molhamfetnah/ai-cup-2026-bird-classification

Python

0

0

🦅 AI Cup 2026 Bird Classification

#

Physics-informed machine learning for wind turbine bird strike prevention

Industrial AI system combining radar trajectory analysis with gradient boosting ensembles for real-time bird species classification. Features 59+ kinematic features, gliding ratio analysis, and turbine safety controller logic.

🎯 System Architecture

#

graph LR
    A[Radar Sensor] --> B[WKB Parser]
    B --> C[Feature Engine]
    C --> D[Kinematic Features]
    C --> E[Radar Features]
    C --> F[Temporal Features]
    D --> G[ML Ensemble]
    E --> G
    F --> G
    G --> H[Bird Classifier]
    H --> I[Safety Controller]
    I --> J{Threat Level}
    J -->|Critical| K[Turbine Shutdown]
    J -->|Elevated| L[Reduce Speed]
    J -->|Low| M[Normal Operation]

📊 Problem Description

#

Multi-class classification of bird radar tracks into 9 categories:

• Clutter - Non-bird radar returns
• Cormorants - Large water birds
• Pigeons - Urban/common birds
• Ducks - Waterfowl
• Geese - Large migratory birds
• Gulls - Coastal/seabirds (58% of dataset)
• Birds of Prey - Hawks, eagles (high-altitude gliders)
• Waders - Shore birds
• Songbirds - Small passerines (19% of dataset)

🚀 Quick Start

#

Setup Environment

#

# Activate virtual environment
source .venv/bin/activate  # Linux/Mac
# or
.venv\Scripts\activate  # Windows

# Install dependencies (already done)
pip install -r requirements.txt

Run Baseline Model

#

# Fast baseline model (~ 2 minutes)
python run_baseline.py

Output: outputs/baseline_submission.csv Performance: Average OOF Log Loss: 0.1495

Run Advanced Ensemble (Optional)

#

# Full pipeline with advanced feature engineering (~ 15-30 minutes)
python run_pipeline.py

Output: outputs/submission.csv

Test Safety Controller

#

# Demo industrial safety logic
python src/safety_controller.py

Output: Threat assessment reports for various bird detection scenarios

📁 Project Structure

#

.
├── data/                       # Competition data
│   ├── train.csv              # Training data (2,601 tracks)
│   ├── test.csv               # Test data (1,872 tracks)
│   └── sample_submission.csv  # Submission format
├── src/                       # Source code modules
│   ├── features.py           # Feature engineering (kinematic, trajectory, radar)
│   ├── train.py              # Ensemble training (XGBoost, LightGBM, CatBoost)
│   └── safety_controller.py  # Industrial turbine safety logic
├── notebooks/                 # Analysis notebooks
│   └── 01_trajectory_visualization.ipynb
├── outputs/                   # Model outputs and submissions
├── models/                    # Saved model checkpoints
├── run_baseline.py           # Quick baseline model script
├── run_pipeline.py           # Full ensemble pipeline
└── requirements.txt          # Python dependencies

🔧 Features

#

Baseline Model (14 features)

#

• Radar signatures: bird size, airspeed, altitude (min/max/range)
• Temporal: duration, hour of day, day of week, month, cyclical encoding
• Trajectory: approximate length from WKB hex string
• Observer: number of birds observed

Advanced Model (59+ features)

#

• Trajectory parsing: WKB geometry decoding for spatial coordinates
• Kinematic features (22 physics-based):
  • Velocity: 3D velocity, horizontal/vertical components, variance, coefficient of variation
  • Acceleration: Mean, max, variance (flight pattern indicator)
  • Flight patterns: Flapping vs gliding detection
  • Climb dynamics: Climb rate, descent rate
  • Gliding ratio: L/D ratio (horizontal distance / vertical drop) - distinguishes Birds of Prey
  • Altitude efficiency: Distance traveled per meter of altitude change
  • Path geometry: Turn radius, curvature, tortuosity
• RCS proxy features: Time interval variance, sampling frequency, track density
• Spatial statistics: Distance, straightness, angular changes
• Temporal patterns: Cyclical time encoding

Safety Controller Module

#

Industrial logic layer for wind turbine bird strike prevention:

• Threat evaluation: Real-time risk assessment by species and proximity
• Species risk profiles: High/medium/low impact classification
• Control actions: Turbine shutdown, speed reduction, or normal operation
• Batch processing: Multi-detection threat prioritization

🎯 Models & Performance

#

Baseline

#

• Algorithm: XGBoost
• Features: 14 basic features
• Training: 3-fold cross-validation
• Speed: ~2 minutes
• Score: 0.1495 OOF Log Loss
• Per-class: Best - Clutter (0.04), Ducks (0.06) | Hardest - Gulls (0.43), Songbirds (0.30)

Advanced Ensemble

#

• Algorithms: XGBoost + LightGBM + CatBoost
• Features: 59+ physics-based features (22 kinematic + RCS proxies)
• Training: 5-fold cross-validation per model
• Ensemble: Average predictions across all models and folds
• Speed: ~15-30 minutes
• Key differentiators: Gliding ratio, velocity variance, RCS proxies

🐳 Docker Deployment

#

Production-ready containerization with multi-stage builds:

# Build and run baseline
docker build --target production -t bird-classifier:latest .
docker run -v $(pwd)/data:/app/data bird-classifier:latest

# Development with Jupyter
docker-compose up jupyter
# Access at http://localhost:8888

# Full training pipeline
docker-compose up trainer

See DOCKER.md for complete deployment guide.

• Expected improvement: Lower log loss through ensemble diversity

📈 Results

#

Per-Class Performance (Baseline)

#

Insights:

• Clutter is easiest to predict (low log loss)
• Gulls and Songbirds are most challenging (high log loss) - likely due to:
  • Large class imbalance (Gulls: 58% of training data)
  • High intra-class variability in flight patterns

📝 Usage Examples

#

Generate Submission

#

# Quick baseline
python run_baseline.py
# → outputs/baseline_submission.csv

# Advanced ensemble
python run_pipeline.py
# → outputs/submission.csv

Load and Analyze Results

#

import pandas as pd

# Load submission
submission = pd.read_csv('outputs/baseline_submission.csv')

# Check prediction probabilities
print(submission.head())

# Verify probabilities sum (should be close to 1 for each row)
print(submission.iloc[:, 1:].sum(axis=1).describe())

🔍 Data Insights

#

• Training samples: 2,601 radar tracks

• Test samples: 1,872 radar tracks

• Class distribution: Highly imbalanced

  • Gulls: 58% (1,503 samples)
  • Songbirds: 19% (483 samples)
  • Pigeons, Waders, Birds of Prey: 3-5% each
  • Clutter, Geese, Ducks, Cormorants: <3% each

• Feature types:

  • WKB-encoded spatial trajectories
  • Radar measurements (size, speed, altitude)
  • Temporal metadata (timestamps)
  • Observer annotations (train only)

🚧 Future Improvements

#

1. Class imbalance handling:

   • SMOTE or class weights
   • Stratified sampling

2. Feature engineering:

   • Fourier transforms of trajectories
   • Statistical moments of movement patterns
   • Weather/environmental features if available

3. Model enhancements:

   • Neural networks for trajectory sequences
   • Attention mechanisms for temporal patterns
   • Stacking with meta-learners

4. Ensemble optimization:

   • Weighted averaging based on validation performance
   • Stacking ensemble (train meta-model on OOF predictions)

📄 License

#

Competition project for educational purposes.

molhamfetnah/ai-cup-2026-bird-classification

Python

0

0