Chengdu UHR dataset
收藏资源简介:
成都UHR数据集是由成都理工大学与中国科学院联合构建的超高分辨率城市绿地提取专用数据集。该数据集包含7700个经人工精细标注的512×512图像块,数据源自2020年获取的77幅Google Earth RGB镶嵌影像,空间分辨率达0.27米,并配合同期Sentinel-2影像衍生的10米分辨率NDVI层。数据集通过严格的分割前裁剪协议生成,确保训练与验证样本间无空间重叠泄露,标注类别融合树木、草坪等植被要素为前景绿地。该数据集专为城市绿地精细分割任务设计,旨在解决超高分辨率影像中植被碎片化边界、阴影干扰及类内差异大等挑战,推动遥感影像智能解译算法的发展。
The Chengdu UHR Dataset is a specialized dataset dedicated to ultra-high resolution urban green space extraction, jointly developed by Chengdu University of Technology and the Chinese Academy of Sciences. It contains 7700 manually meticulously annotated 512×512 image patches, which are derived from 77 Google Earth RGB mosaic images acquired in 2020 with a spatial resolution of 0.27 meters, and is paired with a 10-meter resolution NDVI layer generated from concurrent Sentinel-2 images. The dataset is produced via a strict pre-segmentation cropping protocol to prevent spatial overlap leakage between training and validation samples. Its annotation categories merge vegetation elements including trees and lawns into foreground green space. Specifically designed for the fine-grained urban green space segmentation task, this dataset aims to address challenges such as fragmented vegetation boundaries, shadow interference and large intra-class differences in ultra-high resolution remote sensing images, and promote the development of intelligent remote sensing image interpretation algorithms.
数据集详情总结
- 数据集名称/来源:论文《GMBFormer: An NDVI-Guided Global Memory Bank Transformer for Urban Green-Space Extraction from Ultra-High-Resolution Imagery》中使用的两个数据集:
- 自建成都超高分辨率(UHR)数据集:包含 7,700 张标记好的 512 x 512 像素的图像块。
- 公共ISPRS Potsdam数据集:使用了该数据集的两个减少标签设置版本。
- 数据集任务:用于从超高分辨率遥感影像中提取城市绿地。
- 论文方法:提出 GMBFormer,一个基于 SegFormer 的框架,将 NDVI 解耦为物理信息门控,辅助 RGB 主干网络学习,并通过全局记忆库(global memory bank)进行基于相似性的原型检索,以实现语义复用。
- 性能结果(在相同训练和评估协议下):
- 成都UHR数据集:mIoU 89.25%,mDice 94.31%
- ISPRS Potsdam(减少标签设置1):mIoU 92.17%,mDice 95.92%
- ISPRS Potsdam(减少标签设置2):mIoU 83.72%,mDice 90.86%
- 相关实验:消融研究表明,解耦的NDVI准入、记忆检索、容量和动量共同决定了最终性能。

- 1GMBFormer: An NDVI-Guided Global Memory Bank Transformer for Urban Green-Space Extraction from Ultra-High-Resolution Imagery成都理工大学·地球物理学院; 中国科学院·空天信息创新研究院国家工程研究中心; 中国科学院大学 · 2026年



