What’s in an Earth Embedding? An Explainability Analysis of Location Encoders

Geographic implicit neural representations (INRs) are increasingly used to create 'Earth embeddings'—compact, queryable representations of geospatial data. While these embeddings are effective for downstream tasks like climate modeling or urban planning, they function as black boxes, making it difficult for researchers to understand exactly what geographic or semantic information they capture. This paper introduces a framework to audit these embeddings by decomposing them into human-interpretable components.

The authors propose three complementary methods to interpret location embeddings:

• Sparse Latent Concepts: Using sparse autoencoders (SAEs), the authors disentangle embeddings into a set of sparse, monosemantic neurons. This allows them to identify specific latent features that correspond to distinct geographic or visual patterns.
• Natural Language Concepts: By aligning location embeddings with a text-based semantic space (using the SpLiCE framework), the authors decompose embeddings into a predefined dictionary of geospatial terms, such as 'forest,' 'desert,' or 'urban.'
• Visual Feature Attribution: The authors utilize CLIP Surgery to generate saliency maps, which highlight the specific visual regions (e.g., roads, landmarks) that influence a location encoder's output.

The study finds that these decomposition methods successfully reveal the internal structure of location encoders. Sparse autoencoders were able to reconstruct embeddings with high fidelity while isolating neurons that activate for specific, coherent geographic regions or visual concepts. For example, certain neurons consistently activated for rainforests in the Amazon or archaeological sites in the Middle East. The authors also demonstrate that these methods can identify artifacts in pretraining datasets, such as visual noise in satellite imagery, providing a diagnostic tool for model developers.