Leakage-audited cross-dataset evaluation reveals large generalization gaps in deep learning models for apple and maize disease recognition

Deep learning models for plant disease recognition often reach near-perfect accuracy on curated benchmark images, yet how well this performance carries over to independently collected images is rarely measured directly. We trained five image-classification networks (ResNet50, EfficientNet-B0, DenseNet121, ConvNeXt-Tiny and ViT-B/16) on PlantVillage apple and maize leaf images under one shared protocol and tested each trained model, without adaptation, on independently sourced images: 1,730 apple images from Plant Pathology 2020 and 371 web-sourced maize images. Before training, an exact-file audit removed seven duplicate PlantVillage images and found no exact or perceptual near-duplicates shared between source and external datasets. On held-out PlantVillage images, Macro-F1 ranged from 0.949 to 1.000. On external images it fell to 0.472–0.723 for apple and 0.209–0.375 for maize. Exact McNemar tests with Holm correction identified significant differences in per-image correctness for several model pairs. Because each architecture was trained once with one seed, architecture-specific comparisons are treated as exploratory. The results show that near-perfect benchmark scores should not be interpreted as evidence of field readiness; independent external testing and leakage auditing are essential.