- The paper offers a multi-level classification framework using deep learning to identify bird species, flock formations, and size categories, aiming to enhance aviation safety and does so by deploying both cascade and unified architectures.
- The Unified Classification Approach (UCA) achieves 92.80% end-to-end accuracy, while the Cascade Classification Approach (CCA) reaches 90.77% accuracy, with varying robustness under different environmental conditions.
- The paper demonstrates the feasibility of using deep learning models for species identification in aviation safety, highlighting the need for further validation in real-world airport environments.
Motivation and problem context
Bird strikes remain a persistent operational hazard, with tens of thousands of strikes reported annually in the United States and a per-flight strike rate that has risen even as total traffic fluctuated. The paper's central premise is that avian radar systems, while effective for real-time detection and tracking of birds near airports, cannot identify bird species. This is a substantive limitation because flight speed, altitude preference, and migratory behavior are species-specific, so predictive models used in strategic and tactical deconfliction cannot be tailored without species information. The authors position image-based classification—paired with cameras co-located with avian radars—as the missing component that closes this gap.
The scope of the work is threefold: (1) classify bird species from images of birds in flight, distinguishing them from aircraft; (2) identify flock formation type via a two-stage framework; and (3) estimate flock size category from visual data. The species set is grounded in FAA wildlife strike data from 1990 to 2023, covering 33 high-risk species, with size categories (small, medium, large) adopted from DeTect Inc.'s radar cross-section thresholds (0–70 g, 71–800 g, 801–1700 g).
Methodology
Two architectures are proposed for species identification:
- Cascade Classification Approach (CCA): a three-stage pipeline — binary bird/aircraft classification, then size classification into small/medium/large, then size-specific species classifiers.
- Unified Classification Approach (UCA): a single classifier that directly predicts one of 24 bird classes or an aircraft class.
All models use ResNet50V2 with ImageNet-pretrained weights, trained with Adam, categorical cross-entropy, batch size 32, and early stopping on validation AUC (patience 35 epochs, maximum 100). Benchmark comparisons are made against SVM, Random Forest (RF), and K-Nearest Neighbors (KNN), each tuned by grid search. A notable practical constraint is that benchmark models could not handle 512×512×3 inputs without memory errors, so they were evaluated at 128×128×3 while CNNs used full resolution—an asymmetry that should be kept in mind when interpreting the comparison.
For flock analysis, twelve horizontal formations are considered (Column, Front, Echelon, J, V, Inverted J/V, Closed Line, Branched V, Globular Cluster, Front Cluster, Extended Cluster). A bottom-view classifier handles all formations; if Column is detected, a side-view classifier resolves vertical alignment (Ascending, Descending, Level). Flock size is classified into five bins (5–20 through 81–100 birds), under the stated assumption that kinetic energy and impact severity vary meaningfully between bins but not within them.
Species classification results
The stage-wise results are strong. The binary bird/aircraft classifier achieved 100% test accuracy on balanced classes of 2,400 images each, though the authors correctly note this reflects a controlled dataset. The size classifier reached 94.76% accuracy across three classes using 4,200 images of 24 species (six species were dropped due to insufficient images; visually similar pairs such as Barn/Cliff Swallow were merged).
Size-specific species classifiers substantially outperformed traditional baselines:
| Task |
CNN |
RF |
SVM |
KNN |
| Large birds (7 classes) |
97.62% |
50.19% |
42.46% |
30.95% |
| Medium birds (10 classes) |
96.25% |
32.92% |
32.08% |
21.67% |
| Small birds (7 classes) |
92.86% |
47.14% |
42.14% |
33.57% |
The margin over classical methods is large enough that the conclusion does not hinge on the resolution asymmetry noted above, though it does mean the comparison is not strictly like-for-like. The UCA achieved 92.80% accuracy across 25 classes (24 bird groups plus aircraft) with 200 images per class. Composing CCA stage accuracies via the authors' end-to-end formula yields 90.77%, slightly below UCA. The paper's claim that UCA edges out CCA is modest and well-supported; the counterpoint offered—that CCA is modular and interpretable per stage—is reasonable, though the modularity comes at the cost of error propagation from intermediate stages, which the formulation makes explicit.
Robustness under environmental degradation
A robustness study simulates rain, snow, Gaussian sensor noise, and reduced brightness on test images. Key findings:
- Rain: at 50% intensity, accuracy falls to roughly 59–61% across all size groups.
- Snow: the most damaging condition; at 50% intensity, medium-bird accuracy drops to 31.67%, with large and small at 46.83% and 50.71% respectively. The medium classifier's sensitivity to partial occlusion is explicitly acknowledged.
- Sensor noise: at σ=0.40, accuracy degrades to roughly 59–61%.
- Darkness: notably resilient; the large-bird classifier stays above 96% even at 30% brightness, and medium/small degrade only slightly.
These results temper the headline accuracies: performance is reliable in benign conditions but degrades substantially under heavy precipitation or noise, which is directly relevant to deployment feasibility at airports where such conditions coincide with elevated strike risk.
Flock type and flock size results
The two-stage flock framework achieves 99.17% accuracy on bottom-view formation classification (12 classes) and 96.67% on side-view vertical alignment (3 classes); effective end-to-end accuracy for Column samples is the product of first-stage recall (1.00) and second-stage accuracy. Flock size classification reaches 85.00% overall, with the weakest class being 61–80 birds (F1 = 0.70), suggesting adjacent-bin confusion at mid-range densities.
An important caveat, which the authors state plainly: flock datasets are synthetic, generated programmatically from single pigeon templates rendered on blue backgrounds. The high flock-classification accuracies therefore demonstrate feasibility on controlled geometry rather than generalization to real aerial imagery, and validation on natural flocks remains open.
Practical insights
Several empirical findings carry methodological value beyond this application. Increasing training images per class generally improves accuracy, though not monotonically. Input resolution matters, and interestingly the required resolution scales inversely with bird size: large birds reach ~90% accuracy at 180×180, medium birds need 320×320, and small birds require 384×384. Grayscale inputs perform within about 1–2 points of RGB when channel-duplicated for pretrained ResNet, offering a viable fallback. Finally, the authors observe that fine-tuning pretrained ResNet yields diminishing returns relative to custom architectures because the frozen structure limits flexibility—a candid limitation of their chosen backbone.
Limitations and open questions
The paper is transparent about several constraints. Six of the 33 target species were excluded due to insufficient imagery, and several visually similar pairs were merged, so the deployed class set covers 24 groups rather than the full high-risk list. The benchmark comparison uses lower input resolution for SVM/RF/KNN due to memory limits. The 100% bird/aircraft accuracy reflects a curated, balanced dataset unlikely to be reproduced in the field. The flock classifiers rest entirely on synthetic data. Open questions include whether these accuracies hold on real airport camera feeds with varied backgrounds, distances, and lighting; how the system performs on species excluded here; and whether flock-size bins calibrated on kinetic-energy assumptions translate to actual damage-risk modeling.
Conclusion
This paper delivers a systematic evaluation of CNN-based multi-level classification for aviation safety, covering species identification (via both cascade and unified pipelines), flock formation type, and flock size. The consistent superiority of ResNet50V2 over tuned SVM, RF, and KNN baselines—by margins of 45–75 percentage points—and the demonstrated robustness to low-light conditions support the feasibility of camera-based species identification as a complement to avian radar. The slightly higher end-to-end accuracy of the unified approach (92.80% vs. 90.77%) favors simplicity, while the cascade retains interpretability advantages. The principal unresolved issue is generalization: synthetic flock data and curated species datasets leave the transition to operational airport environments as the critical next step before these classifiers can inform species-specific flight path prediction and deconfliction systems.
