Pests and diseases cause yield losses of 20–30% and are projected to rise sharply to as much as 60%, posing a significant threat to global food security. In response, various studies have applied artificial intelligence; however, they often fall short in leveraging information compression and feature refinement to optimize model parameters effectively and in incorporating Explainable artificial intelligence (XAI) techniques to enhance transparency in smart agriculture. To address these gaps, we introduce a Knowledge distillation (KD) architecture that combines Inception-ResNetV2 (teacher) and ResNet-50 (student) models, enhanced with Grad-CAM-based explainability, to improve both performance and interpretability on a rice pest classification dataset comprising over 3,000 images across 10 classes. The proposed approach achieved accuracy, precision, recall, and F1-score of 82.50%, 82.35%, 82.50%, and 82.13% (more than 9% of the baseline model). Furthermore, the Intersection over Union (IoU) metric indicates that the teacher model effectively distilled relevant knowledge, enabling the student model to learn critical features. Nevertheless, XAI-based analysis also reveals the model’s limitations in handling class imbalance and visually similar features, suggesting directions for future improvement.