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

Researchers have analyzed geographic implicit neural representations (INRs) and found that location embeddings can be broken down into human-interpretable features, revealing geographic structures such as forests and urban features.

A study submitted on June 23, 2026, examined the information content of geographic INRs through their location embeddings^[1]. The researchers decomposed these embeddings into sparse latent concepts, natural language concepts, and visual features. They found that location embeddings can be broken down into human-interpretable representations while retaining high reconstruction capability. Meanwhile, another study on synthetic data augmentation found that while it can extend existing datasets with realistic images, automatic metrics such as FID and KID may not reflect downstream utility^[2]. The study also showed that semantic segmentation models trained on mixed real-synthetic datasets can be unreliable indicators for geospatial data, and that human evaluation is a more reliable indicator of synthetic data quality. The analysis of geographic INRs and synthetic data augmentation highlights the complexities of working with Earth observation datasets.