CoST: Semantic-Aware Urban Understanding via Spatial-Temporal Alignment

Yutian Jiang; Jiabo Liu; Xixuan Hao; Yuxuan Liang

cs.CVAug 21, 2026

CoST: Semantic-Aware Urban Understanding via Spatial-Temporal Alignment

Authors: Yutian Jiang, Jiabo Liu, Xixuan Hao, Yuxuan Liang

Organizations: The Hong Kong University of Science and Technology (Guangzhou)

Abstract

Geospatial representation learning from satellite imagery is a fundamental problem for large-scale urban analysis and real-world applications. Despite recent advances, current methods struggle with cross-region generalization and semantic interpretability due to their reliance on region-specific auxiliary data and the neglect of semantic alignment within multi-temporal urban imagery. Therefore, we present CoST, a novel Contrastive-based Spatial-Temporal framework that aligns spatial context with multi-temporal semantics to extract universal geographic regularities shared across regions. Specifically, CoST explicitly models spatial correlations to capture transferable geographic structures and exploits multi-year urban change semantics to align learned representations with high-level geo-semantics. Extensive experiments demonstrate that CoST consistently achieves superior performance across various downstream tasks and in unseen scenario, yielding an average relative gain of 8.7% over the strongest competing methods across eight city-indicator settings. The code is available in this repo.

Figures & tables

Figure 1. Geospatial representation and its challenges.

Figure 2. Overall framework of the proposed CoST, aligning geographical neighborhood structures with multi-year urban change semantics to learn generalizable and interpretable geospatial representations for various urban analysis tasks.

Figure 3. Joint modeling spatial and temporal dimensions from the perspectives of spatial proximity and temporal heterogeneity.

Figure 4. Overall performance of CoST-R across various downstream tasks.

Dataset	Coverage in Geospace		Satellite
Dataset	Bottom-left	Top-right	Image
Beijing	39.91°N, 116.20°E	40.24°N, 116.53°E	4,218 $\times$ 11
Guangzhou	22.78°N, 113.15°E	23.36°N, 113.51°E	3,726 $\times$ 11
Shanghai	30.87°N, 121.24°E	31.40°N, 121.68°E	3848 $\times$ 11
Shenzhen	22.52°N, 113.84°E	22.67°N, 114.25°E	2,772 $\times$ 11
New York	40.52°N, 73.95°W	40.80°N, 73.79°W	1,728 $\times$ 11

Table 1. Dataset statistics. The first five datasets are used for pre-training, and the Chicago dataset is held out from pre-training for a fair generalization test.

Model	Beijing				Guangzhou				New York				Chicago
	GDP		Population		GDP		Population		GDP		Population		GDP		Population
	R 2	MSE $\downarrow$	R 2	MSE $\downarrow$	R 2	MSE $\downarrow$	R 2	MSE $\downarrow$	R 2	MSE $\downarrow$	R 2	MSE $\downarrow$	R 2	MSE $\downarrow$	R 2	MSE $\downarrow$
SimCLR	0.420	0.584	0.412	0.570	0.248	0.684	0.163	0.467	0.552	0.363	0.608	0.353	0.501	0.561	0.467	0.522
MoCoV3	0.488	0.460	0.504	0.510	0.293	0.781	0.390	0.447	0.535	0.434	0.594	0.391	0.570	0.367	0.597	0.447
DINOv2	0.346	0.605	0.456	0.544	0.151	0.845	0.296	0.771	0.593	0.435	0.622	0.327	0.487	0.499	0.661	0.357
SeCo	0.463	0.551	0.462	0.497	0.320	0.728	0.472	0.491	0.516	0.502	0.606	0.375	0.583	0.342	0.531	0.415

Table 2. Performance on urban indicator prediction (2020). Bold and underline denote the best and second-best results, respectively. Both variants of our model (CoST-R and CoST-V) are highlighted in bold when they achieve the top two positions.

Methods	Classification		Change Detection
	UCM	BigEarthNet	OSCD	LEVIR-CD
	top-1 Acc	mAP	F1 Score	F1 Score
SimCLR	75.47	71.62	0.5069	0.3944
MoCoV3	90.76	80.30	0.5315	0.6220
SeCo	92.99	81.15	0.4953	0.6192
CACo	86.94	77.81	0.4429	0.4881

Table 3. Performance of different models in classification and change detection tasks. The best results are bold , and the second-best results are underlined .

Figure 5. Visualization of embedding arithmetic. It shows that CoST captures high-level semantic concepts via vector operations. Predictable semantic transformations demonstrate that the learned latent space is disentangled and interpretable.

Figure 6. Ablation study of three CoST components: (a) w/o Spatial , (b) w/o Temporal , and (c) w/o ST .

Figure 7. Urban indicator prediction using time-series satellite imagery in Beijing, Guangzhou, and New York City. Sta and Temp denote static and dynamic prediction, respectively.

Figure 8. Domain adaptation of CoST between the source city and target city (Chicago). Color saturation represents GDP levels from light (low GDP) to dark (high GDP).

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Model	Backbone	GDP		Population
Model	Backbone	R 2	MSE $\downarrow$	R 2	MSE $\downarrow$
SimCLR	R50	0.213	0.955	0.284	0.711
MoCoV3	R50	0.283	0.523	0.379	0.443
DINOv2	R50	0.179	0.692	0.293	0.736
SeCo	R50	0.302	0.542	0.298	0.919
CACo	R50	0.362	0.860	0.305	0.734

Appendix

Table 4. Performance Comparison on Shanghai Dataset. Bold and underline denote the best and second-best results, respectively. Both variants of our model (CoST-R and CoST-V) are highlighted in bold when they achieve the top two positions.

Model	Backbone	GDP		Population
Model	Backbone	R 2	MSE $\downarrow$	R 2	MSE $\downarrow$
SimCLR	R50	0.099	0.832	0.296	0.623
MoCoV3	R50	0.155	0.998	0.382	0.569
DINOv2	R50	0.074	0.943	0.316	0.709
SeCo	R50	0.102	0.894	0.274	0.653
CACo	R50	0.086	0.959	0.291	0.578

Appendix

Table 5. Performance Comparison on Shenzhen Dataset. Bold and underline denote the best and second-best results, respectively. Both variants of our model (CoST-R and CoST-V) are highlighted in bold when they achieve the top two positions.

Figure 9. Performance comparison with RS foundation models across indicator predictions. CoST-R, Prithvi-100M, and DOFA are compared using GDP MSE and population MSE.

Geospatial Representation Learning

Geospatial representation learning builds reusable embeddings of places, imagery, maps, mobility, and Earth data for spatial prediction and reasoning.

See all 126 papers

UrbanVerse: Learning Urban Region Representation Across Cities and Tasks
Feb 17, 2026Fengze Sun, Egemen Tanin, Shanika Karunasekera +3Geospatial Representation Learning Spatial Generalization
School of Computing and Information System, University of Melbourne, Melbourne, Australia · School of Computer Science and Engineering, University of New South Wales, Sydney, Australia
UNIGEOCLIP: Unified Geospatial Contrastive Learning
Apr 13, 2026Guillaume Astruc, Eduard Trulls, Jan Hosang +2Geospatial Representation Learning Multimodal Embedding
LASTIG, Univ Gustave Eiffel, IGN, ENSG, France · CNES, France · LIGM, CNRS, Univ Gustave Eiffel, ENPC, Institut Polytechnique de Paris, Marne-la-Vallée, France +1
Geo-Embed: Towards Unified Multimodal Embeddings for Urban Understanding
Aug 4, 2026Jiapeng Li, Yong Li, Junjie Zhou +2Geospatial Representation Learning Cross-Modal Retrieval
1Peking University · 2Beijing University of Posts and Telecommunications

Feb 17, 2026J/K move · Enter open · S save

UrbanVerse: Learning Urban Region Representation Across Cities and Tasks

Fengze Sun, Egemen Tanin, Shanika Karunasekera, Zuqing Li+2

School of Computing and Information System, University of Melbourne, Melbourne, Australia · School of Computer Science and Engineering, University of New South Wales, Sydney, Australia

Geospatial Representation Learning Spatial Generalization Multi-Task Learning Graph Representation Learning

Open paper AskPDF arXiv

Recent advances in urban region representation learning have enabled a wide range of applications in urban analytics, yet existing methods remain limited in their capabilities to generalize across cities and analytic tasks. We aim to generalize urban representation learning beyond city- and task-specific settings, towards a foundation-style model for urban analytics. To this end, we propose UrbanVerse, a model for cross-city urban representation learning and cross-task urban analytics. For cross-city generalization, UrbanVerse focuses on features local to the target regions and structural features of the nearby regions rather than the entire city. We model regions as nodes on a graph, which enables a random walk-based procedure to form "sequences of regions" that reflect both local and neighborhood structural features for urban region representation learning. For cross-task generalization, we propose a cross-task learning module named HCondDiffCT. This module integrates region-conditioned prior knowledge and task-conditioned semantics into the diffusion process to jointly model multiple downstream urban prediction tasks. HCondDiffCT is generic. It can also be integrated with existing urban representation learning models to enhance their downstream task effectiveness. Experiments on real-world datasets show that UrbanVerse consistently outperforms state-of-the-art methods across six tasks under cross-city settings, achieving up to 35.89% improvements in prediction accuracy.

Abstract

Geospatial Representation Learning

Explore similar work

UrbanVerse: Learning Urban Region Representation Across Cities and Tasks

UNIGEOCLIP: Unified Geospatial Contrastive Learning

Geo-Embed: Towards Unified Multimodal Embeddings for Urban Understanding