Short-term prediction of avian hazard risk across airfields using machine learning and weather data
Year: 2026
Type: article
Abstract: This study employed three machine learning algorithms to address limitations in Next Generation Weather Radar (NEXRAD)-based bird activity risk forecasting by modeling eight weather features from the National Oceanic and Atmospheric Administration (NOAA) to predict Aircraft Hazard Advisory System (AHAS) risk categories up to three hours in advance. The prediction target was the AHAS NEXRAD-derived risk score provided by the Department of the Air Force Safety Center, which characterizes bird activity within radar sectors rather than observed bird-strike events. Long short-term memory recurrent neural network (LSTM-RNN), XGBoost, and Bayesian multinomial logistic regression (BMLR) models were evaluated across four US airfields. An ensemble of LSTM-RNN and XGBoost achieved the highest predictive accuracy, ranging from approximately 80% to 93% across the four airfields, with F1 scores of approximately 0.80. McNemar tests indicated statistically significant differences in classification performance among the evaluated models. The results demonstrate the feasibility of using weather-informed machine learning to extend AHAS risk prediction beyond contemporaneous NEXRAD observations and support multi-hour aviation risk assessment across geographically distinct airfields. Such forecasts could supplement existing bird-activity information during preflight planning and operational risk management. The full Python scripts, imputed datasets, and original AHAS risk data used in the analysis are archived in Zenodo (DOIs: 10.5281/zenodo.15760325, 10.5281/zenodo.22310818, and 10.5281/zenodo.22310839).
Language: English
Subfield: Aerospace Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,400
OpenAlex ID: W7133504672
