Deployable deep learning for cross-domain plant leaf disease detection via ensemble learning, knowledge distillation, and quantization

Accurate leaf disease detection via smartphone-based deep learning holds immense potential for mitigating global crop losses. However, significant deployment challenges persist when transitioning from controlled laboratory environments to real-world agricultural conditions. Despite recent advances, three fundamental barriers remain: cross-domain generalization, severe class imbalance, and computational limitations for edge deployment. This study introduces the first open cross-domain benchmark for tomato leaf disease detection, unifying PlantVillage and TomatoVillage datasets into 15 harmonized disease classes to enable reproducible evaluation across domains. We propose a unified optimization approach integrating ensemble learning, knowledge distillation, and quantization across 24 deep learning architectures for edge-compatible disease detection. Strategic data augmentation and ADASYN-based balancing mitigate the severe 75:1 class imbalance, while systematic hyperparameter tuning optimizes model configurations. Our four-model ensemble (DenseNet-121, ResNet-101, DenseNet-201, EfficientNet-B4) achieves 99.15% accuracy via soft-voting. Knowledge distillation transfers ensemble capabilities to compact ShuffleNetV2, maintaining 98.53% accuracy with 163× parameter reduction and 43.6× speedup. INT8 quantization provides 671× compression (1.46 MB) while sustaining 97.46% accuracy, enabling 0.29ms inference. Cross-dataset validation demonstrates robust generalization with only 3.45% performance degradation. Grad-CAM++ and LIME-based explainability confirm biologically grounded attention patterns aligned with plant pathology principles. Finally, field deployment via a multilingual and multi-platform Flutter application validates real-world feasibility, establishing the first scalable framework bridging research and practical agricultural deployment. This work sets a standardized benchmark and extensible methodology for future multi-dataset precision agriculture research.