Classification of Bean Leaf Lesions Using Modified EfficientNetV2 for Implementation in TensorFlow Lite

Bean leaf diseases such as Angular Leaf Spot and Bean Rust significantly threaten the productivity and quality of common beans (Phaseolus vulgaris L.), a critical crop for global food security. While existing approaches like ensemble methods demonstrate high classification accuracy, their reliance on multiple architectures increases computational complexity, making them unsuitable for resource-constrained devices. This research develops a lightweight yet high-performing model for classifying bean leaf diseases using the EfficientNetV2 architecture and its modified variant. Key metrics evaluated include classification accuracy, inference time, and TensorFlow Lite (TFLite) model size, emphasizing suitability for real-time deployment. EfficientNetV2 was fine-tuned on a dataset containing healthy leaves, Angular Leaf Spot, and Bean Rust. To reduce complexity, the architecture was modified by selectively removing specific blocks. The model was optimized for deployment through TFLite conversion and post-training quantization, significantly reducing size and inference time. Grad-Cam visualizations enhanced interpretability by highlighting regions influencing model predictions. Experiments evaluated trade-offs between accuracy and efficiency. The modified EfficientNetV2B0 achieved 97.76% test accuracy, a model size of 6.18 MB, and an inference time of 0.0594 seconds. These results demonstrate improved efficiency compared to larger models and ensemble approaches, while maintaining competitive accuracy. This research highlights the feasibility of deploying lightweight, accurate models for real-time agricultural applications. Future work could explore integrating ensemble techniques or federated learning to further enhance robustness and scalability in agricultural disease detection.