olmoearth-paper-embeddings
收藏资源简介：
OlmoEarth论文表2基础模型嵌入数据集是一个专门为地球观测领域设计的大规模预提取嵌入集合，旨在支持下游任务（如KNN分类、线性探针和微调）的性能评估与比较，避免研究人员重复运行模型编码器进行特征提取。数据集包含26个不同的地球观测基础模型在24个下游任务上生成的嵌入向量，这些任务与论文《OlmoEarth: Stable Latent Image Modeling for Multimodal Earth Observation》中的表2完全对应。数据规模约1.1 TB，以PyTorch的.pt文件格式存储，每个文件包含embeddings（bfloat16张量）和labels（int64张量）键。模型涵盖AnySat、Clay、CopernicusFM、CROMA、DINOv3系列、Galileo系列、Panopticon、Presto、Prithvi v2系列、Satlas、TerraMind、TESSERA以及OlmoEarth自身的变体。下游任务涵盖多种地球观测模态和应用，包括Sentinel-1、Sentinel-2、Landsat-8以及多模态时间序列数据，任务类型包括分类（如m_bigearthnet、m_so2sat等）和分割（如m_cashew_plant、PASTIS等），评估方法包括K最近邻和线性探针。数据集还提供评估设置文件和已知注意事项文档，采用CC BY 4.0许可证。
The Base Model Embedding Dataset from Table 2 of the OlmoEarth paper is a large-scale pre-extracted embedding collection specially designed for the Earth observation domain. It aims to support performance evaluation and comparison of downstream tasks such as KNN classification, linear probing, and fine-tuning, eliminating the need for researchers to repeatedly run model encoders for feature extraction. The dataset contains embeddings generated by 26 distinct Earth observation base models across 24 downstream tasks, which fully correspond to Table 2 in the paper *OlmoEarth: Stable Latent Image Modeling for Multimodal Earth Observation*. With a total size of approximately 1.1 TB, the dataset is stored in PyTorch .pt file format, where each file contains two keys: embeddings (bfloat16 tensors) and labels (int64 tensors). Covered models include AnySat, Clay, CopernicusFM, CROMA, DINOv3 series, Galileo series, Panopticon, Presto, Prithvi v2 series, Satlas, TerraMind, TESSERA, and variants of OlmoEarth itself. The downstream tasks cover diverse Earth observation modalities and applications, including Sentinel-1, Sentinel-2, Landsat-8, and multimodal time-series data. Task types range from classification (e.g., m_bigearthnet, m_so2sat, etc.) to segmentation (e.g., m_cashew_plant, PASTIS, etc.), with evaluation methods including K-nearest neighbor (KNN) and linear probing. Additionally, the dataset provides evaluation setup files and a known issues documentation, and is licensed under CC BY 4.0.
