Cotton pest and disease diagnosis via YOLOv11-based deep learning and knowledge graphs: a real-time voice-enabled edge solution

Abstract

Introduction: High labor costs, limited expert availability, and slow response hinder cotton pest and disease management. We propose a real-time, voice-enabled edge solution that integrates deep learning–based detection with a domain knowledge graph to deliver accessible, field-ready decision support. Methods: We construct a cotton pest–disease knowledge graph with over 3,000 triples spanning seven major categories by fusing expert-curated and web-sourced knowledge. For image recognition, we develop an enhanced YOLOv11 detector compressed via LAMP pruning and a teacher–assistant–student distillation strategy for lightweight, high-performance deployment on Jetson Xavier NX. Detected objects are semantically aligned to graph entities to generate context-aware recommendations, which are delivered through Bluetooth voice feedback for hands-free use. Results: The optimized model has 0.3M parameters and achieves (Formula presented.) = 0.835 at 52 FPS on the edge device, enabling stable real-time inference in field conditions while preserving detection accuracy. Discussion: Coupling a compact detector with a structured knowledge graph and voice interaction reduces dependence on expert labor and speeds response in non-expert settings, demonstrating a practical pathway to scalable, intelligent cotton pest and disease management at the edge.