Prithvi-EO-2.0：面向地球观测应用的多时序通用基础模型
计算机视觉与模式识别
2026-03-10 v3
摘要
本文介绍了Prithvi-EO-2.0，一种新的地空基础模型，相较于其前身Prithvi-EO-1.0在多个方面取得显著改进。模型在NASA的调和化合成和Sentinel-2数据档案中收集的420万全球时序样本（30米分辨率）上进行训练，新增时序嵌入和位置嵌入以提升在各种地空任务上的性能。通过与GEO-Bench进行大规模基准测试，模型在各项任务上的性能均优于前一代Prithvi-EO模型约8%。还与来自不同领域和分辨率（即从0.1米到15米）的六个其他地空基础模型进行基准测试。结果表明，该模型在经典地球观测和高分辨率应用中都表现出极佳的通用性。早期涉及终端用户和专题领域专家（SME）的持续反馈，使模型和数据集设计得以适应灾害响应、土地覆盖和作物映射以及生态系统动态监测等多样化的SME主导的应用。Prithvi-EO-2.0以开源形式在Hugging Face和IBM TerraTorch上提供，附加资源可在GitHub上获得。该项目体现了所有参与组织所采纳的可信开源科学方法。
引用
@article{arxiv.2412.02732,
title = {Prithvi-EO-2.0: A Versatile Multi-Temporal Foundation Model for Earth Observation Applications},
author = {Daniela Szwarcman and Sujit Roy and Paolo Fraccaro and Þorsteinn Elí Gíslason and Benedikt Blumenstiel and Rinki Ghosal and Pedro Henrique de Oliveira and Joao Lucas de Sousa Almeida and Rocco Sedona and Yanghui Kang and Srija Chakraborty and Sizhe Wang and Carlos Gomes and Ankur Kumar and Myscon Truong and Denys Godwin and Hyunho Lee and Chia-Yu Hsu and Rohit Lal and Ata Akbari Asanjan and Besart Mujeci and Disha Shidham and Trevor Keenan and Paulo Arevalo and Wenwen Li and Hamed Alemohammad and Pontus Olofsson and Christopher Hain and Robert Kennedy and Bianca Zadrozny and David Bell and Gabriele Cavallaro and Campbell Watson and Manil Maskey and Rahul Ramachandran and Juan Bernabe Moreno},
journal= {arXiv preprint arXiv:2412.02732},
year = {2026}
}