Accurately detecting rendezvous and proximity operations (RPO) is crucial for understanding how objects are behaving in the space domain. However, detecting closely-spaced objects (CSO) is challenging for ground-based optical space domain awareness (SDA) algorithms as two objects close together along the line-of-sight can appear blended as a single object within the point-spread function (PSF) of the optical system. Traditional machine learning methods can be useful for differentiating between singular objects and closely-spaced objects, but many methods require large training sample sizes or high signal-to-noise conditions. The quality and quantity of realistic data make probabilistic classification methods a superior approach, as they are better suited to handle these data inadequacies. We present CSO classification results using the Gaussian process python package, MuyGPyS, and examine classification accuracy as a function of angular separation and magnitude difference between the simulated satellites. This orbit-independent analysis is done on highly accurate simulated SDA images that emulate realistic ground-based commercial-of-the-shelf (COTS) optical sensor observations of CSOs. We find that MuyGPyS outperforms traditional machine learning methods, especially under more challenging circumstances.