Google Research unveiled S2Vec, a self-supervised framework transforming how AI decodes urban environments. Unlike manual geospatial analysis, S2Vec converts roads, buildings, and infrastructure into AI-readable embeddings using masked autoencoding. This enables models to recognize patterns like human planners, predicting socioeconomic metrics without labeled data.

The system employs two breakthrough techniques: S2 Geometry partitions Earth into scalable cells, while feature rasterization turns locations into multi-layered images. For instance, a cell with three coffee shops and one park becomes a visual "color" code. This rasterization bridges the gap between raw geospatial data and computer vision models, allowing AI to "see" urban landscapes.

Evaluations show S2Vec outperforms image-based baselines in socioeconomic predictions, excelling at zero-shot geographic adaptation. However, environmental tasks like tree cover analysis require combining S2Vec with satellite imagery embeddings. The framework's strength lies in its ability to group neighborhoods like financial districts without explicit definitions, based purely on spatial relationships.

By eliminating manual labeling, S2Vec unlocks global insights at planetary scale. Urban planners can now model infrastructure impacts on neighborhood health, while environmental researchers improve carbon footprint predictions. This aligns with Google's Earth AI mission, supported by complementary tools like the Population Dynamics Foundation Model.