be very well understood, it is preferable to flag and ignore affected pixels rather than trying to recover them. In those cases, it would of course be wasteful for TrailMask to perform all the computations required to obtain all outputs. It is therefore crucial that the user be able to decide which combination of the various available outputs (1–3) they desire.
Approach
Figure 3. Schematic of TrailMask, when running in deep learning mode. If input images are too different from data used to train the pretrained model, but the user does not provide their own training data, TrailMask can rely on ImageSimulate (section 3.3.2) to provide extra training data.
When several frames of the same part of the sky are available, image differencing can be used to identify trail locations, and a simple median-stacking can be sufficient to remove the trail. TrailMask must also handle the case where only a single image is given as input. A simple, algorithmic approach should be available, and be able to produce satisfactory results for outputs 1 and 2. This could be based on the Hough Transform. However, modified methods may be needed to handle curved trails, which are especially likely to occur in space-based observations. Lastly, a more advanced, deep-learning-based method can be used, allowing for output 3 to be produced (and likely improving the quality of outputs 1 and 2). See Fig. 2 and Fig. 3 for schematic descriptions of these approaches.
Other considerations. How will the cutoff transverse to the trail be set? How will the algorithms behave on curved trails? How will ghosting be handled? For non saturated trails, can we assess whether faint sources can be detected under the trail?
A separate program under the TrailMask area might be a spectroscopy analysis tool to detect spectra showing contamination by a satellite spectrum (which will be close in shape to the solar spectrum).
