# From Pixels to Patches: Pooling Strategies for Earth Embeddings

arXiv: https://arxiv.org/abs/2603.02080 · HTML: https://mlsanity.com/paper/2603.02080/

## Key facts

- Cited by 4 papers in the map (data through 14 Oct 2026).
- Builds on 9 of its 13 references that are in the map.
- First posted on arXiv 2 Mar 2026.

## Builds on (most cited first, 9 of 9)

- [EuroSAT: A Novel Dataset and Deep Learning Benchmark for Land Use and Land Cover Classification](https://mlsanity.com/paper/1709.00029/) (2017)
- [Fine-tuning CNN Image Retrieval with No Human Annotation](https://mlsanity.com/paper/1711.02512/) (2017)
- [AlphaEarth Foundations: An embedding field model for accurate and efficient global mapping from sparse label data](https://mlsanity.com/paper/2507.22291/) (2025)
- [Learnable pooling with Context Gating for video classification](https://mlsanity.com/paper/1706.06905/) (2017)
- [In-domain representation learning for remote sensing](https://mlsanity.com/paper/1911.06721/) (2019)
- [Augmenting Convolutional networks with attention-based aggregation](https://mlsanity.com/paper/2112.13692/) (2021)
- [OlmoEarth: Stable Latent Image Modeling for Multimodal Earth Observation](https://mlsanity.com/paper/2511.13655/) (2025)
- [Distribution Shifts at Scale: Out-of-distribution Detection in Earth Observation](https://mlsanity.com/paper/2412.13394/) (2024)
- [Handling Image and Label Resolution Mismatch in Remote Sensing](https://mlsanity.com/paper/2211.15790/) (2022)

## Built on by (most cited first, 4 of 4)

- [How do Self-Supervised Remote Sensing Vision Models Transfer to Downstream Tasks?](https://mlsanity.com/paper/2606.13896/) (2026)
- [Earth Embeddings](https://mlsanity.com/paper/2608.03410/) (2026)
- [Geospatial Foundation Models Capture Health-Relevant Dimensions of Place Beyond Conventional Social Risk Indices](https://mlsanity.com/paper/2609.11689/) (2026)
- [Planetary Feature Fields are Scalable Earth Representations](https://mlsanity.com/paper/2609.37784/) (2026)

## Abstract (arXiv)

Geospatial foundation models increasingly expose pixel-level embedding products that can be downloaded and reused without access to the underlying encoder. In this setting, downstream tasks with patch- or region-level labels require a post-hoc aggregation step that maps dense pixel embeddings to a single representation. The default choice, mean pooling, discards within-patch variability and can underperform under spatial distribution shift. To study this setting, we introduce EuroSAT-Embed: 81,000 embedding GeoTIFFs derived from three foundation models: AlphaEarth, OlmoEarth, and Tessera. Using these fixed embedding products, we benchmark 11 training-free pooling methods and 2 train-set-fitted baselines under both random and geographically disjoint test splits. Richer pooling schemes reduce the geographic generalization gap by over 50% relative to mean pooling and improve accuracy by up to 6% on spatial splits. We recommend a three-tier strategy: (1) mean as a baseline, (2) stats pooling (min/max/mean/std) as the default at 4x the embedding dimension, and (3) covariance pooling for peak accuracy. Across all three embedding products, simple distributional statistics improve spatial-split performance over mean pooling.