AlphaEarth Foundations:遥感基础大模型的潜力与挑战
作者贡献:Author Contributions
秦其明负责本文的全部学术贡献,包括述评观点与框架的提出、相关文献的调研与评述、核心论点的分析与论证、初稿的撰写以及终稿的审定。
QIN Qiming is solely responsible for all academic contributions to this article, including the proposal of perspectives and framework, the review and evaluation of relevant literature, the analysis and argumentation of the core viewpoints, the drafting of the initial manuscript, and the revision and approval of the final version.
|
秦其明(1955—),男,江苏徐州人,博士,教授,主要从事定量遥感与地理信息系统建模研究。E-mail: qmqin@pku.edu.cn |
收稿日期: 2025-09-02
修回日期: 2025-09-08
网络出版日期: 2025-09-25
基金资助
国家自然科学基金项目(42071314)
AlphaEarth Foundations: The Potential and Challenges of Remote Sensing Foundation Models
Received date: 2025-09-02
Revised date: 2025-09-08
Online published: 2025-09-25
Supported by
National Natural Science Foundation of China(42071314)
【目的】随着全球在轨地球观测卫星数量的快速增长,遥感数据呈现爆炸式积累,为地球系统科学研究提供了动态认知全球变化的前所未有机遇;与此同时,也伴生多源异构、标注稀缺、任务泛化不足与数据过载等一系列挑战。【方法】为应对这些瓶颈问题, Google DeepMind 提出了 AlphaEarth Foundations(AEF),通过整合光学、SAR、LiDAR、气候模拟及文本等多模态数据,构建统一的64 维嵌入表征场,实现了跨模态、跨时空的语义一致性的数据融合,并在 Google Earth Engine 等平台开放。【结果】AEF的主要贡献体现在: ① 缓解了长期存在的“数据孤岛”问题,建立了全球一致的嵌入层; ② 通过 vMF 球面嵌入机制提升了语义相似性度量能力,支持高效的检索与变化检测; ③ 将复杂的预处理与特征工程前置于预训练阶段,使下游应用进入“分析就绪”状态,大幅降低了应用成本。论文指出,AEF 的应用潜力释放可分为3个阶段:首先是地表覆盖分类与变化检测;其次是嵌入向量与物理模型深度耦合,推动科学发现;最后有望演化为空间智能基座,成为全球地理空间智能服务的一种基础设施。尽管如此, AEF仍面临若干挑战: ① 嵌入向量可解释性不足,限制了科学归因和因果分析; ② 域迁移与跨场景适应性存在不确定性,极端环境下的鲁棒性有待验证; ③ 性能优势需要更多跨区域、独立实验的实证支撑。【结论】AEF 以其在数据效率和跨任务泛化方面的突破,代表了遥感与地理空间人工智能研究的新方向,为未来地学研究提供了坚实支撑,但其进一步发展依据可解释性、鲁棒性及真实性验证的持续提升,并将64维嵌入向量通过不同途径转化为可广泛使用的数据资源。
关键词: AlphaEarth Foundations; 遥感基础大模型; 多模态融合; 嵌入表征; 时空建模
秦其明 . AlphaEarth Foundations:遥感基础大模型的潜力与挑战[J]. 地球信息科学学报, 2025 , 27(10) : 2283 -2290 . DOI: 10.12082/dqxxkx.2025.250426
[Objectives] With the rapid increase in the number of Earth observation satellites in orbit worldwide, remote sensing data has been accumulating explosively, offering unprecedented opportunities for Earth system science research to dynamically monitor global change. At the same time, it also brings a series of challenges, including multi-source heterogeneity, scarcity of labeled data, insufficient task generalization, and data overload. [Methods] To address these bottlenecks, Google DeepMind has proposed AlphaEarth Foundations (AEF), which integrates multimodal data such as optical imagery, SAR, LiDAR, climate simulations, and textual sources to construct a unified 64-dimensional embedding field. This framework achieves cross-modal and spatiotemporal semantic consistency for data fusion and has been made openly available on platforms such as Google Earth Engine. [Results] The main contributions of AEF can be summarized as follows: (1) Mitigating the long-standing “data silos” problem by establishing globally consistent embedding layers; (2) Enhancing semantic similarity measurement through a von Mises-Fisher (vMF) spherical embedding mechanism, thereby supporting efficient retrieval and change detection; (3) Shifting complex preprocessing and feature engineering tasks into the pre-training stage, enabling downstream applications to become “analysis-ready” and significantly reducing application costs. The paper further highlights the application potential of AEF in three stages: (1) Initially in land cover classification and change detection; (2) Subsequently in deep coupling of embedding vectors with physical models to drive scientific discovery; (3) Ultimately evolving into a spatial intelligence infrastructure, serving as a foundational service for global geospatial intelligence. Nevertheless, AEF still faces several challenges: (1) Limited interpretability of embedding vectors, which constrains scientific attribution and causal analysis; (2) Uncertainties in domain transfer and cross-scenario adaptability, with robustness in extreme environments yet to be verified; (3) Performance advantages that require more empirical validation across regions and independent experiments. [Conclusions] Overall, AEF represents a new direction for research in remote sensing and geospatial artificial intelligence, with breakthroughs in data efficiency and cross-task generalization providing solid support for future Earth science studies. However, its further development will depend on continuous advances in interpretability, robustness, and empirical validation, as well as on transforming the 64-dimensional embedding vectors into widely usable data resources through different pathways.
表1 3个遥感基础大模型技术特色的比较Tab.1 Comparison of the technical features of three remote sensing foundation models |
| 模型名称 | 预训练数据规模与来源 | 技术架构 | 主要创新点 |
|---|---|---|---|
| Prithvi-EO-1.0 (IBM/NASA) [4] | 数百万张影像;基于HLS(Sentinel-2+Landsat,多光谱,6 波段) | Vision Transformer(ViT)+Masked Autoencoder (MAE) | 时空patch化与联合建模;支持变化检测与分类 |
| DOFA (Dynamic One-For-All) [5] | 数千万级图像样本(任意输入通道);多源、多模态(光学、雷达等) | 超网络+动态权重生成器 | 动态适配不同传感器通道,解决跨模态/跨传感器不一致 |
| AEF (AlphaEarth Foundations, Google DeepMind) [2] | 30.5 亿帧影像;光学遥感影像、SAR影像、LiDAR、气候模拟、文本等多模态数据 | 多模态编码器+统一嵌入场,训练成本极高 | 构建统一、连续的嵌入场;支持连续时间建模; 64维向量表征 |
表2 遥感预处理过程常见的问题Tab. 2 Common issues in remote sensing preprocessing |
| 问题 | 原因 | 产生的效果 | 解决途径 |
|---|---|---|---|
| 传感器噪声与老化 | 在轨运行导致电子噪声、辐射漂移、机械磨损 | 地物辐射测量偏差、几何畸变(如条带噪声、亮度漂移) | 传感器在轨定标 + 场地定标(戈壁、盐湖等稳定目标) |
| 大气干扰(云、气体散射与吸收等) | 电磁辐射在大气层传输时发生遮挡、散射、吸收 | 光谱曲线偏移衰减,形成“灰霾效应”或“云污染”,降低反演精度 | 云检测与云去除;大气校正模型(6S、MODTRAN) |
| 地形效应 | 山区地形坡度、坡向导致辐射接收与散射不均;阴影区辐射不足 | 阴影区辐射减弱;坡向朝阳面光谱增强,导致严重光谱失真 | DEM 支持的地形辐射校正 |
| 混合像元与尺度效应 | 传感器分辨率限制,单个像元内包含多种地物类型 | 光谱信息混合,制约精细分类与参数反演 | 混合像元分解(线性/非线性光谱分解) |
| 临近效应(邻近散射) | 地物间多次散射、大气散射将周边信号叠加到目标像元 | 像元波谱受周围环境干扰,降低目标波谱纯度 | 大气校正 + 基于物理模型的校正(如 6S + DEM 模拟) |
注: AEF解决上述问题的具体方法见文献[2]。 |
表3 面向不同任务的AEF嵌入向量的应用方法Tab. 3 Application methods of AEF embedding vectors for different tasks |
| 任务类型 | 适用方法 | 应用举例 | 潜在局限 |
|---|---|---|---|
| 地物分类 | 64 维嵌入作为输入,训练轻量级分类器(随机森林、SVM、梯度提升机等) | 土地覆盖/利用分类、作物类型识别、局地气候区划分 | 可解释性差; 偏差被下游模型放大 |
| 回归分析 | 构建回归器(线性回归、GBDT、浅层神经网络)预测连续变量 | 生物量估算、蒸散发、地表温度、碳通量估算 | 误差归因困难; 物理规律一致性弱 |
| 变化检测 | 计算时序嵌入的余弦/欧氏距离,或训练孪生网络 | 城市扩张、森林砍伐、灾害评估 | 语义变化难以直接解释驱动力; 对缓慢演化现象(如植被退化)不敏感 |
| 异常检测 | 利用隔离森林、One-Class SVM 等算法发现异常点 | 非法采矿监测、突发灾害、植被病虫害 | 嵌入空间“异常”不一定对应真实地物物理特性异常; 易误判稀有、但正常分布地物 |
利益冲突:Conflicts of Interest 本文作者声明不存在利益冲突。
All authors disclose no relevant conflicts of interest.
注:本文根据2025年8月24日第三届北京交叉科学大会“气候变化与遥感交叉应用”前沿交叉论坛发言内容修改与补充。
| [1] |
付琨, 卢宛萱, 刘小煜, 等. 遥感基础模型发展综述与未来设想[J]. 遥感学报, 2024, 28(7):1667-1680.
[
|
| [2] |
Brown,
|
| [3] |
EDITORIAL. Towards responsible geospatial foundation models[J]. Nature Machine Intelligence, 2025,7:1189. DOI:10.1038/s42256-025-01106-7
|
| [4] |
|
| [5] |
|
| [6] |
|
| [7] |
|
| [8] |
The AlphaEarth Foundations team. AlphaEarth Foundations helps map our planet in unprecedented detail[EB/OL]. Google DeepMind, 2025-07-30. https://deepmind.google/discover/blog/alphaearth-foundations-helps-map-our-planet-in-unprecedented-detail/
|
| [9] |
|
/
| 〈 |
|
〉 |