STAND: Semantic Anchoring Constraint with Dual-Granularity Disambiguation for Remote Sensing Image Change Captioning

Technical Summary: STAND (Semantic Anchoring Constraint with Dual-Granularity Disambiguation)

1. Problem Statement

Remote Sensing Image Change Captioning (RSICC) involves describing the differences between two bi-temporal images. The authors identify that RSICC is significantly more challenging than natural image change captioning due to three inherent ambiguities:

• Viewpoint Ambiguity: Top-down perspectives make distinct objects (e.g., a road vs. a roof) look visually similar.
• Scale Ambiguity: Changed objects often occupy a very small portion of the total pixels, making them difficult for standard encoders to capture.
• Knowledge Ambiguity: Precise entity-level identification (e.g., distinguishing between a parking lot, a building, or a container) requires specific domain-specific prior knowledge.

Existing methods often rely on static image encoders or simple mask-filtering, which fail to capture long-range context or inadvertently destroy crucial spatial information.

2. Methodology

The proposed STAND framework adopts a progressive pipeline to resolve these ambiguities through a video-modeling paradigm. It treats the "before," "change mask," and "after" images as a short video clip. The architecture consists of four main components:

• Interpretable Transition Constraint (ITC): To ensure the temporal features are reliable, the ITC module regularizes the transition between the "before" state, the "change" representation, and the "after" state. It uses a bidirectional InfoNCE loss to align synthesized "pseudo-after" features with the actual "after" features, ensuring the model captures genuine temporal dynamics.
• Dual-Granularity Target Disambiguation (DGTD): This module addresses spatial uncertainties through two scales:
  • Micro-level (FRCA): The Frequency-Refocused Complementary Attention uses Discrete Cosine Transform (DCT) to filter out low-frequency noise and refocus attention on high-frequency details, enhancing the response to small-scale objects.
  • Macro-level (CAVD): The Context-Aware Viewpoint Disambiguation incorporates global scene semantics to provide context, helping the model distinguish between visually similar entities (e.g., a container vs. a roof) based on their surroundings.
• Semantic Concept Anchoring (SCA): To tackle knowledge ambiguity, this module extracts category labels from the training set (using a dual-agent LLM verification process). It uses learnable queries to capture object-specific representations and weights them by their probability of change, grounding the visual features in structured linguistic priors.
• Prior Knowledge Guided Difference Decoder: A decoder that progressively fuses the disambiguated difference features with the object-aware semantic representations to generate precise text.

3. Key Contributions

• A Novel Disambiguation Framework: Introduced a systematic approach to handle viewpoint, scale, and knowledge ambiguities in RSICC.
• Frequency-Domain Refinement: Proposed the FRCA module to leverage frequency-domain filtering for better small-object sensitivity.
• Semantic Grounding: Developed the SCA module to bridge the gap between visual changes and domain-specific linguistic categories.
• Temporal Regularization: Introduced the ITC to ensure the video-based encoding captures meaningful transitions rather than just pixel differences.

4. Results

• State-of-the-Art Performance: STAND outperformed existing methods on both the LEVIR-CC and WHU-CDC datasets across all major metrics (BLEU, METEOR, ROUGE-L, and CIDEr).
• Ablation Success: Ablation studies confirmed that the modules are complementary; for instance, combining both macro and micro-level disambiguation (CAVD + FRCA) yielded significantly better results than using either alone.
• Ambiguity Mitigation: Targeted evaluations proved that STAND specifically improves performance in small-scale detection, reduces false positives in viewpoint confusion, and increases entity-level recall for knowledge-heavy tasks.
• Superiority over MLLMs: The authors demonstrated that even fine-tuned general Multimodal Large Language Models (like LLaVA) underperform compared to STAND, highlighting the necessity of task-specific, domain-aware architectures for remote sensing.

5. Significance

This work shifts the focus of RSICC from simple "difference detection" to "semantic disambiguation." By integrating frequency-domain processing, global context, and linguistic priors, STAND provides a robust blueprint for handling the unique geometric and semantic complexities of satellite and aerial imagery. It demonstrates that for specialized domains like remote sensing, compact, well-designed architectures with domain-specific constraints can outperform massive, general-purpose foundation models.