FM-ChangeNet++: Learned Nonlinear Feature Transport for Change Detection

Reliable change detection from repeated Earth observations is important for climate-relevant monitoring tasks including land-use analysis, disaster assessment, ecosystem monitoring, and urban development. However, most change detection (CD) methods compare bi-temporal representations only at their endpoints, while recent flow-matching approaches model feature transport using a fixed linear path. We introduce FM-ChangeNet++, which learns an endpoint-preserving nonlinear feature trajectory through a lightweight residual transport formulation. The residual bends intermediate feature states while preserving the observed pre- and post-temporal endpoints. Because the resulting path is nonlinear, the velocity field is supervised by its instantaneous derivative, providing localized evidence of feature evolution for change prediction. FM-ChangeNet++ combines multi-scale transport estimation, segmentation supervision, and trajectory regularization in a unified architecture. Experiments show consistent improvements over linear flow matching and representative CD baselines.