A Comprehensive Survey and Reference Architecture for AI-Powered Autonomous Drone Systems in Smart Cities

Despite a rapid rise of AI-powered Unmanned Aerial Vehicle (UAV) deployments in smart city environments, current surveys and frameworks lack a unified, protocol-level reference architecture that integrates multi-domain applications, edge AI perception, cognitive reasoning through Large Language Models (LLMs), and regulatory compliance within a single deployable specification. This study presents a comprehensive cross-domain review of AI-powered drone systems for traffic management, delivery, infrastructure inspection, disaster response, and environmental monitoring. The study introduces COMPASS (Cognitive Operations Model for Programmable Autonomous Smart-city Systems), a novel seven-layer technical reference architecture that describes communication protocols (MAVLink 2.0, ROS2/DDS, MQTT 5.0, and NGSI-LD), edge computing hardware recommendations for five drone payload tiers, and quantified performance requirements for safety-critical operations. The key feature of COMPASS is its LLM-based Semantic Middleware Layer, which allows for context-aware decision-making, natural human-drone interaction, and regulatory compliance verification. Comparing COMPASS to many other frameworks reveals that it is the only architecture to simultaneously provide multi-domain coverage, protocol-level specifications, hardware recommendations, LLM integration, and empirically verified benchmarks.