Detection and Classification of Apple Diseases using ML Techniques: A Leakage-Free Deep-Learning Framework with Honest Cross-Dataset Evaluation

Abstract

Many published apple-leaf disease systems report near-perfect accuracy on the PlantVillage collection although augmented copies of the same physical leaf are spread across the training and test partitions, so that the model is tested on images. This study rebuilds the four-class apple subset as a de-duplicated, leakage-free benchmark and measures what the leak actually changes. Retraining the same ResNet50 on the original, leaky split shows that the leak is invisible in the laboratory metric (99.98% versus 99.66% accuracy, and 100% even on test images whose leaf was never seen in training) but not in the field, where the leaky model is less accurate on the external FGVC8 dataset (39.0% versus 43.9% over three seeds). Five ImageNet-pretrained backbones fine-tuned under one protocol all reach 99.35--99.80% in-domain accuracy over five seeds, a statistical tie, so cost rather than accuracy decides deployment. External zero-shot accuracy on two real-field datasets is low and varies between training runs (PlantDoc 59.5±5.7%, FGVC8  43.9±0.5%); most errors are healthy orchard leaves predicted as diseased, and geometric augmentation is the only augmentation family that measurably narrows the gap, whereas GAN-based oversampling of the minority class widens it. Unsupervised domain-adversarial training raises FGVC8 accuracy from 37.2% to 54.4% without field labels, but fine-tuning on only 68 labelled field images reaches 76.7%. A data-efficiency study with three backbones and two fine-tuning budgets shows that leakage-free laboratory pre-training makes adaptation faster (+0.21 macro-F1 with 80 field images under a short 15-epoch budget) but does not raise the accuracy reached once fine-tuning converges. Grad-CAM places 86% of its activation inside the leaf, with the class ordering confirmed by an independent GrabCut segmentation, and a YOLO11 detector localizes diseased leaves in the field at 93.5% mAP@0.5, ahead of YOLOv8 in a matched three-seed comparison (92.9% versus 88.1%).