Meiliu Wu

The integration of citizen science, volunteered geographic information (VGI), and Web/mobile geographic information systems (GIS) has demonstrated significant potential in enhancing disaster response efforts. However, delivering timely, comprehensive and trustworthy information remains a major challenge, particularly when relying on passive data collection from social media. While researchers have developed specialized platforms for natural hazards and advanced models for data analysis, few studies present a holistic lifecycle from stakeholder-oriented design through development, especially with attention to the design phase. To address this gap, this paper introduces an agile and iterative user-centered framework for designing and developing a participatory mobile GIS application for collecting reliable, first-hand observations. A pilot study conducted during real-world hurricane events demonstrated the application’s ability to operate both in real time and offline, enabling the collection of precise geotagged data, categorized labels, and diverse media formats. The results highlight the potential of this active, stakeholder-centered approach to support intelligent disaster response strategies and complement passive and authoritative data sources. This paper advances the integration of citizen science and mobile GIS by providing a framework that follows user-centered design principles to inform future disaster response applications.

Previous studies have attempted to define segregation succinctly but have not reached a consensus. Adding to the unsettled debates was the shift from using place-based to person-based data in segregation studies. Deduced from the literature, the meanings of segregation may be summarized by the two essences of measuring the distribution of and interaction among different population groups, and measures may be characterized by four elements of segregation: dimensions, spatial extent, social extent, and data type. Besides showing a trend from using spatially aggregated place-based data to individual-level person-based data, a review of selected papers confirms the salience of distribution and interaction in measuring segregation and highlights the limitations of existing measures in capturing social interaction between population groups and the impacts of segregation, even in recent studies that utilize person-based data. Finally, we propose a segregation measurement framework to accommodate the increasingly hybrid physical-virtual world using person-based data with three components: distributions of different groups, interactions among groups, and inequality experienced by different groups due to the disparities in accessing resources and opportunities. This review paper charts directions for future studies using segregation measures.

Embedding spatial intelligence into vision-language models (VLMs) has offered a promising avenue to improve geospatial decision-making in complex urban environments. In this work, we propose a novel framework that augments the architecture of Contrastive Language-Image Pretraining (CLIP) with the techniques of spatial-context aware prompt engineering and spatially explicit contrastive learning. By leveraging a diverse set of geospatial imagery (e.g., street view, satellite, and map tile images), paired with contextual geospatial text generated and curated via GPT-4, our approach constructs robust multimodal representations that capture visual, textual, and spatial insights. The proposed model, termed GeospatialCLIP, is specifically evaluated for urban mixed land use detection, a critical task for sustainable urban planning and smart city development. Results demonstrate that GeospatialCLIP consistently outperforms traditional vision-based few-shot models (e.g., ResNet-152, Vision Transformers) and exhibits competitive performance with state-of-the-art models such as GPT-4. Notably, the incorporation of spatial prompts, especially those providing city-specific cues, significantly boosts detection accuracy. Our findings highlight the pivotal role of spatial intelligence in refining VLM performance and provide novel insights into the integration of geospatial reasoning within multimodal learning. Overall, this work establishes a foundation for future spatially explicit AI development and applications, paving the way for more comprehensive and interpretable models in urban analytics and beyond.

Although Vision-Language Models (VLMs) have demonstrated great potential, limited integration of spatial context via textual input has constrained their performance in geospatial analysis, particularly for location-based services (LBS). To this end, we propose Spatial-context Prompt Tuning, tailored to global image geo-localization tasks. Six spatial-context dimensions (i.e., geospatial image types, geo-localization clues, spatial patterns, land use/land cover, urban perception, and urban development) are designed to create Visual Question Answering (VQA) prompts with GPT-4, based on three imagery types (street view images, satellite images, map tiles) sampled from 790 populous cities. Next, we evaluate the efficacy of these dimensions based on a leading open-source VLM – Contrastive Language-Image Pre-training (CLIP). Results demonstrate consistent, task-dependent accuracy improvements in image geo-localization: land-use prompts improve city-level accuracy by 4.5%, and image-type prompts increase country-level accuracy by 6.7% and continent-level accuracy by 3.0%. Our key contributions include: (1) revealing how spatial versus non-spatial context affects prompt-tuned VLM performance, (2) designing reusabGle six dimensions of spatial context that support explainable, context-aware VLMs for geospatial applications, and (3) enhancing geo-localization accuracy across heterogeneous geospatial imagery. This work charts a positive direction for Geospatial Artificial Intelligence (GeoAI)-empowered research, enabling more effective and interpretable VLM applications in geospatial domains.

Understanding urban perception enhances urban intelligence research, aiding sustainable development and smart city. Urban Visual-Spatial Intelligence (UVSI) integrates human and sensor perception. This Perspective explores UVSI’s potential for sustainable development, identifies research gaps, and outlines future priorities. Advances in AI, high-performance computing, real-time data processing, and citizen science could significantly impact UVSI. Key research areas include spatiotemporal data integration, visual analytics, ethical frameworks, and inclusive methodologies for AI-driven urban management.

The assessment of urban wealthiness is fundamental to effective urban planning and development. However, conventional methodologies often rely on aggregated datasets, such as census data, with a coarse-grained resolution at the census tract level, impeding accurate evaluation of wealthiness in individual neighborhoods and failing to capture spatial heterogeneity. This study proposes a novel approach to predict urban wealthiness at a point-scale spatial resolution by utilizing geo-tagged street view images as input for deep learning models, thereby simulating human perception of urban built environments. Using the Place Pulse 2.0 dataset, which contains over 1.2 million pairwise comparisons of 110,988 street view images from 56 cities worldwide for different urban environment perception factors (e.g., safety and wealthiness), we developed deep learning models based on the Swin Transformer and Multi-gate Mixture-of-Experts (MMOE), a multi-task learning architecture. These models extract and integrate visual features of surrounding elements, including buildings, parks, and vehicles, to classify the wealthiness of specific geo-locations into three categories: Impoverished, Middle, and Affluent. To enhance model training and ground truth data, we modified and enhanced the TrueSkill Rating System, used for scoring neighborhoods via pairwise street view image comparisons, by considering temporal decay and spatial autocorrelation factors. These modifications improved the normality of wealthiness score distribution, reducing the standard deviation from 5.385 to 4.302 and skewness from −0.055 to −0.024. Consequently, model performance improved consistently, with accuracy increases observed in Swin Transformer (63 % to 68 %), ViT (54 % to 58 %), and ResNet50 (51 % to 56 %). In addition, proposed MMOE model demonstrates a significant improvement in the differentiation and classification of wealth categories within a three-class classification system (Impoverished, Middle, Affluent). It achieves an overall accuracy of 82 %, outperforming baseline models, Swin Transformer, ViT, and ResNet50, by 14 %, 24 %, and 26 % respectively. Additionally, we compared our model’s predictions with average household income data at the census block group level to elucidate its strengths and limitations. Experimental results demonstrated the efficacy of using geo-tagged street view images for predicting urban wealthiness across diverse geographic and environmental contexts. Our findings also highlight the importance of integrating both quantitative and qualitative evaluations in the prediction of urban environmental factors. By synthesizing human perceptions with advanced deep learning techniques, our approach offers a nuanced understanding of urban wealthiness, providing valuable insights for urban planning and development strategies.

Foundation Models (FMs) have demonstrated significant potential for geospatial analysis. As geospatial big data (e.g., geo-tagged text and images) continues to grow, FMs have evolved into geospatial foundation models (GeoFMs), the FMs capable of geospatial reasoning. However, two critical tasks remain challenging: (1) how to align various geospatial modalities for model training, and (2) how to enhance FMs’ geospatial context awareness. To fill these gaps, this work develops CLIP4Geo, a unified GeoFM that integrates satellite imagery, LiDAR point clouds, geo-tagged points of interest (POIs), and textual descriptions into a shared representation space. Specifically, to empower the model with geospatial intelligence, we design a Geospatial Expert Knowledge System that tailors both textual and visual information and encodes them into shared geospatial embeddings. Furthermore, we propose a novel Geospatial Context Awareness Transformer to guide the cross-attention mechanism, which leverages geospatial context position embedding to effectively capture spatial dependencies across modalities. We also introduce CityVerse, a large-scale, spatially aligned, and multimodal benchmark dataset covering diverse urban regions in the U.S. and China, supporting a range of geospatial downstream tasks (e.g., zero-shot classification, semantic segmentation / localization, and cross-modality retrieval and generation). Extensive experiments demonstrate that CLIP4Geo consistently outperforms state-of-the-art FM baselines, highlighting the importance of expert-guided multimodal alignment and geospatial context embeddings. Our work provides a scalable framework and comprehensive benchmark to facilitate the future development of GeoFMs for various geospatial applications. Code and dataset are available at https://github.com/Ting-Devin-Han/CLIP4Geo.git.

Street-view imagery (SVI) offers a fine-grained lens on traffic risk, yet two fundamental challenges persist: (1) how to construct street-level indicators that capture accident-related features, and (2) how to quantify their causal impacts across different accident types. To address these challenges, we propose Semantic4Safety, a framework that applies zero-shot semantic segmentation to SVIs to derive 11 interpretable streetscape indicators, and integrates road type as contextual information to analyze approximately 30,000 accident records in Austin. Specifically, we train an eXtreme Gradient Boosting (XGBoost) multi-class classifier and use Shapley Additive Explanations (SHAP) to interpret both global and local feature contributions, and then apply Generalized Propensity Score (GPS) weighting and Average Treatment Effect (ATE) estimation to control confounding and quantify causal effects. Results uncover heterogeneous, accident-type-specific causal patterns: features capturing scene complexity, exposure, and roadway geometry dominate predictive power; larger drivable area and emergency space reduce risk, whereas excessive visual openness can increase it. By bridging predictive modeling with causal inference, Semantic4Safety supports targeted interventions and high-risk corridor diagnosis, offering a scalable, data-informed tool for urban road safety planning.

Open-Vocabulary semantic segmentation (OVSS) and domain generalization in semantic segmentation (DGSS) highlight a subtle complementarity that motivates Open-Vocabulary Domain-Generalized Semantic Segmentation (OV-DGSS). OV-DGSS aims to generate pixel-level masks for unseen categories while maintaining robustness across unseen domains, a critical capability for real-world scenarios such as autonomous driving in adverse conditions. We introduce Vireo, a novel single-stage framework for OV-DGSS that unifies the strengths of OVSS and DGSS for the first time. Vireo builds upon the frozen Visual Foundation Models (VFMs) and incorporates scene geometry via Depth VFMs to extract domain-invariant structural features. To bridge the gap between visual and textual modalities under domain shift, we propose three key components: (1) GeoText Query, which align geometric features with language cues and progressively refine VFM encoder representations; (2) Coarse Mask Prior Embedding (CMPE) for enhancing gradient flow for faster convergence and stronger textual influence; and (3) the Domain-Open-Vocabulary Vector Embedding Head (DOV-VEH), which fuses refined structural and semantic features for robust prediction. Comprehensive evaluation on these components demonstrates the effectiveness of our designs. Our proposed Vireo achieves the state-of-the-art performance and surpasses existing methods by a large margin in both domain generalization and open-vocabulary recognition, offering a unified and scalable solution for robust visual understanding in diverse and dynamic environments. Code is available at https://github.com/SY-Ch/Vireo.

While recent studies have started to measure the experienced racial-ethnic segregation across activity space (beyond “residential” segregation), insufficient efforts have been devoted to revealing the experienced segregation levels of the racial-ethnic minorities (e.g., Asian, Hispanic, Native, and Multi-races), mainly due to the lack of a comprehensive probe into various individual-level datasets. This issue leads to an unnoted deficiency in this field – the lack of showing the “directed” interactions between any pair of two groups. To bridge these gaps, this work proposed a unified framework of using social media data to infer both individual’s mobility patterns and user profiles (i.e., race-ethnicity and economic status), to include more racial-ethnic minorities for a more comprehensive estimation of individual experienced segregation. With these inferred information, we developed two novel individual experienced segregation indices, i.e., individual experienced exposure index (EEI) to each race-ethnicity and individual experienced diversity index (EDI). Moreover, we integrated these two indices with the spatial impacts among the mobility-based activity locations based on distance-decay functions, which are often neglected by existing studies. Using Los Angeles-Long Beach-Anaheim as the study case, we discovered several novel, important findings. First, while experienced isolation (i.e., mainly having intra-group interaction) persists among all groups, Asian is most diverse in inter-group interaction. Second, individuals who live closer tend to have similar levels of inter-group interaction. Moreover, exposures to both White and Black are negatively correlated with exposure to Hispanic. On top of that, individuals with a higher economic status are likely less interested in inter-group interaction (except for those from Hispanic), along with more exposures to both White and Black, but less exposure to Hispanic.

Traditional overhead imagery techniques for urban land use detection and mapping often lack the precision needed for accurate, fine-grained analysis, particularly in complex environments with multi-functional, multi-story buildings. To bridge the gap, this study introduces a novel approach, utilizing ground-level street view images geo-located at the point level, to provide more concrete, subtle, and informative visual characteristics for urban mixed land use analysis, addressing the two major limitations of overhead imagery: coarse resolution and insufficient visual information. Given that spatial context-aware land-use descriptions are commonly employed to describe urban environments, this study treats mixed land use detection as a Natural Language for Visual Reasoning (NLVR) task, i.e., classifying land use(s) in images based on the similarity of their visual characteristics and local descriptive land use contexts, by integrating street view images (vision) with spatial context-aware land use descriptions (language) through vision-language multimodal learning. The results indicate that our multimodal approach significantly outperforms traditional vision-based methods and can accurately capture the multiple functionalities of the ground features. It benefits from the incorporation of spatial context-aware prompts, whereas the geographic scale of geo-locations matters. Additionally, our approach marks a significant advancement in mixed land use mapping, achieving point-level precision. It allows for the representation of diverse land use types at point locations, offering the flexibility of mapping at various spatial resolutions, including census tracts and zoning districts. This approach is particularly effective in areas with diverse urban functionalities, facilitating a more fine-grained and detailed perspective on mixed land uses in urban settings.

This study investigated multi-modal learning as a stand-alone solution to image geo-localization problems. Based on the successful trials on the contrastive language-image pre-training (CLIP) model, we developed GEo-localization Multi-modal (GEM) models, which not only learn the visual features from input images, but also integrate the labels with corresponding geo-location context to generate textual features, which in turn are fused with the visual features for image geo-localization. We demonstrated that simply utilizing the image itself and appropriate contextualized prompts (i.e., mechanisms to integrate labels with geo-location context as textural features) is effective for global image geo-localization, which traditionally requires large amounts of geo-tagged images for image matching. Moreover, due to the integration of natural language, our GEM models are able to learn spatial proximity of geo-contextualized labels (i.e., their spatial closeness), which is often neglected by classification-based geo-localization methods. In particular, the proposed Zero-shot GEM model (i.e., geo-contextualized prompt tuning on CLIP) outperforms the state-of-the-art model - Individual Scene Networks (ISN), obtaining 4.1% and 49.5% accuracy improvements on the two benchmark datasets, IM2GPS3k and Place Plus 2.0 (i.e., 22k street view images across 56 cities worldwide), respectively. In addition, our proposed Linear-probing GEM model (i.e., CLIP’s image encoder linearly trained on street view images) outperforms ISN even more significantly, obtaining 16.8% and 71.0% accuracy improvements, respectively. By exploring optimal geographic scales (e.g., city-level vs. country-level), training datasets (street view images vs. random online images), and pre-trained models (e.g., ResNet vs. CLIP for linearly probing), this research sheds light on integrating textural features with visual features for image geo-localization and beyond.