Lightweight CNN and MobileNetV2 with feature selection and explainable AI for efficient insect detection in potato plants

Efficient insect detection in agricultural fields is crucial for crop health and yield optimization. Traditional image-based methods often struggle with generalization and computational cost. This paper presents a novel framework combining feature selection and explainable artificial intelligence (XAI) to enhance detection accuracy, efficiency, and interpretability. Using a curated dataset comprising 3000 annotated images of pests and beneficial insects collected from diverse agricultural environments, we evaluated three feature selection techniques—Mutual Information (MI), Recursive Feature Elimination (RFE), and Principal Component Analysis (PCA). RFE achieved a 40% reduction in feature space, resulting in a 12% improvement in classification accuracy and a 35% reduction in inference time. Our optimized MobileNetV2 model achieved 93.4% accuracy, with a precision of 92.8%, recall of 91.7%, and F1-score of 92.2%. Explainable AI techniques such as SHAP and LIME were employed to visualize and interpret model decisions, thereby ensuring transparency and trust in model predictions. The proposed framework demonstrates superior performance and interpretability over conventional methods, paving the way for real-time, farmer-friendly pest management systems deployable on low-power devices. • Developed a lightweight CNN–MobileNetV2 model for potato plant insect detection. • Applied feature selection (RFE) to improve accuracy and reduce computational cost. • Achieved 93.4% accuracy with 35% faster inference using optimized features. • Employed SHAP and LIME for transparent and interpretable model predictions. • Designed a real-time, low-power solution for farmer-friendly pest management.