IRSAMap is the world's first remote sensing dataset specifically designed for large-scale, high-resolution, and multi-category land cover vector mapping. It aims to support research in geospatial artificial intelligence, semantic segmentation, vector map extraction, and automated cartography from remote sensing imagery.
IRSAMap bridges the gap between raster-based land cover data and vector-based mapping needs by providing precisely labeled, vectorized land cover features across diverse geographical regions. Each instance is annotated with fine-grained geometry and semantic class information.
[2025/09/21] The GeoJSON for land categories in IRSAMap has been updated!!!
- Global Coverage: Diverse regions from multiple continents.
- High Resolution: Vector annotations based on high-resolution satellite imagery.
- Multiple Classes: Rich semantic categories including urban areas, water bodies, forests, farmland, roads, buildings, and more.
- Vector Format: Provided as GIS-ready shapefiles (Shapefile/GeoJSON), enabling easy integration into spatial analysis pipelines.
Overview of the IRSAMap dataset. The top left shows the dataset's annotation process. The central section displays the annotated classes of the IRSAMap dataset. The regional distribution map (top right) highlights the global coverage of IRSAMap imagery. The bottom left compares traditional raster datasets, which only support semantic segmentation tasks, with IRSAMap, which supports multi-class tasks.
If you use the IRSAMap dataset in your work, please cite the following article:
@ARTICLE{11129926,
author={Meng, Yu and Deng, Ligao and Xi, Zhihao and Chen, Jiansheng and Chen, Jingbo and Yue, Anzhi and Liu, Diyou and Li, Kai and Wang, Chenhao and Li, Kaiyu and Deng, Yupeng and Sun, Xian},
journal={IEEE Transactions on Geoscience and Remote Sensing},
title={IRSAMap: Toward Large-Scale, High-Resolution Land Cover Map Vectorization},
year={2025},
volume={63},
number={},
pages={1-19},
keywords={Land surface;Annotations;Roads;Feature extraction;Vectors;Buildings;Remote sensing;Training;Accuracy;Semantic segmentation;Deep learning;high-resolution remote sensing;land cover;object-based modeling;vector mapping},
doi={10.1109/TGRS.2025.3600249}}
You can download the dataset via the following links: