OlmoEarth v1.1: A more efficient family of OlmoEarth models

Abstract: We present a set of improvements to the OlmoEarth family. These improvements allow us to cut compute costs during training ($1.7 \times$ reduction in GPU hours required to train our Base models) and inference ($2.9\times$ reductions in MACs on Sentinel-2 tasks), while maintaining the models' overall performance. All training code is available at github.com/allenai/olmoearth_pretrain.

• The paper demonstrates significant compute reduction by merging bandsets and optimizing input handling.
• It introduces random band dropout and non-linear projection to recover accuracy and stabilize training.
• Revised temporal masking and loss adjustments yield efficiency improvements with minimal impact on overall performance.

Efficient Advancements in Multimodal Earth Observation: OlmoEarth v1.1

Introduction

"OlmoEarth v1.1: A more efficient family of OlmoEarth models" (2605.20804) presents a set of architectural and algorithmic modifications to the OlmoEarth foundation models, focusing on a reduction in computational cost while preserving and—in select benchmarks—even improving performance across a range of Earth observation tasks. The OlmoEarth models are large-scale encoder-decoder ViTs pre-trained with masked image modeling on multi-modal, multi-temporal satellite imagery and derived map products. Version 1.1 introduces specific changes to input handling, masking, and the loss function, and demonstrates that these yield Pareto improvement in computational efficiency versus accuracy trade-offs.

Model Architecture Modifications and Motivations

The transition from OlmoEarth v1 to v1.1 centers on three high-impact changes:

Single band set representation:

OlmoEarth v1 processed bands from Sentinel-2 and Landsat in multiple bandsets grouped by spectral resolution, which led to significant token and compute inflation. OlmoEarth v1.1 merges all bands for each modality into single bandsets, reducing token counts (and thus compute) by 3x for Sentinel-2 and Landsat data. However, naively collapsing bands resulted in a notable $\sim$10% drop in accuracy on tasks like m-eurosat due to information loss in patch embedding and diminished cross-band interactions.

Random band dropout and non-linear projection:

To mitigate representational collapse in the merged band regime, two methods were introduced:

• Random band dropout stochastically removes bands during training, preventing the model from relying on any single band and forcing learning of cross-band dependencies.
• Non-linear patch projection (linear layer with a ReLU-activated hidden dimension) replaces the purely linear patch embed, which stabilizes optimization by reducing gradient norm pathologies seen in the single-bandset setting. This design choice aligns with findings in ViT training for natural images, where early non-linearities facilitate optimization [xiao2021early].

Revised masking and loss function:

OlmoEarth v1.1 removes the two-stage masking with bandset-level suppression, instead always passing map modality targets to the loss. A temporal masking strategy is introduced, which applies mask blocks along the temporal axis to promote temporal reasoning, following evidence that structured spatiotemporal masking boosts downstream generalization [cong2022satmae, szwarcman2024prithvi, tseng2025galileo]. Additionally, the loss omits extremely difficult negatives (those nearly identical to the target) in the patch discrimination contrastive objective, further stabilizing training and yielding slight improvements.

Empirical Results

Pareto-dominant computational efficiency:

OlmoEarth v1.1 establishes a new Pareto frontier of performance vs. inference and training cost. For the Base model, v1.1 reduces MACs by $2.9\times$ compared to v1, and end-to-end pretraining GPU hours decrease by $1.7\times$ (v1 Base: 2,989 hours; v1.1 Base: 1,763 hours), with associated reductions in energy, carbon, and water footprints.

Task-averaged performance:

Comprehensive evaluations on standard kNN and linear probe downstream tasks and selected real-world applications demonstrate that performance is largely maintained:

• Task-averaged kNN/linear probe scores are essentially unchanged between v1 and v1.1 (Nano: $60.1 \to 60.9$, Tiny: $61.9 \to 62.1$, Base: $65.2 \to 64.3$).
• Fine-tuning results are similarly stable (Nano: $73.0 \to 72.8$, Tiny: $77.0 \to 77.0$, Base: $79.0 \to 78.5$).
• In some instances, v1.1 outperforms v1 (e.g., on m-bigearthnet, BreizhCrops, PASTIS), while regressing on others (notably, CropHarvest tasks). These drops are minor relative to the compute savings.

Ablation analyses confirm that the combination of random band dropout and a non-linear projection layer is essential for recovering lost accuracy from the bandset merging. Temporal masking consistently enhances validation scores across diverse tasks, with low sensitivity to its application probability.

Practical and Theoretical Implications

From a practical perspective, the reduction in inference and training compute directly translates to lower cost, faster modeling workflows, and diminished environmental impact, especially when producing large-scale downstream maps or running extensive fine-tuning experiments. The results are particularly salient for users with constrained hardware or strict carbon budgets.

Theoretically, OlmoEarth v1.1 empirically validates a route to efficient masked image modeling for highly multi-modal, multi-band, and multi-temporal data typical in Earth observation. It demonstrates how architectural efficiency can be systematically advanced without trading off accuracy, provided careful input regularization and early non-linearities are applied. The observed robustness of temporal masking also underlines the growing importance of designing pretext tasks sensitive to spatiotemporal context in multimodal transformers.

Importantly, efficiencies achieved in OlmoEarth v1.1 can inform designs for other large-scale geospatial foundation models, autonomous perceptual systems, or multimodal encoders in settings beyond remote sensing.

Potential Future Directions

Future work could extend several directions highlighted by the OlmoEarth v1.1 findings:

• Analysis of model size effects: The disproportionately strong improvements for Tiny variants suggest further study on how early-layer capacity and regularization interact with model depth/width trade-offs.
• Extending to broader input modalities: As data grows in heterogeneity (e.g., addition of radar, LiDAR, thermal), evaluating the efficacy of single-bandset representations with nonlinear projections for even higher-bandwidth inputs is warranted.
• Transfer and scalability: Given the established efficacy at lower computational budget, investigating the transfer, compositionality, and scalability of the OlmoEarth v1.1 pretraining recipe to other geospatial or visual domains is of high value.
• Refinement of masking and contrastive objectives: The positive effects of temporal masking and hard-negative mining present opportunities to develop richer, adaptive masking and contrastive methods that further exploit modality, time, and spatial information.

Conclusion

OlmoEarth v1.1 concretely demonstrates that systematic architectural and algorithmic simplifications—guided by targeted regularization and pretext task design—can yield foundation models that are substantially more efficient without degradation of generalization performance on diverse Earth observation tasks. This approach points towards more accessible, environmentally responsible, and scalable large vision models suitable for planetary-scale analytics, and sets important methodological baselines for the continued evolution of foundation models in multimodal domains.