Systematic Evaluation of Cross-Dataset Domain Shift and Shortcut Learning in Lightweight CNNs for Tomato Disease Classification

Abstract

The reliable deployment of deep learning architectures for automated plant pathology remains a significant challenge due to the pronounced performance gap between laboratory-curated datasets and real-world field conditions. This study presents a rigorous investigation into cross-dataset domain shift and shortcut learning in lightweight convolutional neural networks (CNNs). Using PlantVillage (5,431 images) for source training and PlantDoc (436 images) for target evaluation across four shared tomato disease classes, we assessed ShuffleNetV2, MobileNetV3-Small, and ResNet18 via a seven-phase framework integrating Grad-CAM interpretability and causal masking interventions. Domain drops ranged from 54.2% to 60.1%, with ShuffleNetV2 achieving the best cross-dataset accuracy (44.63%), outperforming the larger ResNet18 (39.07%). Phase-7 experiments showed models maintained up to 35.52% accuracy on backgroundonly pixels—well above the 25% random baseline—strongly suggesting substantial reliance on contextual shortcuts over true disease features. Statistical testing confirmed standard augmentation $(p>0.64)$ provides no significant robustness gains. These findings necessitate a shift towards segmentation-guided learning and domain-invariant feature extraction for practical agricultural deployment.