Aircraft Bird Strike Risk Prediction Using LSTM-RNN, XGBoost, and BNN

This study employed three machine learning algorithms to address limitations in Next Generation Radar (NEXRAD)-based bird strike forecasting by modeling eight weather features from the National Oceanic and Atmospheric Administration (NOAA) to predict aircraft and bird strike risk up to three hours in advance. The target risk feature was developed from the Department of the Air Force Safety Center’s AHAS NEXRAD risk models. An ensemble of LSTM-RNN and XGBoost achieved the highest accuracy (80% to 93%) across four airfields, with an F1 score of approximately 0.80. Entropy-based confidence intervals from the Bayesian Network were the most stable, outperforming the bootstrapping approach used in LSTM-RNN and XGBoost. This study demonstrated that a single ensemble can make stable risk predictions across multiple airfields with reasonable accuracy, providing aviation professionals with a potential tool to improve risk management during preflight planning. GitHub location (https://doi.org/10.5281/zenodo.15760325) contains the full Python script.