2026-10-14T21:06:24Z https://www.scidb.cn/oai

10.57760/sciencedb.33946 2026-06-25T11:05:09Z

2026-06-25 LuoJia-FG doi:10.57760/sciencedb.33946 en LuoJia-FG is a large-scale, attribute-driven, and hierarchically structured multimodal land-cover dataset designed to bridge the semantic gap in high-resolution Earth observation. It is tailored for fine-grained land-cover interpretation and next-generation vision-language models by integrating pixel-level visual information with structured semantic descriptions.The dataset contains 119,619 image-mask-text triplets derived from Gaofen-2, Ziyuan-3, and aerial imagery, with spatial resolutions ranging from 0.5 m to 2.0 m. The samples cover approximately 1.59 million km^2 across six representative provinces and autonomous regions in China, including Anhui, Shaanxi, Hainan, Xinjiang, Gansu, and Qinghai, capturing diverse biogeographical landscapes and heterogeneous land-cover patterns.LuoJia-FG provides a three-level hierarchical taxonomy consisting of 8 primary classes, 52 secondary classes, and 95 tertiary fine-grained classes. Beyond conventional image-mask annotations, it introduces an attribute-coding system that generates attribute-driven textual descriptions based on physical properties such as vegetation structure, canopy density, phenology, land-use function, and spatial context. This design enables the dataset to support not only semantic segmentation but also hierarchical classification, open-vocabulary semantic segmentation, visual question answering, and text-guided geospatial understanding.To improve representativeness, LuoJia-FG adopts a knowledge-guided sampling strategy based on landscape heterogeneity, prioritizing complex and fragmented regions with rich semantic boundaries. Together with the CLIP-guided Hierarchical Classification Network benchmark, the dataset provides a challenging and extensible testbed for developing attribute-aware remote sensing vision-language models and advancing fine-grained Earth observation. Vision–Language Dataset; Fine-grained Land Cover Classification; Attribute-Driven; Hierarchical Taxonomy Yibing Xiong Yibo Wang Xiangyun Hu PUBLIC https://api.github.com/licenses/cc0-1.0 dataset http://www.doi.org/10.1080/20964471.2026.2688652 Science Data Bank