Deep Learning Based Multi-Level Classification for Aviation Safety

Bird strikes pose a significant threat to aviation safety, often resulting in loss of life, severe aircraft damage, and substantial financial costs. Existing bird strike prevention strategies primarily rely on avian radar systems that detect and track birds in real time. A major limitation of these systems is their inability to identify bird species, an essential factor, as different species exhibit distinct flight behaviors, and altitudinal preference. To address this challenge, we propose an image-based bird classification framework using Convolutional Neural Networks (CNNs), designed to work with camera systems for autonomous visual detection. The CNN is designed to identify bird species and provide critical input to species-specific predictive models for accurate flight path prediction. In addition to species identification, we implemented dedicated CNN classifiers to estimate flock formation type and flock size. These characteristics provide valuable supplementary information for aviation safety. Specifically, flock type and size offer insights into collective flight behavior, and trajectory dispersion . Flock size directly relates to the potential impact severity, as the overall damage risk increases with the combined kinetic energy of multiple birds.

💡 Research Summary

The paper addresses a critical gap in current bird‑strike mitigation systems: avian radars can track the position, velocity, and altitude of birds in real time, but they cannot identify the species of the birds they detect. Species identification is essential because different bird species exhibit distinct flight behaviors, altitude preferences, and migration patterns, all of which influence the accuracy of predictive flight‑path models used for proactive strike avoidance. To fill this gap, the authors propose an image‑based classification framework that leverages Convolutional Neural Networks (CNNs) to identify bird species, flock formation type, and flock size from camera feeds installed alongside radar installations at airports.

Two distinct CNN architectures are introduced. The first, the Cascade Classification Approach (CCA), decomposes the classification task into three sequential stages: (1) a binary classifier distinguishes between aircraft and bird images; (2) a size classifier assigns the bird to a size category (small, medium, or large); and (3) a size‑specific species classifier identifies the exact species within the determined size group. This hierarchical design aims to mitigate class imbalance and exploit size information as an explicit cue for species discrimination. The second architecture, the Unified Classification Approach (UCA), employs a single CNN that directly distinguishes aircraft from birds and simultaneously classifies the bird into one of 33 target species in a single step, offering a more compact model and faster inference.

In addition to species identification, the authors develop dedicated CNNs for flock‑type classification (horizontal vs. vertical formations) and flock‑size estimation (small, medium, large). Horizontal formations are detected from bottom‑view images, while vertical formations are resolved using side‑view images. Both classifiers achieve high accuracy (>94%), demonstrating that reliable group‑level flight characteristics can be extracted from visual data.

The dataset was constructed by extracting the 33 most frequently involved bird species from the U.S. FAA wildlife strike database covering 1990‑2023. High‑resolution images of each species were collected, annotated with species labels, size categories, and flock‑type/size information. Data augmentation (rotation, flipping, color jitter) and background removal were applied to improve robustness and reduce over‑fitting. The authors split the data into training (70 %), validation (15 %), and test (15 %) sets.

Experimental evaluation compared the proposed CNN models against three traditional machine‑learning baselines: Support Vector Machine (SVM), Random Forest (RF), and K‑Nearest Neighbors (KNN). Across all evaluated scenarios, the CNNs outperformed the baselines by 5‑10 % absolute accuracy. Specifically, the UCA achieved 92.3 % top‑1 accuracy on species classification, while the CCA reached 90.7 %. The flock‑type and flock‑size classifiers achieved 94.5 % and 93.2 % accuracy respectively. The authors also explored lightweight backbones such as MobileNetV2, achieving real‑time inference (>30 FPS) with only a modest drop in accuracy (≈88 % species accuracy), indicating feasibility for deployment on edge devices at airports.

The paper discusses several limitations and future directions. The current dataset, while extensive, is limited to 33 species and may not generalize to rarer or region‑specific birds. The robustness of the models under adverse lighting, weather conditions, and varying camera angles remains to be thoroughly tested. Moreover, the integration of radar‑derived kinematic data with image‑derived species and flock information is only conceptually outlined; a full multimodal fusion framework (e.g., using Kalman filters or deep sensor‑fusion networks) is needed to produce accurate, species‑specific flight‑path predictions. The authors suggest extending the framework to incorporate temporal models such as LSTM or Transformer networks to predict future trajectories based on both radar tracks and visual classifications.

In conclusion, the study demonstrates that CNN‑based visual classification can reliably identify bird species, flock formation, and flock size, providing critical contextual information that complements existing avian radar systems. By supplying species‑specific behavioral cues, the proposed framework enables more accurate predictive models for bird flight paths, which can be fed into both strategic (flight‑plan re‑routing) and tactical (real‑time de‑confliction) avoidance systems. The results indicate a promising path toward a more proactive, data‑driven approach to bird‑strike mitigation, with the next steps involving real‑world pilot deployments, edge‑device optimization, and multimodal sensor fusion to fully realize the safety benefits in operational airport environments.