Answers are grounded in this paper and cite their sources.
Local displacement, illumination changes, and repeated roof textures can produce false building changes in bi-temporal remote-sensing images. We propose PINF-RCAF to detect change from the discrepancy that remains after local transport and nuisance regression. The Pseudo-Invariant Nuisance Field (PINF) estimates displacement-penalized bidirectional transport, and Pseudo-Invariant Residual Regression (PIRR) removes the locally explained component. Reciprocal Correlation Alignment Fusion (RCAF) combines local correlation, forward–backward cycle agreement, and learned trust to control semantic and boundary corrections. A direct residual path to the decoder preserves evidence for structures without a reliable temporal counterpart. At a fixed threshold of 0.50, three independently seeded training runs yield mean test F1/IoU of 0.9305/0.8701 on FOTBCD, 0.9448/0.8954 on LEVIR-CD, and 0.9722/0.9460 on DSIFN-CD. The mean F1 gains over the matched Control are 3.49, 2.58, and 1.26 percentage points, respectively. Component comparisons identify the gains from local matching and residual correction. Boundary evaluation and controlled validation-set appearance perturbations further show improved contour placement and fewer background responses. These results support the use of correspondence reliability to refine residual change evidence.
No takes yet. Share an insight, caveat, or question.
Ning et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: