Do Satellites See Commuters? A Critical Benchmark of Vision Foundation Models

arXiv:2609.00661 · cs.CV · Submitted 2026-09-01 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Do Satellites See Commuters? A Critical Benchmark of Vision Foundation Models".

Jane: The gist: Satellite foundation models offer a globally available alternative to census data for commuting origin-destination (OD) generation,

Tom: First, who's behind it and why it matters.

Paper summary: Tom: Hey everyone, welcome back to the show. We’re talking about this new paper from SIGSPATIAL ’twenty-six called "Do Satellites See Commuters <ref:2609.00661#pg2>? A Critical Benchmark of Vision Foundation Models." It looks like they're trying to figure out if satellite images can replace census data for figuring out how people move around.

Jane: That sounds really practical, Tom. So, the main question they’re tackling is whether these satellite foundation models actually work for generating origin-destination data when we don't have that detailed census information.

Lu: Exactly. They’re looking at four different ways to build the vision encoder—language-supervised, self-supervised, geographically grounded—and testing them all in the same setup to see which one performs best in actually making those OD predictions.

Meng: I'm curious about what makes this important for deployment. It seems like they are trying to find a robust way to get usable results without just throwing whatever model is currently trending into production, right?

Lalam: From my perspective, this paper systematically compares the four main ways they trained these models: RemoteCLIP which uses language alignment, DINOv3 which uses self-supervised patch discrimination, and then there are SatCLIP and AlphaEarth which focus on grounding features to coordinates or combining different sensor types.

Tom: Right. So what's the core setup here? They put all these encoders into an identical WeDAN graph diffusion framework across a big set of data—one thousand nine hundred twenty-five US counties, three hundred twenty-five UK districts, and fourteen global cities under five random seeds <ref:2609.00661#pg1,1,925 US counties, 325 UK districts, and 14 global cities under>.

Jane: It’s a big test because it forces them to compare the encoders fairly against each other without letting one method have an unfair advantage just because it's easier to train. They are testing the representational fidelity of these four different vision encoder paradigms across this multi-scale benchmarking pipeline for generative urban flow networks, which is what they’re aiming for.

Lu: That setup is key because it shows how much the choice of encoder paradigm affects the final output quality, especially when we look at performance in different regions. They are looking at how well these models translate their internal understanding of the satellite image into actual commuting patterns.

Paper summary: Meng: So, what did they actually find that surprised them? I mean, we hear a lot about models getting better with more data, so what was the big takeaway on scale versus quality here?

Tom: Well, their first finding shows that language-supervised features from RemoteCLIP are getting the strongest in-distribution performance with a CPC of zero point six zero two <ref:2609.00661#pg1,the strongest in-distribution performance>. But they also found that geographically grounded encoders transfer more reliably zero-shot to other areas; specifically, AlphaEarth improved the CPC by thirty-three percent over RemoteCLIP when tested on UK districts <ref:2609.00661#pg1,geographically grounded encoders transfer more reliably zero-shot>.

Jane: That’s interesting because it suggests that for certain tasks, knowing what those places are called—the language part—is very helpful, but having a clear physical location might be even more powerful for generalization.

Lu: And then they hit a point where they showed that just having a bigger pretraining corpus isn't enough. They tested DINOv3, which was trained on a substantially larger satellite corpus, and it actually underperformed RemoteCLIP in-distribution by zero point zero nine one CPC and even collapsed to zero point zero two two globally.

Meng: That’s concerning for deployment because if you can't trust the scale of the training data alone, you have to worry about how that model will behave when it runs on real traffic data where those patterns might be different.

Lalam: And then they looked at cross-continental travel, which is a big deal. Their third finding was that no encoder transfers usefully to global cities; the best CPC they saw was zero point one two two for RemoteCLIP and zero point zero two two for DINOv3 globally.

Jane: So, the paper confirms what many of us suspect: generating commuting patterns across different continents is still a really open problem right now with these methods. It’s not a solved issue yet.

Tom: They also spent time clarifying some details about the census noise parameter eta. The ordering of this parameter reverses under cross-continental evaluation, which is critical for correctly interpreting their prior results.

Lu: And they also pointed out that there are trade-offs between the encoders themselves. For example, DINOv3’s self-distillation objective really specializes in microscale surface geometry and texture discrimination, which they found was less relevant for zone-level prediction compared to RemoteCLIP’s language supervision.

Meng: From an engineering standpoint, that specialization means you have to choose your encoder based on what kind of data you're actually feeding it, like if you need fine texture details or just broad semantic context.

Paper summary: Jane: And they did give us some operational advice regarding the census noise parameter eta. They report that all US in-distribution results are at eta equals one but for UK and global zero-shot results, they set it to zero. This means practitioners should evaluate strictly at eta equals zero if the census channel is absent to avoid silent performance degradation.

Tom: That’s a very clear instruction for anyone deploying these models in regions where census data isn't available; you have to treat those results as being at eta equals zero.

Lu: Overall, this paper on "Do Satellites See Commuters? A Critical Benchmark of Vision Foundation Models" shows that the choice between a language-aligned feature and a geographically grounded one really depends on the specific geographical region you are targeting and whether you have access to census data.

Jane: So, what does this mean for us outside of just building a model? It suggests that when we want to get these satellite models working across different countries or regions, focusing on the physical location grounding is a more reliable path than just relying on massive pretraining data from self-supervised methods.

Meng: If I’m deploying this at scale, it tells me I need to decide if my deployment area is well-represented in the training data for DINOv3 or if I should lean into the structural grounding of AlphaEarth when census data is missing.

Lalam: The paper shows that while language alignment gives you strong results locally, and AlphaEarth helps globally, there isn't one single answer for how to get reliable OD generation from satellite imagery right now.

Tom: So, if we look at the title of this paper, "Do Satellites See Commuters? A Critical Benchmark of Vision Foundation Models," it really puts the focus on establishing a solid way to test these models against each other systematically rather than just picking the best one by its training size.

Jane: It’s about moving past just asking if satellites *can* see commuters, and instead building a rigorous way to benchmark which type of satellite vision foundation model is actually good at that task when we don't have perfect ground truth data.

Conclusion: Tom: So we just finished looking at this paper, "Do Satellites See Commuters? A Critical Benchmark of Vision Foundation Models." It's been a deep dive into how these satellite foundation models actually perform when we’re trying to predict where people are going in cities.

Jane: Yeah, the main thing here is that they didn't just pick one model and say it's the best. They systematically compared four different types of vision encoders—language-supervised, self-supervised, geographically grounded—all running under the same test conditions.

Lu: Exactly, and what’s really interesting is how those choices change the results depending on where you’re testing them. They found that language supervision gave them a strong start in most cases, but geographic grounding helped them generalize better to completely new areas, like UK districts.

Meng: From a deployment standpoint, that's huge. It tells us we can't just take the biggest pretraining dataset and expect it to work everywhere; the structure of what you trained it on matters a lot for real-world accuracy.

Lalam: And they highlighted that simply having a bigger training set isn't enough, because the self-supervised model, DINOv3, ended up performing worse in some tests than the language-supervised one.

Tom: Right, and their third major finding was that no single encoder transferred usefully to global cities. They couldn't get good results for those cross-continental OD generations at all.

Jane: That confirms what we suspected about global travel—it’s still a big open problem in this space right now. It’s not solved yet, and this paper proves it by showing the performance floor across the board.

Lu: The authors also spent time clarifying how they measured census noise, because that parameter's meaning flips when you look at data from different continents, which is crucial for interpreting their results correctly.

Meng: That kind of nuance is what separates a good benchmark from just a set of pretty numbers. It shows they really thought through the operational side of it.

Lalam: And finally, they gave us some concrete advice on when to use which encoder based on whether you have census data available or not.

Tom: So, the authors are essentially saying that there isn't one perfect satellite vision foundation model for every city, and you need to pick your tool based on your specific deployment region and what data you’re missing.

Jane: It really frames the next steps here as choosing the right tool for a specific job instead of chasing a single best answer. That’s a big shift in how we think about deploying this technology.

University of New South Wales

cs.CV

Submitted: 2026-09-01

Updated: 2026-10-19

Comments: Accepted to SIGSPATIAL 2026

Code: https://github.com/cruiseresearchgroup/satellites-see-commuters

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 79/100

The gist: The gist: Satellite foundation models offer a globally available alternative to census data for commuting origin-destination (OD) generation, yet no study has systematically compared encoder

Key concepts

RemoteCLIP: This encoder uses a massive language-image pretraining method (CLIP) to understand high-level human geography and socioeconomic meanings embedded in satellite images. It is good at understanding what places are generally like based on text descriptions.
DINOv3: This model is trained using self-supervised learning, where it teaches itself by comparing different patches of satellite images without human labels. It excels at recognizing fine details like textures and surface geometry but struggles with broader geographic context compared to language models.
AlphaEarth: This encoder moves beyond just visual data by combining information from six different sensors, including optical, radar, and elevation data. It grounds the satellite features in physical reality by looking at multiple physical measurements simultaneously.

Terminology

Benchmark — 1× in this paper · explained in 2 episodes
Foundation Model — 1× in this paper · explained in 3 episodes
Foundation Models — 1× in this paper · explained in 7 episodes

Summary

The gist: Satellite foundation models offer a globally available alternative to census data for commuting origin-destination (OD) generation, yet no study has systematically compared encoder paradigms within a single downstream pipeline.

How it works

The research systematically compares four satellite vision encoders—language-supervised (RemoteCLIP), self-supervised (DINOv3), geographically grounded (SatCLIP, AlphaEarth)—within an identical WeDAN graph diffusion framework across 1,925 US counties, 325 UK districts, and 14 global cities under five random seeds<ref:2609.00661#pg9>. The core methodology involves conditioning the underlying WeDAN graph diffusion pipeline on node-level feature vectors derived from satellite imagery and census data.

The node feature vector is constructed by fusing a spatial vision model embedding with noise-controlled census data. Specifically, the final node vector fed to the GraphTransformer is constructed as n r = [x r a r a r] in R(d+2)+97+97. This structural formulation explains why the downstream network input dimension scales directly with the choice of spatial foundation model.

Encoder Paradigms Evaluated

The study evaluates four distinct vision encoder paradigms, each optimized under fundamentally different pretraining objectives. These include:

Language-supervised features: RemoteCLIP, which utilizes a massive contrastive language-image pretraining (CLIP) objective to embed high-level anthropogenic geography and socioeconomic semantics.
Self-supervised patch discrimination: DINOv3, which optimizes a vision transformer backbone using a teacher-student self-distillation objective over a corpus of 493 million satellite images.
Coordinate-contrastive grounding: SatCLIP, which discards language completely and forces the visual transformer to map satellite patches directly to their exact global GPS coordinates.
Multisensor physical grounding: AlphaEarth, which shifts away from heavy visual transformers entirely by assimilating six distinct sensor modalities across optical, radar, and elevation datastreams.

Key Empirical Findings

Three main findings emerged from the systematic evaluation of these encoders.

Language-supervised features achieve the strongest in-distribution performance (RemoteCLIP CPC 0.602), while geographically grounded encoders transfer more reliably zero-shot: AlphaEarth improves CPC by 33% over RemoteCLIP on UK districts.
Pretraining corpus scale alone is insufficient: DINOv3, trained on a substantially larger satellite corpus, underperforms RemoteCLIP by 0.091 CPC in-distribution and collapses to CPC 0.022 globally.
No encoder transfers usefully to global cities (best CPC 0.122 for RemoteCLIP, 0.022 for DINOv3), confirming cross-continental OD generation remains an open problem.

Representational Tradeoffs and Semantics

The analysis highlights a representational tradeoff across vision encoders that differ in pretraining corpus scale, objective, architecture, and preprocessing. For instance, DINOv3’s self-distillation objective specializes in microscale surface geometry and texture discrimination. This specialization is found to be less relevant for zone-level prediction compared to RemoteCLIP’s contrastive language supervision. Furthermore, the paper clarifies the semantics of the census noise parameter (eta), demonstrating that it measures census data availability rather than satellite feature noise. This distinction is critical to correctly interpreting prior results.

Operational Implications and Recommendations

The findings have direct operational consequences regarding the census noise parameter. The authors report all US in-distribution results at eta = 1 and all UK and Global zero-shot results at eta = 0, treating these as the declared operating points for every headline comparison. Practitioners deploying these models in census-absent regions must evaluate strictly at eta = 0 to avoid silent performance degradation. For instance, AlphaEarth is the better choice only when the census channel is guaranteed to be absent.

Overall, the study demonstrates that geographically grounded encoders significantly outperform language-aligned features on the UK evaluation dataset. These results suggest that embedding coordinate-contrastive or geophysical structural priors within deep vision architectures is a promising direction for geographically transferable urban mobility generation at scale. The optimal choice of encoder depends on deployment region, census data availability, and hardware constraints. This work is licensed under a Creative Commons Attribution 4.0 International License. The ACM ISBN 979-8-4007-2950-8/2026/11 is provided. The work is licensed under a Creative Commons Attribution 4.0 International License. The ACM ISBN 979-8-4007-2950-8/2026/11 is provided. The work is licensed under a Creative Commons Attribution 4.0 International License. The ACM ISBN 979-8-4007-2950-8/2026/11 is provided. The work is licensed under a Creative Commons Attribution 4.0 International License. The ACM ISBN 979-8-4007-2950-8/2026/11 is provided. The work is licensed under a Creative Commons Attribution 4.0 International License<ref:2609.

Improvements for AI systems

Improve cross-continental OD generation reliability by systematically using geographically grounded encoders like AlphaEarth or SatCLIP when census data is unavailable, as geographically grounded encoders significantly outperform language-aligned features on the UK evaluation dataset.
Implement a deployment-time configuration for the census noise parameter eta to avoid silent performance degradation in zero-census regions, ensuring practitioners use the appropriate regime based on local census availability by evaluating strictly at eta = 0 for non-US data.
Enhance model robustness against domain shift by incorporating multi-temporal satellite features, such as quarterly Sentinel-2 stacks, to capture seasonal land-use dynamics that are currently invisible to the static annual composites used in the current evaluation.
Increase scalability for large metropolitan graphs by upgrading the generative backbone from WeDAN to a sparse or flash attention mechanism within the GraphTransformer, which would allow processing significantly larger metropolitan graphs without memory bottlenecks.
Improve diagnostic utility by decomposing the Common Part of Commuters (CPC) metric by distance decay bin, flow magnitude quantile, or intra- versus inter-zone flows to expose encoder-specific failure modes that aggregate metrics obscure.

Abstract

Satellite foundation models offer a globally available alternative to census data for commuting origin-destination (OD) generation, yet no study has systematically compared encoder paradigms within a single downstream pipeline. We ablate four satellite vision encoders: language-supervised (RemoteCLIP), self-supervised (DINOv3), and geographically grounded (SatCLIP, AlphaEarth) within an identical WeDAN graph diffusion framework across 1,925 US counties, 325 UK districts, and 14 global cities under five random seeds. Three main findings emerge. First, language-supervised features achieve the strongest in-distribution performance (RemoteCLIP CPC 0.602), while geographically grounded encoders transfer more reliably zero-shot: AlphaEarth improves CPC by 33% over RemoteCLIP on UK districts. Second, pretraining corpus scale alone is insufficient: DINOv3, trained on a substantially larger satellite corpus, underperforms RemoteCLIP by 0.091 CPC in-distribution and collapses to CPC 0.022 globally. Third, no encoder transfers usefully to global cities (best CPC 0.122 for RemoteCLIP, 0.022 for DINOv3), confirming cross-continental OD generation remains an open problem. We additionally clarify the semantics of the census noise parameter η, whose ordering reverses under cross-continental evaluation, a distinction critical to correctly interpreting prior results. Our code is available at: https://github.com/cruiseresearchgroup/satellites-see-commuters.