Research map / Architectures and efficiency / RNNs and normalization
Decomposing Earth Embeddings with Sparse Autoencoders
Vitus Benson, Fanny Yang, Markus Reichstein
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Applies sparse autoencoders to decompose Earth observation foundation model embeddings into more interpretable sparse features.
Problem
Earth observation foundation model embeddings are opaque because each dimension entangles unrelated physical concepts, hurting interpretability and robustness for practitioners and scientists.
Approach
Proposes using sparse autoencoders as a principled way to decompose dense Earth embeddings into sparse, disentangled concept-like features.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use trained SAE instead of PCA at matched sparsity | Earth embeddings · AlphaEarth reconstruction L0=16 | fraction of variance explained: 0.70 -> 0.78 (frozen decoder) | Improved |
| train the SAE decoder instead of freezing a random dictionary (ablation) | BatchTopK SAE · AlphaEarth reconstruction L0=16 | unexplained variance (%): 22 -> 2.7 | Improved |
| train the SAE decoder instead of freezing random dictionary (ablation) | BatchTopK SAE · Tessera reconstruction L0=16 | unexplained variance (%): 47 -> 8.4 | Improved |
| use SAE features instead of raw dimensions | Earth embeddings · AlphaEarth concept alignment | median groups aligned: 12 -> 88% of features align with at most one group | Improved |
| train the decoder instead of freezing a random dictionary (ablation) | SAE · AlphaEarth concept alignment | share aligned with at most one group: 88% vs 89% | No clear effect |
| use SAE features instead of raw dimensions | Earth embeddings · AlphaEarth spatial scale | within-patch variance share: 9% -> 23-32% | Improved |
| train decoder instead of freezing random dictionary (ablation) | SAE · AlphaEarth spatial scale | median correlation length: 23 km (frozen) vs 20-33 km | No clear effect |
| use SAE features for concept detection instead of raw dimensions | AlphaEarth · concept detection | mean folded AUROC: 0.84 -> 0.75 | Hurt |
| compute AoA distance in SAE code instead of raw embedding | ridge regressor on canopy height · OOD error ranking | Spearman rho: +0.17 (AlphaEarth), +0.21 (Tessera) median gain | Improved |
| compute AoA distance in SAE code instead of raw embedding | ridge regressor on elevation · OOD error ranking | Spearman rho | No clear effect |
| use trained SAE instead of untrained random sparse code for OOD ranking (ablation) | AoA distance · canopy height | Spearman rho gain: random code gains as much on AlphaEarth; +0.07 on Tessera | No clear effect |
| increase dictionary size M from 512 to 2048 (ablation) | SAE at L0=16 · reconstruction | FVE: <0.01 change | No clear effect |
Experimental setup
Builds on: sparse autoencoders
Closest papers by meaning
- No Location Left Behind: Measuring and Improving the Fairness of Implicit Representations for Earth Data (ICLR 2025 · similarity 0.89)
- GeoCore-9B: Towards Geo-Aware Generative Foundation Models in Earth Observation (NeurIPS 2026 · similarity 0.87)
- DUNIA: Pixel-Sized Embeddings via Cross-Modal Alignment for Earth Observation Applications (ICML 2025 · similarity 0.87)
- InstructSAM: A Training-free Framework for Instruction-Oriented Remote Sensing Object Recognition (NeurIPS 2025 · similarity 0.86)
- AnySat: One Earth Observation Model for Many Resolutions, Scales, and Modalities (CVPR 2025 · similarity 0.86)
- Learning without Exact Guidance: Updating Large-scale High-resolution Land Cover Maps from Low-resolution Historical Labels (CVPR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
