Schematic overview of our SSIM-based deep learning pipeline for satellite-trail detection. Each group consists of 30 time-consecutive images of the same field at a fixed wavelength (original data). For training, synthetic linear trails are injected into trail-free images to construct a balanced positive–negative sample (synthetic data). For each target image, we compute SSIM maps between this image and the remaining images in the group, which helps suppress the static astronomical background and enhance transient linear features. The original image and its SSIM maps are then fed into a ResNet-34 classifier, which outputs the probability (﹩{\hat{p}}_{\mathrm{trail}}﹩) that a satellite trail is present in the target image.