Counterfactual False-Positive Mining for Appearance-Robust Change Detection

Appearance differences between acquisition dates can increase change responses at unchanged locations. Standard supervision on augmented pairs does not explicitly isolate this additional response. We define an input-level counterfactual response as the logit difference produced by the same fixed EMA teacher at the same pixel when the original time-2 image is replaced by a specified appearance intervention. This paired model response holds the scene geometry, label, teacher parameters, and operator realization fixed; it is not a causal reconstruction of the physical acquisition process. We propose counterfactual false-positive mining with reference prediction preservation (CFPM-RP). An exponential-moving-average (EMA) teacher compares an original pair with a pair undergoing a controlled appearance transformation. Positive logit increments within stable unchanged regions define candidate locations. A smooth one-sided loss penalizes the perturbed student response relative to the original teacher prediction, with auxiliary support weighting, segmentation supervision, and changed-region constraints. The auxiliary paths are used only during training, retaining the deployed detector architecture. Experiments cover LEVIR-CD, WHU-CD, and SYSU-CD, eight appearance perturbations, an equal-budget continuation, eight additional comparison methods, and a second backbone. CFPM-RP attains perturbation-averaged F1 values of 91.444%, 71.181%, and 77.842% on the three datasets, respectively. Relative to Generic-FT, false-positive counts decrease by 10.37%, 11.15%, and 11.27%. The results show that CFPM-RP reduces appearance-induced false positives while retaining competitive change-detection performance.