A Validation-Driven Explainable Deep Ensemble Framework for Image-Based Saffron Adulteration Detection

Saffron (Crocus sativus L.), one of the world’s most valuable spices, is highly vulnerable to adulteration due to its premium market price and the limitations of conventional analytical methods for rapid, non-destructive authentication. Although recent deep learning-based approaches have reported promising accuracy, many rely on single models or naïve ensembles and lack rigorous validation and statistical reliability assessment. This study proposes a validation-driven and explainable deep ensemble framework for image-based saffron adulteration detection. Multiple pretrained convolutional neural networks (DenseNet169, ResNet50, and VGG16) are integrated using a validation-driven weighted ensemble strategy in which fusion weights are computed exclusively from validation performance within the training folds and fixed prior to evaluation on the held-out fold, thereby preventing information leakage between model selection and performance assessment. The proposed framework achieved 98.61% classification accuracy, 98.17% F1-score, and 98.61% AUC, outperforming the best individual base model by up to 1.4% in F1-score. Stratified five-fold cross-validation demonstrated stable performance, with a mean accuracy of 97.81% ± 0.53, confirming robustness across data splits. Statistical validation using McNemar’s test (p < 0.05) and 5 × 2 cross-validated significance testing verified that performance improvements over constituent models are statistically reliable. Grad-CAM-based explainability and background-invariance analysis further demonstrated that predictions are driven primarily by intrinsic filament-level characteristics, with only a marginal (~0.9%) performance reduction under ROI-cropped evaluation. The proposed framework provides a robust, interpretable, and statistically validated solution for saffron authentication and offers methodological insights for reliable image-based food adulteration detection under limited data conditions.