Aviation Contrail Detection & Tracking
AI pipeline for ground-based contrail detection, tracking, and flight attribution
9/10
Thesis Grade
End-to-End
Pipeline
>90% acc
Detection
71.75%
Attribution
Problem
Contrails and contrail-cirrus account for roughly 40% of aviation's total warming effect, yet no validated pipeline existed for ground-based contrail monitoring with flight attribution.
Approach
Built a 3-stage AI pipeline: Mask R-CNN (ResNet-101 FPN) for contrail detection and classification, multi-object tracking evaluated with DeepSORT vs Norfair (Norfair selected - MOTA 43.3%, IDF1 67.9%), and a novel 4D-to-2D projection system for attributing tracked contrails to specific flights. Processed data at scale using Azure Databricks and Apache Spark.
Detection
Mask R-CNN segmentation
- ResNet-101 FPN backbone
- Contrail type classification (short, thin, wide, cirrus)
- >90% class accuracy
Tracking
Multi-object tracking
- DeepSORT vs Norfair evaluation
- Norfair selected: MOTA 43.3%, IDF1 67.9%
- Stable track identities across frames
Attribution
Novel 4D→2D projection
- Geometric & convolution-based methods
- 3D flight paths → 2D camera images
- 71.75% overall accuracy
Results
First validated end-to-end ground-based contrail monitoring pipeline. Thesis graded 9/10. Achieved >90% class accuracy for detection, 71.75% overall attribution accuracy (95–100% for short contrails), and two novel attribution methods (geometric projection and convolution-based).
Detection & tracking demonstration