Deep Learning for Cross-Domain Data Fusion in Urban Computing: Taxonomy, Advances, and Outlook

Motivation

• Urbanization intensifies sustainability challenges including energy consumption, pollution, socio-economic disparities, and traffic issues.Urban computing applies sensing technologies and computing infrastructure to analyze urban data and address such challenges.Source: §1 · p. 1

Related Surveys

• Earlier surveys focused on specific aspects, not extensively covering cross-modal fusion or newer contrast-based and generation-based approaches popular since 2021, and none provided an up-to-date urban-computing taxonomy.Other surveys covered GAN-based spatio-temporal fusion, self-supervised multimodal representation learning in the general domain, or specific urban applications.Source: §1 · p. 3

Survey Scope

• The survey reviews deep-learning-based cross-domain data fusion for urban computing and aims to document advances, resources, applications, and research directions.The authors position it as the first comprehensive, systematic survey focused on this topic.Source: §1 · p. 3

Taxonomy

• The taxonomy spans five urban data-source types, four fusion-method families, and applications including planning, transportation, economy, public safety, society, environment, and energy.The method families are feature-based, alignment-based, contrast-based, and generation-based.Source: §1 · p. 3

Dataset Compilation

• The survey compiles and categorizes datasets by source, temporal coverage, and spatial distribution, and reviews application scenarios, contributions, and limitations.It also examines practical contributions and limitations across downstream applications.Source: §1 · p. 3

Future Research Outlook

• The outlook covers privacy protection, open benchmarks, application diversification, efficiency, and potential urban-computing uses of large language models.The survey examines LLM applications and proposes potential approaches for urban computing.Source: §1 · p. 3

2. Taxonomy

• The survey structures cross-domain urban data fusion around data, fusion methods, and applications.Table 1 compares related surveys by scope, modalities, and topical coverage.Source: §2 · p. 3, Table 1 · p. 4
• Urban data is categorized by five sources and four modalities, with public datasets also surveyed.Sources are geographical, traffic, social media, demographic, and environmental; modalities include spatio-temporal, visual, textual, and other data.Source: §2 · pp. 3–4
• Fusion methods are grouped into feature-based, alignment-based, contrast-based, and generation-based categories, with recent urban LLM applications also reviewed.The survey further subdivides each category according to model properties.Source: §2 · p. 4
• Multimodal urban applications are classified into urban planning, transportation, economy, public safety and security, social, environment, and energy.The survey examines the superiority of multimodal fusion methods across these downstream task categories.Source: §2 · p. 4

3.1. Overview

• The survey groups datasets into six source classes—geographical, traffic, social network, demographic, environment, and other, including healthcare—and summarizes open datasets in Table 2.Source: §3.1 · p. 4, Table 2 · p. 5, Figure 3 · p. 6
• Approaching 70% of surveyed papers use geographical or transportation data as a primary modality; around 11% incorporate demographic and environment data.Source: §3.1 · p. 4, Figure 4 · p. 7
• Beijing and New York datasets are most frequently used, followed by Chicago, Singapore, and Shanghai; most surveyed datasets originate in China and the United States.Source: §3.1 · p. 4, Figure 5 · p. 7

3.2. Geographical Data

• Geographical data supports spatial modeling; POIs represent area semantics, satellite images depict Earth’s surface, and street-view images provide ground-level perspectives.Source: §3.2 · p. 7

3.3. Traffic Data

• Traffic data is the second-largest category and spans trajectories, flow, road networks, and miscellaneous data such as logistics, safety, and transportation recommendations.Source: §3.3 · pp. 7–8

3.4. Social Media Data

• Social-media inputs include geo-tagged tweets, photos, and videos, used for POI associations, mobility patterns, recommendations, and context-aware location learning.Source: §3.4 · p. 9

3.5. Demographic Data

• Population statistics, crime records, and land-use datasets support socioeconomic prediction, regional embeddings, and region-representation evaluation.Source: §3.5 · p. 9

3.6. Environment Data

• Environmental sources include meteorological data, satellite-derived tree coverage, and air quality, with hourly measurements from 2,296 stations in 302 Chinese cities.Source: §3.6 · pp. 9–10

• The section defines and explains four types of multimodal fusion.Source: §4 · p. 10
• Each fusion type is explored through more specific categorizations and illustrative examples.Source: §4 · p. 10
• Table 3 provides a comprehensive summary of fusion models.Source: §4 · p. 10

• Table 3 summarizes deep-learning-based cross-domain data fusion models in urban computing.Its notation identifies traffic, geographical, social media, demographic, and environmental data sources.Source: Table 3 · p. 11
• Feature-based fusion directly combines features using operations such as addition, multiplication, concatenation, or graph operations, with complexity typically linear in feature dimensions.The category includes sensor, visual, and textual features.Source: §4.1 · pp. 10, 12
• Alignment-based fusion transforms diverse representations into a shared feature space or structure for semantic consistency, with complexity typically quadratic in input sequence length.Attention can weight input parts; image-text alignment is one example.Source: §4.1 · p. 12
• Contrast-based fusion uses positive and negative sample pairs to improve sample-level feature discriminability, with complexity often quadratic in the number of samples.The survey distinguishes this sample-level comparison from alignment-based feature-level alignment.Source: §4.1 · p. 12
• Generation-based fusion generates one urban modality conditioned on the same or other modalities, using approaches including mask modeling, diffusion, and LLM-enhanced techniques.Generative-model training and optimization are described as highly complex, with complexity generally quadratic or higher.Source: §4.1 · p. 12

4.2. Feature-based Data Fusion

• Feature-based fusion is simple and efficient for real-time applications such as transportation, but its coarse integration may underperform when tasks require accurate alignment of diverse datasets.Source: §4.2 · p. 12

4.2.1. Feature Addition and Multiplication

• Elementwise addition sums corresponding modality features, while elementwise multiplication combines them coordinate by coordinate; examples include revenue prediction and citywide passenger-demand prediction.The cited examples use RoD, taxi, and POI features in product form, and passenger demand, meteorological data, and time metadata for a joint representation.Source: §4.2.1 · p. 12

4.2.2. Feature Concatenation

• Concatenation appends modality features to preserve their individual information in a fused representation for downstream tasks.Examples include crowd-flow forecasting with temporal flow maps and external factors, and visual fusion of satellite imagery, street-level imagery, and taxi trajectories.Source: §4.2.2 · pp. 12–13

4.2.3. Graph-based Data Fusion

• Multi-view graphs represent different relationships among shared nodes, such as distance, mobility, or semantics, and graph convolution can fuse these views for urban forecasting.Geng et al. represent 1 km × 1 km regions as nodes and use multiple graphs to encode pairwise regional correlations.Source: §4.2.3 · p. 13, Figure 11 · p. 14
• Because separate graph views lack direct cross-view node interaction, heterogeneous and compound multi-linear graphs connect urban entities or combine edge relations across views.Zhang et al. use compound random-walk paths that combine relationships from different views; SHNs are presented as a way to encapsulate complex cross-modal relationships.Source: §4.2.3 · p. 14, Figure 12 · p. 14

4.3. Alignment-based Data Fusion

• Alignment-based fusion aligns semantically related content through cross-modal attention or unified encoders, offering precise and flexible integration but requiring more computation.The survey notes this can make alignment-based methods more suitable for offline transportation and urban-planning tasks.Source: §4.3 · p. 14, Figure 13 · p. 15

4.3.1. Attention-based Alignment

• Cross-attention projects modality features into query, key, and value spaces, normalizes cross-modal scores into weights, then uses those weights to compute weighted value sums.Urban examples include attention-based fusion of road attributes and adaptive weighting of human-mobility and region-attribute views.Source: §4.3.1 · p. 15

4.3.2. Encoder-based Alignment

• Encoder-based fusion feeds multiple modalities into a shared encoder, encouraging it to learn inter-modal relationships and dependencies.Examples include DeepTransport using GPS and transportation data in a deep LSTM, and RegionDCL using a Transformer encoder for building features and POI information.Source: §4.3.2 · pp. 15–16

4.4. Contrast-based Data Fusion

• Contrast-based fusion uses instance or batch contrast to improve discrimination through negative samples, but requires careful negative selection and large batches or datasets.The survey describes it as suited to data-abundant downstream tasks, including transportation, environment, and economy.Source: §4.4 · p. 16
• InfoNCE maximizes similarity for positive pairs while minimizing it for negative pairs, and CLIP applies this principle to image-text representations.The loss uses K negative samples; the survey presents contrastive fusion in instance- and batch-based forms.Source: §4.4 · p. 16

4.4.1. Instance Contrast-based Fusion

• Instance contrast fuses knowledge across views of the same example, capturing fine-grained patterns through augmented pairs such as transformed images, POIs, roads, or trajectories.Examples include rotation, grayscale, and flipping for urban images, and insertion, deletion, or replacement for POIs.Source: §4.4.1 · p. 16

4.4.2. Batch Contrast-based Fusion

• Batch contrast adds global context by distinguishing samples within a batch, including geographically guided image comparisons and cross-modal or cross-view urban representations.UrbanCLIP pairs satellite imagery with LLM-generated descriptions; UrbanVLP extends visual clues across satellite and street-view images and addresses text quality.Source: §4.4.2 · pp. 16–17, Figure 16 · p. 17

4.5. Generation-based Data Fusion

• Generation-based fusion produces desired data from specified conditions, across input and output formats, but its computational and training demands make it unsuitable for real-time tasks.The survey notes its fit for large datasets and identifies autoregressive, mask-modeling, diffusion-based, and LLM-enhanced types.Source: §4.5 · p. 17

4.5.1. Autoregressive Model

• Autoregressive models predict future data from historical context, with multimodal attention enabling queries, keys, and values to draw on different modalities.Urban examples include GeoMAN’s LSTM and temporal-attention decoder for sensor prediction and image classification using visual and contextual information.Source: §4.5.1 · p. 17

4.5.2. Mask Modeling-based Fusion

• Mask-modeling fusion masks some or all input modalities and integrates information from other modalities to reconstruct the masked data.Urban examples use masked-language-modeling objectives for location embeddings, delivery-address robustness, travel search, and toponym knowledge.Source: §4.5.2 · p. 18, Figure 17 · p. 18

4.5.3. Di

• Diffusion-based fusion conditions generation on additional modalities; DiffusionSat combines geospatial metadata, timestamps, and text for satellite-image tasks.DiffSTG extends DDPM to spatio-temporal graphs, while DiffUFlow infers fine-grained flows from coarse inputs and DiffTraj generates trajectories through stepwise denoising.Source: §4.5.3 · p. 18, Figure 18 · p. 19

4.5.4. LLM-enhanced Data Fusion

• LLM-enhanced fusion uses LLMs for feature encoding and information interaction, with urban applications spanning geospatial knowledge, remote sensing, time series, and agents.The survey characterizes this as a specific instance of encoder-based alignment.Source: §4.5.4 · p. 19
• Time-LLM aligns time series with natural language by reprogramming series into text-prototype representations before feeding them to a frozen LLM.The figure caption identifies the text-prototype step as preceding input to the frozen model.Source: §4.5.4 · p. 19, Figure 20 · p. 20

5. Application Perspective

• The application taxonomy covers seven domains and their common downstream tasks, while urban planning uses multi-source deep learning to support city- and region-level work.The survey frames this perspective as a review of how fusion methods facilitate these applications.Source: §5 · p. 20, §5.1 · p. 20, Figure 21 · p. 20

5.1.1. City-level Planning

• City-level planning uses multi-source data to sense urban dynamics, including vibrancy, social events, and road-network change, with fusion methods supporting event prediction and road-attribute extraction.DuARE combines GPS trajectories and satellite images and is reported to update China’s national road network by 100,000 km monthly.Source: §5.1.1 · pp. 20–21

5.1.2. Region-level Planning

• Region-level planning combines POIs, imagery, mobility, and other sources for region detection and embeddings that support land-use inference and related planning tasks.Studies also address evolving natural or functional boundaries, including cross-city transfer and boundary sensing for urban villages and POIs.Source: §5.1.2 · pp. 21–22
• 91.6% overall accuracy was reported for identifying urban villages in Shenzhen using satellite, street-level imagery, and mobility data.The cited study’s ablation found the additional vision data crucial for region detection.Source: §5.1.2 · p. 21

5.2.1. Public Transportation Service

• Public transportation applications are grouped into safety, flow control, and efficiency, covering road risks, crowd and vehicle movement, and demand or delivery-time prediction.Examples combine trajectory with environmental sensing for obstacle detection and use multi-view spatial, temporal, and semantic features for taxi-demand prediction.Source: §5.2.1 · pp. 22–23
• Flow-control studies address traffic regulation and mobility prediction, using spatio-temporal or contrastive methods to model urban flows and movement patterns.The survey connects flow control with public policy, resource scheduling, and urban planning.Source: §5.2.1 · p. 23

5.2.2. Transportation Recommendation

• Transportation recommendations include route, multimodal, and trip planning, with multimodal systems combining travel modes and transition instructions for personalized journeys.One framework integrates time-dependent multi-view transportation graphs using hierarchical multi-task route representation learning.Source: §5.2.2 · pp. 23–24

5.3. Economy

• Urban-economy studies use POI density, satellite imagery, and mobility data to characterize regional conditions and estimate property prices or economic indicators.Other approaches combine intra- and inter-region geographic structure or add satellite imagery to real-estate price prediction.Source: §5.3 · p. 24
• 10% improvement in urban economic-indicator prediction was reported for combining satellite and street-view images, compared with studies before 2022.The survey also describes POI, satellite, and mobility data as indicators of regional economic conditions.Source: §5.3 · p. 24

5.4. Public Safety and Security

• Mobility, region attributes, and crime records support regional crime prediction, while vehicle trajectories and road-environment data help identify traffic-violation-prone locations [124].Traffic-safety studies also fuse taxi orders, points of interest, and weather to predict accidents, and use multi-view features for fine- and coarse-grained accident-risk prediction.Source: §5.4 · p. 25
• Traffic trajectories, road networks, satellite imagery, meteorology, earthquake records, and text reports support road-risk analysis and post-disaster mobility modeling [188].DeepMob combines trajectories, earthquake records, text reports, and transportation-network data to analyze and predict behavior and mobility after disasters.Source: §5.4 · p. 25

5.5. Social

• For location-based social-network recommendations, multi-graph models represent user–POI or friend connections, with contrastive learning used to embed spatio-temporal trajectory features.The passage links these methods to the non-Euclidean spatial structure of geolocation data.Source: §5.5 · p. 25

5.6. Environment

• Multi-source air-quality prediction fuses station history with meteorological, forecast, and other-station data; graph neural networks model spatio-temporal relations for city-level weather forecasting.The air-quality model predicts the next 48 hours at a single station point.Source: §5.6 · p. 25

5.7.1. Energy E

• Urban energy fusion informs planning and transport decisions; charging-station siting models use taxi GPS, station data, and vehicle-range information, with waiting time considered in one approach.The passage states that considering waiting time could reduce energy used for charging.Source: §5.7 · pp. 25–26, §5.7.1 · p. 26

5.7.2. Energy E

• Fuel-efficient route recommendation combines trajectory timing and driving factors in a transformer with a genetic algorithm, while DeepFEC fuses vehicle context and road traffic data to predict road-level energy consumption.DeepFEC uses a deep convolutional residual model for spatial dynamics and a Bi-LSTM for temporal information.Source: §5.7.2 · p. 26
• A multi-source personal-car trajectory dataset with user information supports meta-learning vehicle-energy prediction that can be fine-tuned to a driver’s preference [209].The cited application is in Ann Arbor, Michigan.Source: §5.7.2 · p. 26

LLM-enhanced Application

• LLM use in urban fusion remains early: models struggle with spatial relationships and orientations, and most public LLMs are text-based with limited performance in other modalities.The survey notes emerging graph-learning work and multimodal LLMs as promising directions.Source: §6 · pp. 26–27

Agent-based Simulation

• Earlier agent-based urban simulations rely on human-defined rules or goals; LLM-driven agents such as Urban Generative Intelligence are proposed for modeling urban dynamics from cross-domain data [317].The passage frames this as a potential approach to simulating urban systems’ complex dynamics.Source: §6 · p. 27

Multi-modal Causal Learning

• Urban cross-domain causal learning is still early, with complex cross-modal causality and the lack of a universally accepted accurate and efficient representation approach identified as challenges.The authors anticipate further work on multimodal causal learning to improve interpretability of urban systems.Source: §6 · p. 27

Multi-source Data Privacy

• Multi-source urban data can expose sensitive information through model memorization, while institutions’ and governments’ reluctance to share data impedes deployment; differential privacy and federated learning are proposed directions.The stated challenge is protecting privacy while retaining the benefits of cross-domain fusion.Source: §6 · p. 27

Open Benchmark

• An open benchmark must address heterogeneous formats and modalities through standardized data formats, unified evaluation metrics, and shared benchmarks for model comparison.Source: §6 · p. 27

Downstream Task Diversity

• Research concentrates on transportation and urban planning, while the authors call for broader coverage of economy, society, and environment applications.They connect broader task coverage with understanding diverse conditions in urban settings.Source: §6 · p. 27

Computation Efficiency

• Current urban research often overlooks computational efficiency, which the authors say hampers real-world deployment; they propose compression, efficient training, and efficient architectures as directions.Examples include distillation, low-rank decomposition, quantization, prompt tuning, hardware-assisted attention acceleration, and mixture-of-experts architectures.Source: §6 · p. 27

7. Conclusion

• The survey reviews cross-domain fusion for urban computing, examining data sources and modalities and classifying methods as feature-based, alignment-based, contrast-based, or generation-based.Source: §7 · p. 27
• It categorizes urban applications into seven types and outlines challenges and promising future research directions.Source: §7 · p. 28

Appendix A. Github Link

• The appendix provides a link to the comprehensive and up-to-date paper list.Source: Appendix A · p. 28