谷歌把整个地球装进大模型！实时观测，按天更新

Core Viewpoint - Google DeepMind has launched the AlphaEarth Foundations model (AEF), which enables high-precision mapping of the Earth, addressing challenges of data overload and information inconsistency [1][3][10]. Group 1: Model Functionality - AEF acts like a virtual satellite, integrating massive amounts of Earth observation data, including optical satellite images and climate simulations, to create customizable geographic representations [3][14]. - The model utilizes a spatiotemporal precision encoder (STP) to capture long-distance geographic relationships and temporal dynamics, allowing for detailed mapping even with sparse labeled data [16][18]. - AEF separates the observation data support period from the effective mapping period, enabling reliable predictions even when direct observation data is unavailable [19][21]. Group 2: Data Integration and Processing - AEF can process over ten types of input data, including optical images, radar, and climate data, breaking down barriers between different data types [23]. - The model generates compact embedding vectors that represent complex Earth surface information, with a storage requirement only one-sixteenth of other AI systems [26]. - AEF employs a teacher-student model and contrastive learning strategies, enhancing the semantic information of features through text alignment training [26]. Group 3: Performance and Applications - AEF consistently maintains high accuracy across various tasks, achieving a balanced accuracy of 0.82 in land cover classification, outperforming the next best model at 0.69 [30][31]. - The model excels in scenarios with scarce labeled data, showing a 24% lower average error rate compared to tested models [32]. - AEF is utilized in diverse applications, including land cover classification, biophysical variable estimation, and change detection, with precise differentiation of land types [34][35]. Group 4: Data Utilization and Impact - The satellite embedding dataset powered by AEF is one of the largest of its kind, containing over 1.4 trillion footprints annually, and is used by organizations like the UN Food and Agriculture Organization [40][41]. - This dataset aids in projects like the "Global Ecosystem Atlas," which aims to map and monitor global ecosystems, crucial for identifying conservation priorities [40][41]. - AEF contributes to addressing critical issues such as food security, deforestation, and water resources, highlighting the role of AI models as public infrastructure [43][44].