A Corruption-Robustness Benchmark for Mango Fruit Disease Classification

Chemoecology · Springer Nature · 2026-11-05 · DOI: 10.1007/s10341-026-02078-7

Аннотация

Mango diseases reduce both yield and fruit quality. Manual inspection is slow and carries a real risk of error. Accuracies above 95% are routine in the mango disease classification literature, but the studies differ in how they split the data, preprocess the images, and configure training, so the numbers are not comparable, and none of them reports what happens once image quality degrades. We evaluated 15 methods on the MangoDHDS dataset using a single protocol: the same five folds, preprocessing, and training configuration. The 15 cover nine architectures released between 2014 and 2024, four variants that isolate multi-scale fusion and attention modules, a linear probe on a frozen foundation model, and a color-histogram difficulty baseline. All nine architectures fall between 93.33% and 95.80%, a range no wider than the fold-level deviation. The ablation shows that adding multi-scale fusion and an attention module together yields no measurable gain: 46 predictions change, 23 for the better and 23 for the worse (McNemar, p = 1.00). None of the 14 deep methods beats the best single model, and 17 images defeat every method at once; 15 of these images belong to a single class pair. Across 11 corruptions generated with the ImageNet‑C protocol, the same models are spread by 17.5 points, and the difference is attributable to the architecture family rather than the pre-training regime or model capacity. The lightweight architecture normally selected for field deployment is the one that fails most often under corruptions simulating field conditions. On this dataset, clean test accuracy is not a sufficient criterion for model selection; behavior under corruption is discriminative and points to a different choice.

Читать у издателя →

Похожие статьи