Research map / Vision-language and multimodal / Human and embodied modeling

RSCC: A Large-Scale Remote Sensing Change Caption Dataset for Disaster Events

NeurIPS 2025 · Vision-language and multimodal › Human and embodied modeling · approach: Medical vision-language models

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces RSCC, a 62k-pair pre/post-disaster remote sensing change caption dataset generated with QvQ-Max from xBD/EBD building-damage labels, and benchmarks MLLMs plus a fine-tuned Qwen2.5-VL-7B model.

Problem

Remote sensing datasets lack bi-temporal image pairs with detailed textual change descriptions for disaster events, limiting vision-language models for disaster-aware temporal understanding.

Approach

Generate detailed change captions for xBD and EBD bi-temporal disaster image pairs with a visual reasoning MLLM (QvQ-Max), using building damage masks as visual prompts and damage-level text as in-context information, followed by metadata-based post-correction and expert verification.

Key results

• Fine-tuned Ours (7B) on RSCC reaches 58.52% ST5-SCS and 14.99% ROUGE in the zero-shot setting, beating Qwen2-VL (7B) at 45.55% ST5-SCS and the other open baselines on the RSCC test subset.
• Adding building-damage textual prompts raises ST5-SCS substantially for most general models (e.g., Phi-4-Multimodal 34.55% to 67.62%), with performance saturating across model sizes.
• In human preference voting, QvQ-Max ground-truth captions win 80.7% (vs InternVL3) to 99.0% (vs CCExpert) of comparisons.

Measured effects reported in the paper

Surprising findings

• Training-free scaling correction decoding (VCD, DoLa, DeCo) gave no obvious boost on change captioning, with DoLa and VCD underperforming across sizes.
• Kimi-VL (3B) outperformed larger models such as Qwen2-VL (7B) on ST5-SCS, suggesting model size does not fully determine performance.

Limitations

• Captions may be vague or hard to verify, since the labels are imperfect and image pairs are complex.
• Only text similarity metrics are used because image-to-text captioning metrics are designed for single images and fail in the multi-image setting.
• BLEURT is strongly biased by text length, so it was not a reliable metric.

Experimental setup

Datasets: xBD EBD

Benchmarks: RSCC test subset

Metrics: ROUGE METEOR BLEURT ST5-SCS (Sentence T5-XXL sharpened cosine similarity)

Baselines: LLaVA-NeXT-Interleave BLIP-3 LLaVA-OneVision Qwen2-VL Pixtral Phi-4-Multimodal Kimi-VL InternVL 3 TEOChat CCExpert

Builds on: Qwen2.5-VL xBD LLaVA-NeXT-Interleave Shtedritski et al. marking-based visual prompting

Closest papers by meaning

• DisasterM3: A Remote Sensing Vision-Language Dataset for Disaster Damage Assessment and Response (NeurIPS 2025 · similarity 0.90)
• RSRCC: A Remote Sensing Regional Change Comprehension Benchmark Constructed via Retrieval-Augmented Best-of-𝑁 Ranking (NeurIPS 2026 · similarity 0.87)
• MONITRS: Multimodal Observations of Natural Incidents Through Remote Sensing (NeurIPS 2025 · similarity 0.87)
• Flood Insights: Integrating Remote and Social Sensing Data for Flood Exposure, Damage, and Urgent Needs Mapping (AAAI 2024 · similarity 0.85)
• SkyScript: A Large and Semantically Diverse Vision-Language Dataset for Remote Sensing (AAAI 2024 · similarity 0.85)
• The STVchrono Dataset: Towards Continuous Change Recognition in Time (CVPR 2024 · similarity 0.85)