Bird Strike Risk in Aviation: A Calibrated Threshold-Optimized Machine Learning Decision Support

Bird–aircraft strikes are rare but high-impact events that pose persistent safety and operational challenges in aviation. Existing approaches often focus on predictive accuracy, with limited integration of probability calibration, rare-event precision, and decision-making under operational constraints. This paper proposes a calibrated threshold-optimized decision-support framework that anticipates short-term bird strike risk and translates probabilistic forecasts into constrained intervention plans. The framework combines (i) a calibrated occurrence model for short-term strike probability, (ii) a conditional intensity model for expected strike magnitude, and (iii) a constrained planning module that selects a limited number of high-risk time windows per day. The approach is evaluated using EASA bird strike data (2019–2024), enriched with environmental, meteorological, astronomical, tide, and traffic-proxy features. After this stricter temporal reconstruction, the calibrated occurrence model reaches ROC-AUC ≈0.636 and PR-AUC ≈0.050 across forward chain folds, while the OOF aggregate gives ROC-AUC =0.648 and PR-AUC =0.047 against a forward label prevalence of 0.0246. Under a strict daily alarm budget (K=2), OOF operational Precision@K is 0.063 and Recall@K is 0.075, with a median credited lead time of 90 min. The results highlight the importance of calibrated probabilities and decision-oriented evaluation, showing that effective rare-event management requires not only prediction, but also integration with operational planning.