Biotic and abiotic stresses reduce crop yields, yet existing image‑based diagnostic models are often crop‑specific, computationally intensive, and poorly interpretable. We present: (i) a multi-crop diagnostic platform integrating lightweight convolutional neural networks (CNNs) with management recommendations, (ii) a hybrid ultralightweight CNN architecture with triple attention mechanisms: SE (Squeeze-and-Excitation), CBAM-Lite (convolutional block attention module), and coordinate attention, optimized for efficient edge deployment; and (iii) a crop transferability index (CTI), a single scalar index, to quantify cross-crop generalization. This framework leveraged eight transfer‑learning CNNs to detect 34 stress classes across 11 crops from RGB images. We assembled 47263 RGB images (44, 623 field & 2, 640 web-scraped), randomly stratified split into training/validation/test sets (70/15/15), and applied 5-fold cross-validation on the training set for model tuning. Gradient‑weighted class activation saliency maps (Grad‑CAM) and Platt-scaling prediction confidence scores were used to localize stress‑related regions and enhance model interpretability. Achieving 99. 9% test accuracy, the hybrid model, ConvNeXtV2ₐtto, significantly outperformed baseline CNNs, matched ResNet18weight, and offered 8-14 times fewer parameters than conventional backbones with the most expeditious CPU inference (27. 7 FPS). Statistical analysis (Friedman test with Nemenyi post hoc, p < 0. 001) confirmed the superiority of the hybrid, ConvNeXtV2ₐtto and ResNet18weights backbones. The CTI up to 45. 11% and edge-device performance metrics further reinforce these findings, with MobileNetV2, V 2 PlantNet and SqueezeNet1₁ remaining competitive alternative due to their lower computational cost. However, expanding work on zero-shot and unsupervised domain adaptation is needed, as Fréchet inception distance explains only 10–11% of CTI variance (R² = 0. 10–0. 11).
Basnet et al. (Wed,) studied this question.