Spatial statistics is a central component of spatial analysis because it provides methods for identifying, modeling, and explaining spatial patterns and processes. A spatial process refers to a phenomenon whose values, relationships, or interactions vary across geographic space rather than occurring independently of location.

Broadly, spatial regression models can be divided into two major groups. The first includes global spatial regression models, such as spatial lag, spatial error, Spatial Durbin, and related specifications. These models account for spatial dependence by recognizing that nearby observations may influence one another. However, they still estimate a single coefficient for each predictor across the entire study area. As a result, they can model spatial dependence but do not directly reveal how the relationship between a predictor and an outcome varies from place to place.

The second group includes spatially varying coefficient models, whose primary objective is to allow regression relationships to vary across space. Instead of estimating one coefficient per predictor for the entire study region, these models estimate location-specific coefficients, enabling researchers to examine how the effect of a predictor changes among observations or geographic locations. This broader class includes frequentist approaches, Bayesian spatially varying coefficient models, spatial filtering approaches, and, more recently, methods based on machine learning and deep learning. Together, these approaches provide increasingly flexible ways to characterize spatial heterogeneity and complex geographic relationships.

In the second group, also known as local spatial regression models, two broad types of models have been developed within each methodological class: single-scale and multiscale models. Single-scale models assume that all predictor–outcome relationships operate over the same spatial scale, typically represented by a common bandwidth or neighborhood structure. In contrast, multiscale models allow each predictor to operate at its own spatial scale, recognizing that different processes may vary over different geographic extents. This distinction is important because some relationships may be highly localized, while others may reflect broader regional or even near-global spatial patterns.

My research on spatial regression models focuses on advancing the conceptual understanding of how relationships between variables vary across space and how different forms of spatial relatedness can be represented within local modeling frameworks. In particular, I have worked on extending geographically weighted regression by incorporating not only geographic proximity but also similarity in attribute space, leading to the development of Similarity-Based Geographically Weighted Regression (SGWR) and its multiscale extension (M-SGWR). Conceptually, this work is motivated by the idea that spatial processes may be shaped not only by where observations are located, but also by how similar or connected they are in other dimensions. Geographic distance represents only one form of proximity, whereas meaningful relationships may also emerge through social, functional, network, behavioral, or attribute-based proximity, depending on the context. My study in this domain aimed to expand the concept and theory of proximity beyond the predominantly geographic interpretation used in conventional spatial analysis.

My research in GeoAI and the application of Large Language Models (LLMs) to spatial analysis includes both algorithm and tool development in collaboration with colleagues. During my master’s studies, I focused on developing algorithms for multiple geographical feature label placement, including point, polyline, and polygon features. Geographical feature label placement is a well-known NP-hard optimization problem in cartography, meaning that the computational difficulty grows rapidly as the number of geographic features and possible label positions increases. For maps containing many points, polylines, and polygons, identifying a high-quality labeling configuration requires evaluating a very large number of possible arrangements while avoiding label overlap, ambiguity, and other cartographic conflicts. As a result, exact solutions quickly become computationally impractical for large datasets, making heuristic, metaheuristic, and parallel-computing approaches important for improving both solution quality and computational efficiency.

My work on multiple geographical feature label placement (MGFLP) has focused on improving both the flexibility and computational efficiency of automatic map labeling for point, line, and polygon features. In one study, my collaborators and I introduced a multiple-candidate-position framework in a two-degrees-of-freedom space, extending candidate placement beyond conventional point-feature approaches and providing a more general formulation for labeling mixed geographic features. Building on this, we developed a parallel map-segmentation algorithm that decomposes the labeling problem into smaller subproblems and processes them simultaneously across CPUs using Message Passing Interface (MPI). The optimization strategy combined discrete differential evolution and genetic algorithms, and the study demonstrated strong scalability and substantial computational gains for large labeling problems. More recently, we proposed an MPI-based parallel genetic algorithm that combines fixed-position and sliding label-placement models. This approach was designed specifically to address the high computational cost of MGFLP while maintaining cartographic quality across mixed feature types. The results showed substantial reductions in runtime together with fewer label–label and label–feature conflicts compared with earlier approaches. Together, these three studies represent progression from expanding the conceptual space of candidate label positions, to decomposing the optimization problem for parallel processing, and finally to developing a scalable MPI-based framework for efficient, high-quality label placement.

More recently, I have become interested in applying large language models (LLMs) to spatial analysis and extending this work into public health and epidemiology. In one study, I evaluated the feasibility of using GPT-4 to support systematic-review screening for HIV medication-adherence research, where the goal was to identify relevant studies from a large pool of scientific records using predefined inclusion and exclusion criteria. The experiment compared several prompt-engineering strategies, including standard input-output prompting, zero-shot chain-of-thought prompting, few-shot chain-of-thought prompting, and an ensemble majority-voting approach that combined the outputs of the different strategies. The study used 1,680 records drawn from a larger set of 15,839 unique publications, including 39 studies that had been manually identified as relevant by trained reviewers and were treated as the benchmark for evaluation. The results showed that prompt design had a substantial influence on screening performance: standard prompting achieved a recall of 0.82, while zero-shot chain-of-thought increased recall to 0.97, and both few-shot chain-of-thought and majority voting identified all 39 relevant studies, achieving a recall of 1.00. Among the tested approaches, majority voting produced the strongest overall balance between precision, recall, and F-score, although it also required more computational time and cost.

My collaborative research on GIS Copilot, autonomous geospatial data retrieval, and spatial text-to-SQL has focused on a broader question: how can artificial intelligence make geospatial analysis more accessible, efficient, and interactive without separating users from the underlying analytical process? Through these projects, we have explored how large language models and agent-based systems can translate natural-language questions into structured geospatial tasks, identify and retrieve relevant spatial data, and connect users with GIS operations that would otherwise require substantial technical knowledge. In the GIS Copilot work, our goal has been to investigate how conversational AI can support spatial analysis by helping users move from a natural-language question toward an executable GIS workflow. Related work on autonomous geospatial data retrieval addresses another major barrier in spatial analysis: finding appropriate datasets, understanding their structure, and preparing them for subsequent analysis. Rather than treating data discovery as a separate manual step, we explored how autonomous agents could assist with searching, interpreting, and retrieving geospatial information as part of a more integrated workflow. My collaboration on spatial text-to-SQL extends this idea to spatial databases, where natural-language questions can be translated into structured queries that incorporate geographic relationships and spatial operations. We also conducted a study on the capability of LLMs for floodwater-depth estimation. Conceptually, I see these projects as different components of an emerging autonomous GIS ecosystem. Together, they move beyond using LLMs simply as conversational interfaces and toward systems that can coordinate data discovery, query formulation, spatial analysis, and interpretation.

In recent years, particularly following the COVID-19 pandemic, human mobility science has received growing attention because of the increasing availability of large-scale mobility data collected through smartphones, location-based services, and commercial data providers. These data provide new opportunities to study when, where, and how people move across space and interact with different places. Compared with traditional mobility data sources such as travel surveys or census-based commuting records, smartphone-based mobility data can capture population movement at much finer spatial and temporal resolutions and across very large geographic areas. Human mobility science is therefore increasingly important for understanding a wide range of social, environmental, and public-health processes. Mobility patterns can reveal how populations access workplaces, healthcare facilities, recreational spaces, businesses, and other destinations, while also providing insight into exposure to environmental hazards, infectious diseases, and neighborhood conditions beyond residential locations. During the COVID-19 pandemic, mobility data became particularly valuable for examining changes in population movement and their relationships with disease transmission and public-health interventions.

At the same time, these emerging datasets introduce important methodological challenges. Commercial mobility data may not represent all population groups, places, or geographic areas equally. Consequently, understanding data coverage, representativeness, privacy, and potential bias is essential before mobility data can be reliably used for scientific inference and policy decisions. Part of my research has focused on evaluating these limitations. In one study, we examined the representativeness of commercial mobile-location data across different geographic scales and over time in the United States, demonstrating that data coverage and population representation can vary considerably depending on location, spatial resolution, and demographic context. This work highlighted that mobility datasets should not automatically be treated as representative samples of the underlying population, particularly when they are used to compare communities or infer population-level behavior. More recently, we have extended this research from population representation to the places themselves. In this work, we evaluate whether points of interest (POIs) captured by commercial mobility data adequately represent the broader landscape of real-world places, whether coverage differs systematically across categories and communities, and whether visitors observed at those locations reflect the populations they are intended to represent. The study examines multiple dimensions of bias, including POI coverage, category-specific representation, urban–rural and socioeconomic differences, and visitor and visit representation. Together, these studies emphasize that mobility-data bias is multidimensional: it may arise from who is observed, which places are captured, where those places are located, and how frequently observed users interact with them. Understanding these layers of representation is essential for determining when mobility data provide reliable evidence and when resulting analyses may reproduce or amplify existing geographic and population inequalities.

Spatial knowledge and skills are critical to public-health research because health outcomes are rarely independent of place. Where people live, work, travel, and spend time determines their exposure to social, environmental, economic, and healthcare conditions that can shape health risks and opportunities. The relationship between people and place is also reciprocal: populations modify their environments through development, mobility, land use, and social activity, while those same environments influence behavior, exposure, accessibility, and ultimately health. Spatial approaches therefore help public-health researchers move beyond asking only who is affected to also ask where, why there, and under what geographic conditions. GIS, spatial statistics, mobility analysis, and related geospatial methods can be used to identify disease clusters, measure environmental and neighborhood exposures, evaluate access to healthcare and other resources, study population movement, detect spatial inequalities, and understand how health relationships vary across locations. By explicitly incorporating geographic context, spatial knowledge provides a more complete understanding of population health and can support more targeted, place-sensitive public-health interventions and policies.

I have been involved in a range of interdisciplinary public-health studies where spatial thinking, geospatial data, and quantitative methods are used to understand how health outcomes vary across places and populations. My work has included infectious diseases such as HIV/AIDS, gonorrhea, chlamydia, and respiratory infections, as well as chronic and population-health outcomes including lung cancer, cardiovascular disease, obesity, mental health, and healthcare utilization. Across these studies, I have examined how neighborhood conditions, residential segregation, population mobility, and place visitation relate to geographic disparities in health. For example, my research has investigated spatiotemporal patterns of chlamydia and gonorrhea in relation to residential segregation, the spatial association between gay-bar visitation and HIV prevalence across major U.S. cities, and the use of large-scale human mobility data to study lung cancer at the census-tract level. I have also contributed to research on smartphone-based place visitation and coronary heart disease, neighborhood-level health estimation, and racial disparities in acute-care utilization. A common theme across this work is the importance of place, mobility, and social context in shaping health. Rather than treating health outcomes as spatially isolated, I use spatial statistics, mobility data, and geospatial big data to examine how environmental and social exposures, connectivity, and population behavior vary geographically. This perspective has also motivated my growing interest in integrating GeoAI and LLM-based tools into epidemiological research and public-health decision support.

These two figures illustrate state-level temporal trends in gonorrhea and chlamydia incidence rates per 100,000 population across the United States from approximately 2013–2021. The figures highlight substantial geographic heterogeneity in both the magnitude and trajectory of disease incidence. Gonorrhea generally shows increasing trends in many states, although the timing and rate of increase vary considerably. Chlamydia exhibits more stable or gradually increasing patterns before 2020, followed by noticeable declines in many states around the COVID-19 pandemic period and partial rebounds thereafter. For full details, please refer to the complete article.