FVT: Feature-to-Video Transition Modeling for Semantic Change Detection

Abstract

Semantic change detection aims to identify changed regions and recognize their semantic categories from remote sensing images acquired at two different times. One of the key challenges is to model cross-temporal change relationships effectively. However, most of the existing methods struggle to maintain stable semantic representations in complex cross-temporal scenarios, and the change processes are often treated as discrete. This leads to inconsistent semantic understanding across time and makes it difficult to capture the progressive nature of real-world changes. To further overcome the limitations of existing input-level dynamic change modeling, we propose FVT, a semantic change detection network integrating self-supervised priors with the pair-to-video (P2V) paradigm, explicitly modeling change transitions at the feature level. Specifically, we introduce a P2V-based feature-level temporal evolution strategy transforming bitemporal features into continuous video-like feature sequences. Furthermore, a bidirectional video feature module is designed to characterize directional “from–to” semantic transitions, together with a difference-frame enhancement mechanism to emphasize fine-grained interframe changes. Extensive experiments on the SECOND and Landsat-SCD datasets demonstrate that FVT achieves state-of-the-art performance compared with 12 representative methods.