ChrisRPL/satellite-disruption-triage-aux-v1-1

卫星中断分类v1.1（辅助）数据集是一个用于从卫星图像中进行宏观尺度民用中断分类的视觉语言模型（VLM）研究的公共辅助资源。它明确不是一个标准基准，也不是专家标记的核心事实，更不是生产卫星分类系统的直接替代品。数据集包含322个成对的卫星图像（基线和当前）示例，来自17个全球民用灾害事件，并带有结构化的JSON分类输出。数据集分为训练集（231个示例）和评估集（91个示例），采用事件保留策略确保评估集的事件家族不出现在训练集中。数据集支持两种模态：光学到SAR（跨模态）和光学到光学（同模态）。每个示例包含一个基线图像、当前图像和结构化的JSON输出，包括动作、类别、理由和边界框等信息。数据集的主要局限性包括算法生成的标签、类别不平衡、模态差距以及许可证限制等。The Satellite Disruption Triage v1.1 (Auxiliary) dataset is a public auxiliary resource for vision-language model (VLM) research on macro-scale civilian disruption triage from satellite imagery. It is explicitly not a canonical benchmark, not expert-labeled core truth, and not a drop-in substitute for a production satellite triage system. The dataset contains 322 examples of paired satellite images (baseline + current) from 17 global civilian disaster events, annotated with structured JSON triage outputs. It is split into train (231 examples) and eval (91 examples) sets with an event-held-out policy ensuring no source-event family appears in both splits. The dataset supports two modalities: optical-to-SAR (cross-modality) and optical-to-optical (same-modality). Each example includes a baseline image, current image, and structured JSON output with fields like action, category, rationale, and bounding box. Key limitations include algorithmic labels with template rationales, class imbalance, modality gaps, and license restrictions.