REOBench
收藏资源简介:
REOBench是一个用于评估地球观测基础模型鲁棒性的综合基准,涵盖了基于掩码图像建模、对比学习和视觉语言预训练范式训练的广泛模型。数据集专注于高分辨率光学遥感图像,广泛应用于城市规划、灾害响应等关键领域。该基准包括12种图像退化类型,包括基于外观和几何的扰动。数据集的构建过程包括系统地评估了众多模型,以揭示当前地球观测基础模型在实际世界扰动下的脆弱性,并为开发更鲁棒和可靠的模型提供了可操作的见解。
REOBench is a comprehensive benchmark for evaluating the robustness of Earth Observation (EO) foundation models, covering a wide range of models trained under paradigms including masked image modeling, contrastive learning, and vision-language pre-training. This benchmark focuses on high-resolution optical remote sensing images, which are widely applied in critical domains such as urban planning and disaster response. It includes 12 types of image degradations, covering both appearance-based and geometry-based perturbations. The construction of this benchmark involves systematically evaluating numerous models, aiming to reveal the vulnerability of current Earth Observation foundation models under real-world perturbations and provide actionable insights for developing more robust and reliable models.
REOBench数据集概述
数据集简介
- 名称:REOBench
- 全称:Benchmarking Robustness of Earth Observation Foundation Models
- 目的:评估地球观测基础模型的鲁棒性
- 数据内容:高分辨率光学遥感图像
- 应用场景:城市规划、灾害响应等
关键特性
- 覆盖任务:6种广泛研究的遥感图像理解任务
- 视觉中心任务
- 视觉语言任务
- 扰动类型:12种
- 外观损坏(如噪声、模糊、雾霾)
- 几何畸变(如旋转、缩放、平移)
- 严重程度:不同级别
数据集获取
- 下载地址:https://huggingface.co/datasets/xiang709/REOBench
- 加载方式: python from datasets import load_dataset fw = load_dataset("xiang709/REOBench", streaming=True)
实验支持
- 语义分割:使用mmsegmentation
- 目标检测:使用mmrorate
- 分类:提供相关代码
- 视觉语言任务:提供评估代码(Caption, VQA, Visual Grounding)
许可信息
- 许可证:CC-BY-4.0
- 许可链接:https://creativecommons.org/licenses/by/4.0/deed.en
相关项目
引用信息
bibtex @misc{li2025reobenchbenchmarkingrobustnessearth, title={REOBench: Benchmarking Robustness of Earth Observation Foundation Models}, author={Xiang Li and Yong Tao and Siyuan Zhang and Siwei Liu and Zhitong Xiong and Chunbo Luo and Lu Liu and Mykola Pechenizkiy and Xiao Xiang Zhu and Tianjin Huang}, year={2025}, eprint={2505.16793}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2505.16793}, }
致谢
- 基于以下数据集构建:
- 使用工具:




