Explicit Semantic Transition Modeling for Open-Vocabulary Remote Sensing Change Detection
Abstract
Open-vocabulary change detection (OVCD) aims to detect land-cover changes associated with semantic concepts, but existing methods usually infer change semantics from independently predicted endpoint labels. Such endpoint-centric formulations do not directly model semantic change as a directed from-to transition and may suffer from accumulated endpoint errors. In this article, we propose an explicit semantic transition modeling framework for open-vocabulary remote sensing change detection. The proposed method uses frozen DINOv2 and SegEarth-OV3 to extract bi-temporal visual features, semantic probability maps, and text prototypes. A lightweight proposal branch identifies ROIs that may contain changes, within which the model directly predicts pixel-level directed semantic transitions. To handle the large transition space, a compact Top- $k$ candidate set is first constructed from ROI semantic distributions, and a residual candidate supplement module (RCSM) further supplements potentially omitted transitions. Moreover, a consistency-aware semantic transition refinement module (CSTRM) calibrates transition logits through transition-marginal consistency (TMC) and improves spatial coherence through transition-aware refinement. Experiments on six remote sensing change detection datasets demonstrate that the proposed formulation consistently outperforms representative unsupervised, foundation-model-based, and open-vocabulary methods. The code will be available at https://github.com/liuyk97/OV-SCD