Learning to Explain Air Traffic Situation

A research team has proposed a machine-learning framework designed to decode how air traffic controllers perceive complex traffic situations, using real-world surveillance data from Incheon International Airport in South Korea [1]. The framework, detailed in a paper submitted on 15 Feb 2025 and revised through 25 Jun 2026, employs a Transformer-based multi-agent trajectory model to capture both the spatio-temporal movement of aircraft and the social interactions between them [1]. By deriving attention scores from the model, the researchers can quantify the influence of individual aircraft on overall traffic dynamics, offering what they describe as explainable insights into a controller's mental picture of the airspace [1]. The model was trained on surveillance data collected from the terminal airspace around Incheon International Airport, the main international gateway serving Seoul [1]. The airport opened on 29 March 2001, replacing Gimpo International Airport for most international traffic, and was constructed on reclaimed land between Yeongjong and Yongyu Islands [9]. Previous attempts to model controller strategies often centered on specific tasks or pairwise aircraft interactions, neglecting the comprehensive dynamics of an air traffic situation [1]. The new framework addresses this gap by treating the traffic scene as a multi-agent system where attention mechanisms highlight which aircraft are most salient to the evolving situation [1]. This approach intersects with the broader field of dynamic decision-making, which studies how people use experience to control complex, real-time systems that change due to their own actions or external events [5]. The work arrives as aviation stakeholders continue to examine tools that can augment human performance in high-stakes environments. While the paper does not report operational deployment, the authors suggest the framework could potentially support and enhance controller decision-making and situational awareness [1]. The four versions of the submission ranged in size from 1,562 KB to 5,367 KB, reflecting iterative refinements over sixteen months [1].