The detection of moving sources in astronomical data is the backbone of many planetary astronomy projects. To date, this task relies on costly visual vetting to confirm moving sources amongst the much more numerous stationary sources. This is especially true of surveys which search stacks of sequential images that have been shifted at rates of motion relevant to the bodies of interest. When the shift rate matches that of a moving source, a point source is revealed. This process provides a search depth that is comparable to the point-source depth that would be had from a single sidereal stack. As sources are not visible in individual frames, the so-called shift’n’stack technique comes at the cost of not being able to link detections as a source moves through the frames. This results in maximal human search cost, even after applying modern processing techniques such as image subtraction to remove most of the stationary chaff. Here we present a new machine learning technique to identify high probability candidate moving sources, geared specifically for shift’n’stack data. We make use of a multi-layer resnet to perform the binary classification task: good or not good? Our network is trained on artificial sources that were injected into the data before image subtraction, themselves incorporating a rate of motion matching the objects of interest. We have applied this network in a search for Kuiper Belt Objects (KBOs) in data acquired with the Hyper Suprime-cam on the Subaru telescope, as part of a search for targets for NASA’s New Horizons Kuiper Extended Mission. The network’s classification performance is extremely good, resulting in a reduction of spurious candidate sources by more than three orders of magnitude. An entire night’s worth of search data requires roughly only one hour of human vetting. We find a detection efficiency >80% for r<25.5, with a limiting magnitude of r~26.5-26.8 (depending on image quality) despite the fact that these data were acquired at a galactic latitude of ~10 deg (see Figure 1). A handful of the >200 newly detected KBOs are bright enough to be imaged directly by the Long Range Reconnaissance Imager on board the spacecraft, and are scheduled to be observed during 2022-25. Our results show a promising new avenue for moving object detection that has the potential to greatly increase the depth of upcoming large surveys such as the Vera Rubin Observatory Legacy Survey of Space and Time.

Figure 1: Example detection efficiency of 2022 Subaru/HSC data searches. The pre-ML vetted shift’n’stack sources are shown in blue, and possess a ~300:1 false positive rate. The post-ML results classified by the resnet is shown in orange, and the final human confirmed sources in green. A near-zero false positive rate is found after human+resnet vetting, down to the limiting magnitude (r~26.5).