Data-driven runway and taxiway exits prediction of landing aircraft: A case study at Hartsfield–Jackson Atlanta International Airport

Airport surface operations increasingly limit system-wide performance at high-throughput hubs. We study the airport surface arrival taxi-in problem at Hartsfield–Jackson Atlanta International Airport (KATL) and propose a two-stage, data-driven decision aid that emulates controller workflow: Stage I predicts the runway exit selected by an arriving aircraft; Stage II predicts whether, conditional on that exit, the aircraft will cross the active departure runway at a designated point or take the end-around taxiway. Models are trained on ASDE-X surface trajectories, aircraft characteristics, ramp destinations, short-horizon traffic rates (arrivals, departures, crossing usage), and airport weather across multiple look-back windows. We benchmark nine classification methods spanning linear models, neural networks, and tree-based ensembles, with Random Forest, XGBoost, LightGBM, and CatBoost as primary candidates, and report accuracy as well as imbalance-aware metrics (macro-F1, precision–recall curves, confusion matrices, Brier score, Expected Calibration Error). Across east and west flows, XGBoost and LightGBM outperform Random Forest. Stage I achieves accuracies of 0.86–0.89 with macro-F1 of 0.40–0.50; Stage II achieves 0.70–0.74 accuracy with macro-F1 of 0.28–0.55. Feature importance analyses indicate that approach speed is the dominant driver of exit choice, while departure rate, crossing rate, ramp destination, and (for west flow) the taken exit are the strongest predictors of crossing vs. end-around. Although overall accuracies are high, minority classes exhibit lower recall, which motivates explicit imbalance handling (e.g., class weighting, focal objectives) for operational deployment. A class overlap analysis using t-SNE and UMAP identifies feature-space inseparability as the primary bottleneck for minority-class prediction. The framework is positioned as a controller-support tool intended to enhance ATCO situational awareness through calibrated, explainable predictions while preserving human judgment and operational responsibility for the final routing decision. The results provide interpretable, airport-specific insight into when tactical runway crossings are favored relative to the end-around and illustrate how predictive models can support controller situational awareness and surface-flow efficiency within future human–machine collaboration environments.