Machine learning-based risk classification of space debris conjunction events

In the case of the growing, alarming threats posed by space debris in the low Earth orbit, conventional approaches of examining the probability of space crashes at the Time of Closest Approach (TCA) are not performing up to the mark. This paper presents a machine learning pipeline to deal with the dreadful imbalance of classes present in conjunction assessment. Our hybrid approach is a Random Forest classification using SMOTE oversampling and F2-optimal threshold optimisation on 14, 657 Conjunction Data Messages. The baseline model was accurate in 99.44% and had zero recall on the risky events. Our hybrid scheme has a perfect recall (1.00) with the risk of collision risks, but at the cost of zero false negatives at the expense of higher false positives (above average trade-off in the context of safety-critical). The system will be implemented within a multi-tier operational architecture and consumes 95 per cent less computation, but still has full capability with regard to risk detection.