SSTC Summer Research Internship · Curtin University · 2025–2026

Overview

The Desert Fireball Network (DFN) operates a continent-wide array of all-sky cameras that continuously photograph the night sky to detect and track meteorites for recovery. Each camera captures a wide variety of transient optical events — meteors, satellites, lightning, lens flares, and artifacts — all mixed together in raw data.

Before any scientific analysis can happen, events need to be classified.

This internship project explored what light curve features distinguish these event types, built both rule-based and machine learning classifiers, and developed a second pipeline for detecting fragmentation events within meteor light curves.

Project management:

• Managed a 10-week research project with defined deliverables (classification pipeline, performance metrics, fragmentation detection algorithms)

• Coordinated with DFN researchers to understand requirements and validate results

• Delivered final presentation to stakeholders including research leads and technical staff

Part 1 — Classification Pipeline

Starting from raw JSON event files, I built a full processing pipeline: extract 18 features per event from the light curve and detection metadata, write to CSV, generate pair plots, and classify.

To understand the feature space I generated and manually reviewed 153 pair plots — every combination of 18 features — sorting them into useful, possible, and useless for class separation. This informed a set of conditional threshold rules for each event type.

I then automated feature space analysis using PCA, building a script that tests every combination of features for k=2–10 clusters and identifies the subset with the best class separation. This ran iteratively as the dataset grew.

Two Classifiers, Compared

Manual rule-based classifier: Conditional thresholds derived from pair plot and PCA analysis, encoded as a multi-class decision tree in Python. Required iterative refinement — particularly for long, faint satellite signals that kept falling into the wrong bucket.

RandomForest ML classifier: Trained using scikit-learn with StratifiedKFold and GroupKFold cross-validation to handle class imbalance and prevent data leakage.

Class Manual F1 RandomForest F1

Meteor 0.90 0.95

Satellite 0.95 0.94

Lens flare 0.95 0.80

Lightning 0.75 0.00

Overall 0.834 0.867

The ML model performed well overall but failed completely on lightning detection due to class imbalance. The manual classifier was more robust on harder classes. The honest conclusion: both approaches have limitations, and a larger, more balanced dataset is needed before either is production-ready.

Part 2 — Fragmentation Detection

A second research thread explored detecting fragmentation events in meteor light curves — the brightness spikes caused when a meteorite breaks apart during atmospheric entry.

I implemented two detection methods and compared them against the BLADE algorithm (Silber, 2025):

Mean brightness rolling window — averages flux over a sliding window to identify FRED (Fast Rise Exponential Decay) events. Sensitive but noisy.

Standard deviation of gradient — detects rate-of-change events using adaptive noise masking. Better handles the different noise profiles of the Forrest and Perenjori cameras, but harder to compare across events.

I proposed a hybrid approach combining BLADE's normalisation with gradient-based detection, preserving reproducibility while improving noise handling.

Tools & Stack

Python · scikit-learn · pandas · NumPy · matplotlib · Bash pipeline scripts · JSON → CSV processing · PCA · GMM clustering · RandomForestClassifier · StratifiedKFold / GroupKFold

What This Project Demonstrates

Working with space surveillance infrastructure: DFN is a continent-wide autonomous sensor network operating in harsh environments — exactly the kind of distributed systems used in satellite constellations and ground station networks.

Algorithm development and validation: Understanding how to process noisy sensor data, handle edge cases, and validate algorithmic performance under real-world conditions translates directly to aerospace applications.

Research project management: I can manage projects with ambiguous requirements, communicate technical findings to stakeholders, and deliver results on deadline.