anonymoususer202021/HierGeo-Data

HierGeo-Data是一个大规模卫星图像-字幕数据集，包含464,553个图像-字幕对，覆盖140个国家和超过2,000个不同命名地点。该数据集是作为HierGeo地理条件检索框架的训练语料库引入的，并提交至NeurIPS 2025。它通过空间连接Core-Five亚米级（0.25–0.3米）卫星图像与Foursquare的1.14亿个已验证POI数据，使用InternVL3-38B模型生成基于POI元数据（如国家、城市、地点类别）的字幕，以编码地理背景、物理外观和对象级内容，并经过视觉质量过滤（如云覆盖、传感器伪影、空间熵）以保留信息丰富的图像块。该数据集是唯一提供具有地理基础、多方面字幕的亚米级遥感图像的大规模数据集，覆盖全球范围。HierGeo-Data is a large-scale satellite image–caption dataset of 464,553 image–caption pairs spanning 140 countries and more than 2,000 distinct named locations. It is the training corpus introduced alongside the HierGeo geographically conditioned retrieval framework, submitted to NeurIPS 2025. The dataset is created by spatially joining the Core-Five sub-meter (0.25–0.3 m) satellite archive with 114 million verified Foursquare POIs across 140 countries, generating captions using InternVL3-38B conditioned on verified POI metadata (country, city, place category) to encode geographic context, physical appearance, and object-level content, and filtering for visual quality (cloud cover, sensor artifacts, spatial entropy) to keep only informative patches. It is the only large-scale EO dataset providing sub-meter imagery with geographically grounded, multi-aspect captions across a truly global footprint.