ReA-OVCD: Reliability-Aware Open-Vocabulary Change Detection via Semantic and Spatial Refinement

Traditional remote sensing change detection often relies on predefined categories, limiting its utility in dynamic, real-world scenarios. Open-Vocabulary Change Detection (OVCD) aims to identify land cover changes using arbitrary text prompts, but existing methods struggle with a trade-off: instance-level comparisons miss fine-grained semantic details, while pixel-level comparisons are prone to instability, false positives, and boundary artifacts due to semantic ambiguity.

The authors propose ReA-OVCD, a training-free framework that utilizes a frozen SAM-3 model to parse bi-temporal images. The framework employs two primary refinement modules to ensure reliability:

1. Semantic Change Reasoning (SCR): Instead of relying on discrete label flips, this module analyzes the continuous semantic distribution and response intensity. It identifies true changes as structured shifts in class probabilities, effectively suppressing incidental noise.
2. Boundary-aware Change Refinement (BCR): This module addresses spatial artifacts near object edges. It evaluates candidate change regions by checking for support from reliable interior pixels, filtering out spurious responses caused by geometric misalignment or segmentation uncertainty.

ReA-OVCD consistently outperforms state-of-the-art OVCD approaches across multiple standard remote sensing datasets (LEVIR-CD, WHU-CD, DSIFN, and SECOND). The method achieves F1-score improvements ranging from 2.13% to 9.75%. By integrating instance-level precision with semantic-level global consistency, the framework maintains high computational efficiency while significantly reducing false positives and boundary noise.

This research provides a robust, training-free solution for identifying land cover changes without the need for task-specific fine-tuning. By explicitly modeling the reliability of semantic and spatial predictions, the framework bridges the gap between the flexibility of open-vocabulary models and the high-precision requirements of remote sensing applications.