OrbitAware: AI Asteroid Detector

BY: Shishir Bahubali, Rishi Kodungallur, Revanth Guda, and Sohil Singh

01

03

02

04

TABLE OF CONTENTS

The Problem

Demonstration

Our Solution

Scalability

The Problem

01

1-2 Kilometers

The size of an asteroid that could cause worldwide damage (NASA 2014)

Difficulty in Detection

• Stray asteroids floating in space can be difficult to detect, especially if they are smaller in size and are approaching at different angles

​

​

​

Collision Risk

Satellites traveling on long distance journeys through space could be at risk of asteroids in asteroid belts “as big as 940 kilometers (about 583 miles) across” (NASA 2014).

The Solution

02

AN AI MODEL TO CLASSIFY THE DANGER LEVEL OF ASTEROIDS TO HELP SPACE EXPLORATION

Technical Details

• Using data sets from Nasa’s Jet Propulsion Laboratory Database on existing asteroids, we set up python code incorporating AI to be trained on recognizing the danger status through dimensions of certain asteroids

AI Models

Logistic Regression

A supervised machine learning algorithm that uses binary classification to predict the probability of an outcome

Random Forest Classifier

Machine learning algorithm that combines the output of multiple decision trees to reach a single result

LightGBM

an ensemble learning framework, specifically a gradient boosting method, which constructs a strong learner by sequentially adding weak learners in a gradient descent manner

Metrics

• To choose between the 3 different machine learning algorithms we looked at multiple different metrics
• Accuracy:
• a metric that measures how often a machine learning model correctly predicts the outcome
• Calculated by: correct predictions divided by all predictions
• Precision:
• the quality of a positive prediction made by the model
• A positive is when the correct prediction is “Yes”
• Calculated by: true positives divided by total number of positive predictions
• Recall:
• a metric that measures how often a machine learning model correctly identifies positive instances (true positives) from all the actual positive samples in the dataset
• F1 Score
• the harmonic mean of the precision and recall of a model
• By combining these two metrics it offers a more correct reliability of the model

Logistic Regression

• A logistic regression predicts the probability of an outcome to binary classify data
• This regression is very simple but effective
• However, it sacrifices accuracy because of the fact that it is less complex than other algorithms
• Our logistic regression was our second best algorithm after training

Random Forest Classifier

• This algorithm uses decision trees to classify data inputs
• It combines the outputs from different trees to reach a decision
• One unique thing about this classifier is that it can handle both regression and classification problems
• However because this data is numerical it would need a regression algorithm to solve and since Random Forest Classification does both it sacrifices some of the metrics to do this
• This caused it to be our worst algorithms in terms of metrics

LightGBM

• Light GBM is a fast, distributed, high-performance gradient boosting framework based on decision tree algorithm
• Because of this LightGBM requires intense training time to acquire high level results
• This made training our model with this algorithm take a long time but lead to great results
• This was our best algorithm that we tested and is what will be used with our model that space agencies can use

Demonstration

03

04

Scalability

As an open source code, OrbitAware software can be sent to different space agencies, who can modify and use the code to make it their own.

​

​

• LightGBM can be further trained to recognize specific asteroid patterns, connecting to spacecraft navigation systems to reroute paths through large asteroid belts