A novel ensemble approach for crop disease detection by leveraging customized EfficientNets and interpretability

Abstract

Crop leaf diseases present a persistent and serious threat to agricultural productivity and food security, especially in agro-based countries. An effective resolution of this issue demands the development of automated methods for the timely detection and management of crop diseases. In this work, we present a novel ensemble technique for the automatic detection of crop diseases, utilizing four diverse datasets: corn, potato, wheat, and tomato, each containing images of both healthy and disease-affected crop leaves. While previous studies often employed basic transfer learning (TL) techniques, we aimed to improve TL performance by systematically integrating different versions of EfficientNet and customizing their architectures with additional layers. A key contribution of our research is a novel model selection method for ensemble learning, which goes beyond traditional accuracy metrics by addressing misclassifications and class-specific shortcomings. We developed a tailored approach using misclassification counts and Hamming Loss to redefine the model selection process, identifying the most suitable EfficientNet models for each dataset. We applied Gradient Class Activation Mapping (Grad-CAM) to visualize the model’s prediction process and integrated Shapley Additive Explanations (SHAP) to enhance interpretability by providing detailed insights into feature contributions. Thus, we introduced an efficient and transparent technique for automatic crop disease detection, achieving over 99% accuracy, precision, recall, and F-Score across all datasets, significantly outperforming existing methods.