Neel Prasad

Research · IEEE CHANDICON 2026

Predicting space debris collision risk with machine learning

The paper: Machine Learning Prediction of Space Debris Collision Risk Using Orbital Data. It asks whether a model can tell high-risk satellite situations from low-risk ones using only public orbital data.

Neel Prasad (Round Rock High School) with Subham Kumar Jalan and Kavitha Sharanappa Gudadur (iResearcher). Presented at the 2026 IEEE Chandigarh Subsection International Conference, hosted by Punjab Engineering College, Chandigarh, August 20 to 22, 2026.

Read the paper (PDF)See the certificate

At a glance

0.99
accuracy, precision, recall and F1 for both Random Forest and Gradient Boosting
138 of 139
test cases classified correctly by the Random Forest (109 low-risk, 30 high-risk)
Inclination
the orbital parameter that mattered most in predicting risk

Why it matters

The number of satellites and debris objects in low Earth orbit has grown fast, and more objects mean more chances of collision. A collision can damage working satellites and create new debris, and even tiny fragments are dangerous because of their speed.

Most existing collision prediction relies on mathematical calculations and fixed warning thresholds. They still work, but they are getting harder to apply as the amount of orbital data grows. This paper tests whether machine learning can help with satellite monitoring and space traffic management.

How it was done

  1. 01

    Data

    Two-Line Element (TLE) orbital records from public CelesTrak repositories. They are openly available and widely used in satellite tracking, so others can reproduce the work.

  2. 02

    Preprocessing

    Converted the raw line-format TLE text into a table with Pandas. Pulled out inclination, eccentricity, mean motion, right ascension of the ascending node, argument of perigee and mean anomaly. Collision risk labels came from a weighted orbital risk score.

  3. 03

    Analysis

    Distribution plots, correlation analysis, scatter plots and boxplots to understand the data and spot unusual values, then feature importance to see which parameters drive the prediction.

  4. 04

    Models

    Random Forest and Gradient Boosting classifiers sorted cases into low-risk and high-risk, trained on a training split and judged on a held-out test split with accuracy, precision, recall, F1, confusion matrices and ROC curves.

The orbital data

Most satellites in the dataset sit in low, nearly circular orbits, and their inclinations fall into planned bands.

Histogram of satellite mean motion values, peaking near 14.5 with smaller counts toward 13 and 15.5
Fig. 1. Distribution of satellite mean motion values.
Histogram of orbital eccentricity, with most satellites very close to zero and a long thin tail up to about 0.05
Fig. 2. Distribution of orbital eccentricity. Most orbits are close to circular.
Histogram of orbital inclination showing two tight groups, one near 74 degrees and a smaller one near 86 degrees
Fig. 3. Distribution of orbital inclination. Satellites sit in distinct orbit bands.

What relates to risk

Inclination showed the strongest link to collision risk. Eccentricity showed a wider spread in the high-risk group.

Correlation heatmap of orbital parameters and collision risk, with inclination showing a correlation of 0.74 with collision risk
Fig. 4. Correlation heatmap of the orbital parameters and collision risk. Inclination has the strongest positive relationship.
Scatter plot of inclination against mean motion, with low-risk points in a band near 74 degrees and high-risk points in a band near 86 degrees
Fig. 5. Inclination against mean motion, colored by collision risk class.
Boxplots of eccentricity for low-risk and high-risk classes, with the high-risk class spread more widely and several outliers
Fig. 6. Eccentricity across the two risk classes. The high-risk group spreads wider.

How the models did

Both models performed almost identically. On the Random Forest confusion matrix, one high-risk case was missed and there were no false alarms.

Confusion matrix for the Random Forest model: 109 low-risk cases correct, 0 false alarms, 1 high-risk case missed and 29 high-risk cases correct
Fig. 7. Confusion matrix of the Random Forest classifier.
ROC curve for the Random Forest model hugging the top-left corner, showing near-perfect separation
Fig. 8. ROC curve of the Random Forest model.
Bar chart of Random Forest feature importance, with inclination far above mean motion, eccentricity, RAAN, argument of perigee and mean anomaly
Fig. 9. Feature importance in the Random Forest model. Inclination matters most.
Bar chart comparing Random Forest and Gradient Boosting, both with scores near 1.0
Fig. 10. Random Forest and Gradient Boosting score almost identically.
Table I. Performance comparison of the two models
ModelAccuracyPrecisionRecallF1-score
Random Forest0.990.990.990.99
Gradient Boosting0.990.990.990.99

Limits and next steps

The paper is open about one limit. The risk labels come from a simplified weighted score of orbital parameters, not from real collision events. That helps explain how clean the results are, so they show the approach is workable, not that it is ready for operations.

  • Use larger datasets.
  • Use real conjunction event records.
  • Try graph-based learning models.
  • Add time-dependent orbital prediction.

Presented at IEEE CHANDICON 2026

Certificate of Presentation from IEEE CHANDICON 2026, awarded to Neel Prasad for presenting the paper Machine Learning Prediction of Space Debris Collision Risk Using Orbital Data
Certificate of Presentation, IEEE CHANDICON 2026. Click to open full size.

← Back to all projects