Mixture of Geographical Experts:
Disentangling Earth

Moien Rangzan^1,2 · Gregory Duveiller² · Maha Shadaydeh¹ · Markus Reichstein² · Joachim Denzler^1
1Computer Vision Group, Friedrich Schiller University Jena    ²Max Planck Institute for Biogeochemistry, Jena

_{Routing weights of MoGE's six experts on FMoW (stronger color means higher weight). The regions are learned from coordinates alone.}

Note

Official implementation of MoGE (GCPR 2026). The code will be available soon. Star or watch the repository to get notified.

TL;DR

Most domain generalization methods assume that one predictor P(Y | X) holds everywhere. In Earth observation it often does not: a cluster of trees can be a natural forest in one region and a plantation in another. MoGE is a geo-routed Mixture-of-Experts that separates what is invariant everywhere from what varies across space.

• Discovers its own domains. A router driven by latitude and longitude, not by the image, activates a few geo-specialized experts. They self-organize into continuous, concept-consistent regions, so no hand-crafted domains are needed.
• Two models in one. An always-active shared expert learns features that hold everywhere. On its own (MoGE-SE), it is a domain-invariant model from the same training run.
• Robust to location-class bias. A probabilistic group-robust objective over router-induced soft groups keeps experts from relying on harmful regional priors.
• Adapts cheaply. Tuning only the gate (~5% of the weights) on 150 labeled samples adapts MoGE to an unseen continent.

Method

MoGE replaces the feed-forward network of selected ViT blocks with a geo-routed mixture of experts:

$$f(x,m) = \underbrace{\gamma\, f_0(x)}_{\text{invariant}} + \underbrace{(1-\gamma)\sum_{k=1}^{E}\alpha_k\big(\phi(m)\big)\, f_k(x)}_{\text{specialized}}$$

(a) The location embedding φ(m) is matched against a learned expert codebook by cosine similarity, and the top-k experts are mixed. The shared expert f₀ is always active. (b) A supervised contrastive loss makes f₀ invariant across the discovered domains. (c) A CKA penalty keeps the specialists diverse. (d) Two MoGE blocks sit at layers 9 and 11 of a ViT-Tiny, with 6 experts and top-3 routing.

Results

Both benchmarks use the same architecture and hyperparameters, with PG-DRO switched on for MapInWild. Values are mean ± std, with the best in bold. The table shows the strongest baselines.

Method	FMoW		MapInWild
	WRA	Acc.	WGA	WCA	Acc.
ERM	32.43 ±1.67	53.69 ±0.37	52.18 ±4.38	70.40 ±3.43	77.16 ±1.73
GroupDRO	30.70 ±0.80	49.06 ±0.37	57.25 ±0.75	73.87 ±1.05	75.70 ±0.75
Fish	33.08 ±0.29	44.99 ±0.72	52.02 ±1.85	69.20 ±1.14	76.72 ±1.39
ERM + LE	35.83 ±0.51	53.23 ±0.72	48.76 ±9.12	70.58 ±3.43	74.48 ±1.46
D3G + LE	34.60 ±1.27	50.76 ±0.12	55.16 ±6.33	70.01 ±5.31	77.82 ±1.25
DCP + LE	35.97 ±3.98	52.30 ±1.08	51.39 ±7.98	64.43 ±4.91	75.42 ±1.46
MoGE-SE (ours)	33.01 ±0.34	48.60 ±0.41	57.83 ±4.37	74.32 ±1.07	77.50 ±0.96
MoGE (ours)	38.85 ±2.50	55.25 ±0.57	60.91 ±1.38	74.20 ±1.51	77.20 ±0.53

_{WRA, WGA, WCA: worst-region, worst-group and worst-class accuracy. LE: location encoder. MoGE-SE: the shared expert alone.}

Transfer to an unseen continent. We train with Asia held out (FMoW-LAO), then tune only the gate on a few labeled Asian samples while the backbone and experts stay frozen. With 150 samples (1% of the validation set), MoGE already beats ERM and the domain-generalization baselines.

_{GT x%: gate tuned on x% of the Asian validation set. MoGE-SC: MoGE with a frozen, pretrained SatCLIP encoder. MoGE-Ex: MoGE without any adaptation.}

Citation

@inproceedings{rangzan2026moge,
  title     = {Mixture of Geographical Experts: Disentangling Earth},
  author    = {Rangzan, Moien and Duveiller, Gregory and Shadaydeh, Maha and
               Reichstein, Markus and Denzler, Joachim},
  booktitle = {Pattern Recognition (Proceedings of DAGM GCPR 2026)},
  publisher = {Springer Nature},
  year      = {2026}
}

Contact

Questions are welcome. Open an issue or email Moien Rangzan at mrangzan@bgc-jena.mpg.de.