OpenEarthSensing: Large-Scale Fine-Grained Open-World Remote Sensing Benchmark
    [paper] | [dataset]
    Home page of the large-scale fine-grained open-world remote-sensing datasets and benchmark OpenEarthSensing (OES) for various open-world remote-sensing downstream tasks, mainly including evaluating the ability of models to detect semantic shifts, adapt to covariate shifts, and continuously update the parameters without forgetting learned knowledge. OES includes 189 scene and object categories, covering the vast majority of potential semantic shifts that may occur in the real world. To provide a more comprehensive testbed for evaluating the generalization performance, OES encompasses five data domains with significant covariate shifts, including two RGB satellite domains, one RGB aerial domain, one multi-spectral RGB domain, and one infrared domain.

    🎉 News

    [08/2025] 🔥 The OES dataset will be presented at the Low-Altitude Intelligent Perception Computing and Applications Forum of China Multimedia 2025

    [08/2025] 🔥 The OES dataset is available on kaggle.

    [07/2025] 🔥 The Second version of the OES paper is available on arXiv

    [02/2025] 🔥 The first version of the OES paper is available on arXiv.

    ✨ Highlights

    • Multiple and diverse domains: OES comprises five sub-datasets with five distinct domains, enabling it to serve as a testbed for various generalization tasks. We randomly select 2,000 images from each domain and utilize GeoRSCLIP to extract features. The t-SNE visualization is presented with each color representing a different domain. Notably, even though sub-dataset 1 and sub-dataset 2 both originate from satellite imagery, there is a significant domain shift due to the varying capturing conditions. Furthermore, satellite, aerial, and infrared images display considerable differences as well. These domain shifts highlight the significant evaluation values and challenges posed by OES.
    • Wide span of scales: To accommodate the scale variations present in remote sensing images, OES has been curated to include a diverse range of data comprising 152 scenes and 37 objects for classification. To delve deeper into the intricacies of scale diversity within the OES dataset, Qwen-VL-chat is employed to evaluate the image scales associated with both scene and object categories. The distribution of OES across different scales is visually represented. The extensive spectrum of scale variations within OES introduces a novel challenge to the realm of remote sensing recognition.
    • Multiple coarse categories: OES comprises 10 coarse-grained categories including Vegetation, Agriculture, Aviation, Waterbody & Facilities, Resource Acquisition & Utilization, Land Transportation, Nature & Climate, Infrastructure, Industrial Facilities and Residential Building, which effectively cover the majority of scenarios encountered in remote sensing applications. Each coarse-grained category is further divided into 10 to 27 fine-grained subcategories, culminating in a total of 189 distinct classifications.

    👀 Overview of All Sub-Datasets

      📖 Citation

      If you use OpenEarthSensing in your research, please cite our work:

      @article{xiang2025openearthsensing,
  title={OpenEarthSensing: Large-Scale Fine-Grained Open-World Remote Sensing Benchmark},
  author={Xiang, Xiang and Xu, Zhuo and Deng, Yao and Zhou, Qinhao and Liang, Yifan and Chen, Ke and Zheng, Qingfang and Wang, Yaowei and Chen, Xilin and Gao, Wen},
  journal={arXiv preprint arXiv:2502.20668},
  year={2025}
}