74.90% accuracy across the three classes. To assess resilience to human labeling errors, a common source of error in field-collected datasets, we simulated noise by randomly swapping pseudo-centroid labels, which resulted in only a 5.44% decline in accuracy, demonstrating that the framework degrades gracefully rather than catastrophically under noisy conditions. This robustness, combined with the interpretability of the pseudo-centroid approach, makes the framework practical for forest rangers and conservation workers to prune or treat infected limbs based on direct visual similarity ranking, without requiring deep machine learning or biological expertise. Given the environmental and economic importance of cork production, this approach offers a scalable, field-ready tool for monitoring the health of cork oak forests. Future work will expand the framework to larger, more diverse datasets and additional pathogens, broadening its applicability across cork-producing regions.