Tabular-to-Image Deep Learning for Multi-Factor Aviation Bird Strike Risk Prediction
- 1School of Graduate Studies, College of Aviation, Embry-Riddle Aeronautical University, Daytona Beach, FL 32114, USA
- 2Department of Computer Science, University of North Carolina at Charlotte, Charlotte, NC 28223, USA
Abstract
Wildlife strikes are a persistent aviation safety concern because bird strike risk is shaped by dynamic interactions among weather, bird movement, airport operations, seasonality, and time of day. Although traditional machine learning methods have been used to model structured aviation safety data, they may not fully capture the temporal and cross-feature relationships that contribute to elevated risk. This study proposed a time-aware tabular-to-image deep learning framework for multi-factor aviation bird strike risk prediction. Hourly weather observations from Houston Intercontinental Airport (IAH) were integrated with NEXRAD radar-based bird risk data to construct a three-class prediction task: Low, Moderate, and Severe risk. The final dataset included 43,732 hourly observations. These classifications represent radar-derived avian hazard levels rather than confirmed bird strike events or damage severity. To preserve temporal structure, tabular weather, time, and train-derived risk prior features were transformed into 20 × 48 feature-by-time images representing a 48 h historical window. These images were used as inputs to a residual squeeze-and-excitation convolutional neural network (CNN) and compared against traditional statistical, machine learning, and baseline CNN models. The proposed 48 h weather history Res-CNN achieved 79.74% accuracy and a macro F1-score of 78.74%, compared with 80.02% accuracy and 79.09% macro F1 for LightGBM. Although the proposed model did not outperform the strongest tree-based baseline, it achieved competitive performance and substantially outperformed the original 4 × 5 feature grid CNN. These findings demonstrate the feasibility of incorporating structured temporal information into tabular-to-image deep learning while highlighting the continued effectiveness of tree-based methods for structured aviation risk data.
Related
- Conformal Prediction and Conditional Coverage for Tabular Foundation Models
- SurvFM enables tabular foundation models for right-censored survival prediction
- Tabular foundation models are not a shortcut to calibrated clinical risk prediction: a pre-registered simulation against the true risk
- LLM knowledge for improved Out-of-Distribution Tabular Prediction - A Cross-Validated Regularisation Approach
- Model-based Bootstrap for Offline Policy Evaluation in Tabular Reinforcement Learning
BibTeX
@article{pan2026tabulartoimage,
title = {{Tabular-to-Image Deep Learning for Multi-Factor Aviation Bird Strike Risk Prediction}},
author = {Bill Deng Pan and Yupeng Yang and Jason Powell and Dahai Liu},
journal = {Mathematics},
year = {2026},
month = nov,
doi = {10.3390/math14214015},
url = {https://www.mdpi.com/2227-7390/14/21/4015},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Tabular-to-Image Deep Learning for Multi-Factor Aviation Bird Strike Risk Prediction
AU - Bill Deng Pan
AU - Yupeng Yang
AU - Jason Powell
AU - Dahai Liu
PY - 2026
DA - 2026/11/03
JO - Mathematics
DO - 10.3390/math14214015
UR - https://www.mdpi.com/2227-7390/14/21/4015
AB - Wildlife strikes are a persistent aviation safety concern because bird strike risk is shaped by dynamic interactions among weather, bird movement, airport operations, seasonality, and time of day. Although traditional machine learning methods have been used to model structured aviation safety data, they may not fully capture the temporal and cross-feature relationships that contribute to elevated risk. This study proposed a time-aware tabular-to-image deep learning framework for multi-factor aviation bird strike risk prediction. Hourly weather observations from Houston Intercontinental Airport (IAH) were integrated with NEXRAD radar-based bird risk data to construct a three-class prediction task: Low, Moderate, and Severe risk. The final dataset included 43,732 hourly observations. These classifications represent radar-derived avian hazard levels rather than confirmed bird strike events or damage severity. To preserve temporal structure, tabular weather, time, and train-derived risk prior features were transformed into 20 × 48 feature-by-time images representing a 48 h historical window. These images were used as inputs to a residual squeeze-and-excitation convolutional neural network (CNN) and compared against traditional statistical, machine learning, and baseline CNN models. The proposed 48 h weather history Res-CNN achieved 79.74% accuracy and a macro F1-score of 78.74%, compared with 80.02% accuracy and 79.09% macro F1 for LightGBM. Although the proposed model did not outperform the strongest tree-based baseline, it achieved competitive performance and substantially outperformed the original 4 × 5 feature grid CNN. These findings demonstrate the feasibility of incorporating structured temporal information into tabular-to-image deep learning while highlighting the continued effectiveness of tree-based methods for structured aviation risk data.
SN - 2227-7390
ER -