OT on the Map: Quantifying Domain Shifts in Geographic Space
| Authors | Haoran Zhang et al. |
| Year | 2026 |
| Field | Machine Learning |
| arXiv | 2604.16220 |
| Download | |
| Categories | cs.LG |
Abstract
In computer vision and machine learning for geographic data, out-of-domain generalization is a pervasive challenge, arising from uneven global data coverage and distribution shifts across geographic regions. Though models are frequently trained in one region and deployed in another, there is no principled method for determining when this cross-region adaptation will be successful. A well-defined notion of distance between distributions can effectively quantify how different a new target domain is compared to the domains used for model training, which in turn could support model training and deployment decisions. In this paper, we propose a strategy for computing distances between geospatial domains that leverages geographic information with Optimal Transport methods (GeoSpOT). In our experiments, GeoSpOT distances emerge as effective predictors of cross-domain transfer difficulty. We further demonstrate that embeddings from pretrained location encoders provide information comparable to image/text embeddings, despite relying solely on longitude-latitude pairs as input. This allows users to get an approximation of out-of-domain performance for geospatial models, even when the exact downstream task is unknown, or no task-specific data is available. Building on these findings, we show that GeoSpOT distances can preemptively guide data selection and enable predictive tools to analyze regions where a model is likely to underperform.
Engineering Breakdown
Plain English
This paper addresses a critical problem in geographic machine learning: models trained on data from one region often fail when deployed in another region due to distribution shifts, yet there's no principled way to predict when cross-region adaptation will succeed. The authors propose GeoSpOT, a method that uses Optimal Transport theory combined with geographic information to quantify the distance between geospatial domains, enabling engineers to measure how different a target region is from training regions. This distance metric can guide decisions about whether to retrain models, apply domain adaptation techniques, or use different architectures for specific geographic areas. The work directly addresses the uneven global data coverage problem that affects computer vision and machine learning systems deployed across regions.
Core Technical Contribution
The core novelty is combining Optimal Transport (OT) with geospatial structure to create a domain distance metric that respects geographic properties rather than treating regions as unlabeled data points. Previous domain adaptation work either ignored geographic structure or used simple proxy measures (like pixel-level distribution differences) that don't account for spatial continuity and regional coherence. GeoSpOT treats geographic space as a structured metric space where closer regions should naturally have smaller domain distances, then computes OT distances that incorporate this geographic constraint. This is the first principled, theoretically grounded approach to quantify domain shifts specifically in geographic space, moving beyond generic domain divergence measures like CORAL or Maximum Mean Discrepancy that don't leverage spatial information.
How It Works
The input is geospatial data from multiple regions (e.g., satellite imagery, census data, climate measurements) along with their geographic coordinates and a learned feature representation of each region from a trained model. The method first constructs a cost matrix between regions that combines feature-space distance (how different the learned representations are) with geographic distance (how far apart regions are spatially), weighted by a hyperparameter that controls the influence of geography. Optimal Transport then solves for the minimum-cost coupling between source and target region distributions, which quantifies the effort needed to transform one region's data distribution into another's while respecting geographic structure. The output is a scalar distance value for each target region relative to training regions, where smaller values indicate the model should generalize well and larger values signal distribution shift requiring adaptation. The geographic weighting ensures that two regions far apart are expected to have larger domain distances, and nearby regions with large feature-space differences are flagged as high-risk gaps in training coverage.
Production Impact
For teams deploying computer vision or ML models across geographic areas (agriculture, climate, urban planning, disaster response), this provides a concrete pre-deployment check: compute GeoSpOT distance to flag regions where the model is likely to underperform, triggering decisions to gather local data, fine-tune, or use region-specific models. The distance metric can be integrated into model serving pipelines as a runtime check—when a prediction request comes from a new geographic area, compute its distance to training regions and either apply adaptive inference or trigger model retraining workflows. This reduces costly failures in remote regions where ground truth is expensive to collect; instead of blind deployment, engineers can quantify risk upfront. The computational cost is moderate (solving OT is polynomial in number of regions) and done offline during model evaluation, not at inference time, making it practical for production systems.
Limitations and When Not to Use This
The paper assumes that geographic distance is a meaningful proxy for domain similarity, which breaks down when regions are geographically close but have very different data characteristics (e.g., two adjacent countries with different data collection standards or two cities with different weather patterns). The method requires labeled or representative data from multiple regions to compute meaningful distances, making it less useful in scenarios with only one source region or extremely sparse geographic coverage. The choice of weighting between feature distance and geographic distance is a hyperparameter that must be tuned, and the paper doesn't provide clear guidance on how to set this for new domains, potentially requiring expensive validation. The abstract indicates the full experimental results are missing, so the actual empirical performance, scalability to hundreds of regions, and comparison against baselines are not yet visible in this excerpt.
Research Context
This work builds on decades of domain adaptation research (DANN, CORAL, OT-based DA methods) but extends it into the geographic domain where spatial structure matters—prior work treated domains as unordered categories rather than points in continuous space. It connects to the broader literature on geographic machine learning fairness and out-of-distribution generalization, where uneven data coverage (more labeled data in wealthy regions) creates systematic generalization failures in underrepresented areas. The paper opens a research direction for spatially-aware domain metrics that could extend to other structured domains (temporal shifts, hierarchical organization) and provides a foundation for principled geographic model selection and active learning strategies that prioritize data collection in high-risk regions.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::