Interpretable Prototype-Based Classification of Bark Symptoms of Charcoal Canker and Bleeding Canker in Cork Oak (Quercus suber L.)

Cork oak (Quercus suber L.) woodlands are ecologically and economically important and are threatened by a decline syndrome in which the endophytic ascomycete Biscogniauxia mediterranea (De Not.) Kuntze (charcoal canker) and the soil-borne oomycete Phytophthora cinnamomi Rands (root rot and bleeding cankers) play prominent roles. This study investigates whether a small set of real, representative bark images can provide an interpretable, field-oriented aid for classifying bark symptoms. We assembled 968 publicly available source photographs (361 asymptomatic, 434 charcoal-canker and 173 bleeding-canker images) labeled from their source captions as asymptomatic, charcoal-canker symptoms, or bleeding-canker symptoms; the labels describe visible symptoms and were not confirmed by laboratory diagnosis. To balance the classes, 227 geometrically augmented copies of bleeding-canker photographs were added (bringing the total image count to 1195). Features were extracted with a frozen MobileNetV3-Small network, and per-class k-means or k-medoids clustering was used to select the real images closest to the cluster centers (“pseudo-centroids”), which were then used as references for k-nearest neighbors (KNN) classification. In 30 repeated splits that kept each source photograph and all its augmented copies on the same side of the split and tested on source photographs only, the practical configuration with four reference images per class reached a mean accuracy of 73.6 ± 3.4%, and the configuration with ten per class reached 77.8 ± 2.8%. On the original single split of 239 images, in which augmented copies of some training photographs also occur in the test set, the corresponding accuracies were 72.8% (95% CI 66.8–78.1%) and 79.1% (73.5–83.8%). Exact McNemar tests and repeated splits showed no statistically significant differences among the three KNN variants, and geometric or AutoAugment-type augmentation of the minority class did not change accuracy. Swapping the labels of two reference images reduced accuracy by 9.8 ± 3.8 percentage points with four references per class and by 4.8 ± 2.1 points with ten. Bleeding-canker symptoms were the most difficult class, reflecting their high visual heterogeneity, overlap with the other classes, and small sample size. Using all training images as references raised accuracy to 91.1%, quantifying the cost of a compact card set. The reference cards are therefore best viewed as a visual screening and training aid that complements, rather than replaces, established field keys and laboratory confirmation.