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

Ashiq Shukoor Iqbal; Wilson Wongso; Flora D. Salim

Source-linked AI summary

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

Ashiq Shukoor Iqbal, Wilson Wongso, Flora D. Salim

arXiv:2609.00661v1Sep 1, 2026cs.CV

TL;DR

The paper addresses the lack of systematic evidence on which satellite vision encoder best supports commuting OD generation and cross-continental transfer. It embeds four encoder paradigms in the same WeDAN graph diffusion pipeline and evaluates them across US, UK, and global settings. RemoteCLIP leads in-distribution, geographically grounded encoders transfer better to the UK, and no encoder transfers usefully to global cities.

Problem
Existing OD-generation work lacks a systematic comparison of satellite vision encoders within one downstream pipeline, despite cross-continental transfer remaining unaddressed.
Method
The paper evaluates four vision encoder paradigms in an identical WeDAN graph diffusion framework across 1,925 US counties, 325 UK districts, and 14 global cities using five seeds.
Results
RemoteCLIP leads US performance at CPC 0.602, while AlphaEarth leads UK zero-shot transfer at CPC 0.515, a 33% relative improvement over RemoteCLIP at η=0.
Takeaways & Limitations
Language supervision is strongest in-distribution, whereas geographic and physical grounding provides greater resilience for cross-continental zero-shot transfer.
Takeaways & Limitations
Static annual satellite composites cannot capture temporal variation in commuting flows such as rush-hour congestion, seasonal employment, and post-pandemic transit changes.

Abstract

from arXiv · show

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. Training scripts and evaluation logs will be released.

1 Introduction

OD generation can reduce dependence on costly, infrequently updated, or restricted mobility data, but existing work has not systematically compared satellite encoder paradigms in a fixed pipeline. This paper evaluates four encoders across domestic and zero-shot geographic settings.

Motivation: Traditional OD data sources face financial, operational, update-frequency, scalability, privacy, and commercial-access constraints.The motivation contrasts census registries, household surveys, and proprietary telecommunication logs.
Prior satellite-conditioned generation: GlODGen uses high-resolution satellite imagery and globally uniform WorldPop statistics instead of downstream tabular census parameters.Its regional tiles are processed through RemoteCLIP before graph-based OD generation.
Research gap: Existing frameworks treat RemoteCLIP as fixed, despite alternative encoders using self-supervised, coordinate-contrastive, and multisensor pretraining objectives.These encoders also differ in representation dimension and preprocessing.
Research gap: The literature lacks a systematic comparison of vision encoders within one generative graph pipeline, leaving the necessity of scene-level text-aligned features unresolved.The paper frames encoder choice as an open empirical question.
Study design: The study evaluates four encoder paradigms with the same WeDAN pipeline across 1,925 US counties, 325 UK districts, and 14 global cities using five random seeds.Matched-seed paired t-tests support the evaluation framework.
Contributions: The contributions include controlled encoder ablation, clarification of census-noise semantics, analysis of pretraining-scale tradeoffs, and geographically grounded transfer gains.The introduction reports DINOv3 CPC = 0.511 versus RemoteCLIP CPC = 0.602 and stronger UK transfer for AlphaEarth.

2 Background and Related Work

OD modeling progressed from parametric spatial-interaction models to pairwise deep learning, graph neural networks, and graph diffusion. The remaining challenge is reliable zero-shot transfer of satellite-conditioned generative mobility models across continents.

Classical models: Gravity and Radiation models impose parametric assumptions based on population, distance, job opportunities, or intervening opportunities.These approaches provide classical baselines for mobility-flow generation.
Pairwise deep learning: DeepGravity and GMEL map neighborhood features to decoupled origin-destination pairs, but pairwise optimization overlooks networkwide transit dynamics.This limitation motivated graph-based approaches.
Graph neural networks: ODCRN represents urban environments as cohesive topological structures and captures spatial constraints through graph message passing.This marks the shift from independent pairs to structured networks.
Graph diffusion: WeDAN generates complete OD matrices through reverse denoising of continuously Gaussian-corrupted flow graphs, establishing CPC ≈0.59 on the full 3,233-area LODES benchmark.Its diffusion formulation captures local anomalies and macro-scale spatial structure.
Satellite-conditioned generation: GlODGen conditions graph diffusion on satellite-derived node embeddings to reduce reliance on localized census registries, while retaining RemoteCLIP as its backbone.This benchmark directly tests whether that fixed backbone is necessary.
Vision foundation models: Remote sensing foundation models differ in representational properties because they optimize distinct pretraining paradigms.Table 1 organizes the evaluated architectural profiles.
Zero-shot transfer: Cross-continental generative mobility transfer remains structurally unaddressed because built form, street networks, population density, transit, and zoning vary sharply across regions.North American developments and European historic grids illustrate the challenge.
Zero-shot transfer: The paper presents a first systematic US-to-UK and US-to-global benchmark for how shifting visual and structural topologies affect satellite-conditioned graph diffusion.Domestic training can cause conditioning layers to overfit regional features.

3 Problem Formulation & Architecture

The architecture treats OD generation as conditional graph diffusion: satellite and population features condition a GraphTransformer that denoises flow matrices. Census availability is controlled separately through η, which governs census conditioning rather than satellite-feature noise.

3.1 OD Generation as Conditional Graph Diffusion: A city is modeled as a directed graph whose nodes are geographic regions and whose edges are commuting flows in the target OD matrix.The matrix is generated as a DDPM and sampled through denoising at inference.
3.1 OD Generation as Conditional Graph Diffusion: The forward process adds Gaussian noise across T timesteps, while the reverse GraphTransformer removes noise conditioned on node-level socio-environmental and structural features.Conditioning guides predicted mobility flows toward region-specific characteristics.
3.1 OD Generation as Conditional Graph Diffusion: Table 1 compares the architectural and pretraining profiles of the four evaluated spatial vision foundation models.The benchmark isolates representation-space differences within the downstream pipeline.
3.2 Satellite-Conditioned Node Features: The base regional feature combines mean-pooled spatial vision embeddings with log-transformed WorldPop population and land-area statistics.H3 resolution 9 tiles are processed independently before regional pooling.
3.2 Satellite-Conditioned Node Features: Training fuses satellite features with a noise-controlled localized census vector, where the true census input has 97 dimensions and Gaussian noise is explicitly defined.The full node vector also includes a predicted census representation.
3.2 Satellite-Conditioned Node Features: Setting if-ImgAttrAug=0 zeros the predicted census vector, so the effective GraphTransformer input excludes that augmentation component.The downstream input dimension therefore depends on the vision encoder dimension.
3.3 GraphTransformer Denoising Mechanics and Training Loss: GraphTransformer layers use multi-head attention with learned query, key, value, and edge-projection components to update node states and process flow edges.The denoising objective trains the network to predict injected Gaussian noise.
3.3 GraphTransformer Denoising Mechanics and Training Loss: DDIM inference uses a deterministic 25-step trajectory, averages 50 runs, and anchors generated flows to fixed satellite and demographic conditioning.The census parameter η is resampled during training to represent varying census availability, not satellite-feature noise.

4 Experimental Setup

The benchmark tests in-distribution and zero-shot generalization across US, UK, and global geographic scales using CPC alongside error metrics. All encoder variants share standardized spatial preprocessing, downstream diffusion settings, and repeated-seed inference.

Datasets: The evaluation spans 1,925 US counties, 325 UK districts, and 14 global metropolitan areas to stress-test localized performance and zero-shot transfer.The geographic scales are designed to probe increasingly heterogeneous settings.
Metrics: CPC measures spatial allocation accuracy by comparing generated and ground-truth OD flow matrices.It is the primary metric across experimental boundaries.
Metrics: CPC is supplemented by RMSE, NRMSE, and MAE to assess numeric scaling behavior.CPC focuses on spatial overlap, while the error metrics track arithmetic discrepancies.
Preprocessing: Satellite imagery is retrieved at zoom level 15 and partitioned into uniform H3 resolution 9 hexagons averaging approximately 0.1 square kilometers.This standardizes sampling across irregular administrative boundaries.
Encoder processing: Each encoder processes tiles independently, then mean-pools their latent embeddings into one regional visual representation.The four encoders produce representations with different internal dimensions and properties.
Implementation: All variants use the same WeDAN GraphTransformer and cosine diffusion configuration, including four attention layers, hidden dimension 32, and 250 diffusion timesteps.The shared downstream setup isolates the effect of encoder choice.
Implementation: Inference uses deterministic DDIM sampling over a compressed 25-step trajectory and averages 50 independent runs for each city.Training, validation, and evaluation are repeated over five random seeds.

5 Results and Analysis

The benchmark shows encoder performance depends strongly on evaluation region and census availability: RemoteCLIP leads in-distribution, geographically grounded encoders transfer better to the UK, and all encoders perform poorly globally. Census-noise behavior reverses across domains, while computational constraints and backbone limitations complicate deployment.

US In-Distribution Performance: 0.598 ± 0.008 CPC is RemoteCLIP’s highest overall mean across five seeds and all η levels.At η = 1 in the US, RemoteCLIP reaches CPC 0.602 ± 0.009 and recovers 94.9% of the full-census WeDAN baseline.
US In-Distribution Performance: 493 million pretraining images do not prevent DINOv3 from underperforming when its self-distillation objective misaligns with zone-level commuting prediction.Mean-pooling tile-level geometric features can discard the zone-level mobility signal required by the downstream task.
UK Zero-Shot Transfer: 0.515 ± 0.015 CPC makes AlphaEarth the strongest UK encoder at η = 0, narrowly ahead of SatCLIP at 0.513 ± 0.034.Across all η levels, SatCLIP and AlphaEarth average 0.495 ± 0.032 and 0.455 ± 0.118, respectively, because AlphaEarth degrades more sharply as η approaches 1.
UK Zero-Shot Transfer: 0.329 ± 0.060 is DINOv3’s UK aggregate CPC, and its performance remains insensitive to η within seeds.This pattern indicates minimal mobility signal from its geometric features under out-of-distribution deployment.
Global Zero-Shot Generalization: 0.122 ± 0.014 is the best Global CPC, achieved by RemoteCLIP, while DINOv3 reaches only 0.022 ± 0.006.RemoteCLIP, AlphaEarth, and SatCLIP differ by less than 0.025 CPC globally, indicating encoder choice is secondary at this transfer distance.
Census Noise Sensitivity and Operational Implications: η measures census availability: US performance improves as η increases, whereas UK and Global performance degrades toward η = 1.At η = 1, zero-valued census vectors create an out-of-distribution input; deployment in census-absent regions should therefore use η = 0.
Practical Encoder Selection: AlphaEarth requires no local GPU because feature extraction runs in Google Earth Engine, while SatCLIP uses roughly one-third the VRAM of ViT-L models.RemoteCLIP and DINOv3 have identical ViT-L extraction costs, so their performance gap reflects pretraining objective rather than model capacity.
Practical Encoder Selection: No single encoder dominates every deployment setting; region, census availability, and hardware constraints determine the preferred choice.The experiments fix WeDAN as the generative backbone, whose full attention scales quadratically and restricts deployment to cities with fewer than approximately 1,000 zones.

6 Limitations and Future Work

The evaluation is constrained by static inputs, dataset splitting and domain shift, fixed backbone scalability, reproducibility dependencies, and aggregate metrics. Future work targets richer temporal and traffic signals, scalable architectures, adaptation, and uncertainty reporting.

Limitations: Static annual satellite composites cannot capture temporal variation in commuting flows, including rush-hour congestion, seasonal employment shifts, and post-pandemic transit changes.
Limitations: A reduced 80/10/10 split introduces a systematic 2–3% CPC deficit, while North American training data leaves substantial UK and global domain shifts.RemoteCLIP’s CPC drops 37% from the US to the UK, from 0.602 to 0.388.
Limitations: SatCLIP’s optical-patch design limits exploitation of multi-spectral geophysical signals that AlphaEarth captures through multi-sensor fusion.
Limitations: DINOv3’s large pretraining corpus does not ensure task alignment because self-distillation emphasizes patch-level texture discrimination over zone-level functional land use.
Limitations: Fixing WeDAN constrains conclusions about encoder quality; batch size one limits cross-city learning, and quadratic attention restricts deployment to cities below approximately 1,000 zones.Alternative score-based diffusion and flow-matching backbones remain to be investigated.
Future Work: Future work should add traffic or multi-temporal satellite features, scale attention, fine-tune embeddings, adapt with few-shot target data, and report diffusion-based uncertainty.These directions include traffic indicators, temporal encoders, sparse or flash attention, LoRA, meta-learning, and variance across 50 DDIM draws.

7 Conclusion

Across US counties, UK districts, and global cities, encoder choice determines whether satellite-conditioned OD generation performs in-domain or transfers across continents.

RemoteCLIP leads US county performance with a CPC of 0.602 but generalizes fragiley across continents.Its cross-continental weakness is associated with morphological distribution shifts.
AlphaEarth leads UK zero-shot transfer with a peak CPC of 0.515 at η=0, improving 33% over RemoteCLIP.Geographical and physical grounding provides greater resilience for cross-continental transfer.
DINOv3 shows that massive pretraining dataset volume alone cannot compensate for insufficient task-specific domain alignment.Its empirical limitations include near-zero performance on global cities.
Coordinate-contrastive and geophysical structural priors are identified as promising directions for geographically transferable urban mobility generation.

A Complete Per-Region Results

The results are organized as performance metrics by census noise parameter η for US, UK, and global evaluations.

Table 6 reports US performance metrics by η level, averaged over seeds.
Table 7 reports UK performance metrics by η level, averaged over seeds.
Table 8 reports global performance metrics by η level and notes that n=4 seeds were used, with Table 3 referenced.