A statistically validated stacking ensemble of CNNs and vision transformer for robust maize disease classification

Early disease detection in crops remains an age-old problem in the quest for global food security. While deep learning has transformed image-based diagnosis, most existing models in maize leaf disease follow single-view approaches, offer limited validation, and neglect the accuracy-computational efficiency trade-off. In this regard, we suggest a robust framework to combat these issues with a heterogeneous stacking ensemble. Our suggested model combines four various convolutional neural networks DenseNet201, InceptionV3, NASNetMobile and VGG19 with a Vision Transformer. The model benefits from both local feature extraction and more global contextual analysis in trying to learn a more holistic representation of disease markers. We extensively evaluated this ensemble on a challenging dataset of 15,995 field images collected in Ethiopia and Kaggle dataset, utilizing stratified fivefold cross-validation for stable assessment. The model exhibited stable performance with a mean validation accuracy of 99.13% across validation folds a statistically significant improvement p < 0.05 over the best individual model, as assessed by a paired t-test. The ensemble recorded 99.15% accuracy on an independent test set, outperforming state-of-the-art lightweight models such as MobileNetV3 97.62%. The findings set a new baseline for maize leaf disease classification. The most important contribution of this work is the statistically validated excellence of the heterogeneous ensemble, along with transparent analysis of computational requirements and an extensive overfitting countermeasure. While the computational demands make this solution more amenable to server-side rather than edge device implementation, it is still incredibly reliable and generalizable as a diagnostic tool. A prototype was also developed and warmly received by users, further demonstrating its real-world applicability to drive data-driven sustainable agriculture.