Projects
❏ Boston: The Green Space, Blue Space, and Urban Heat Island
Summary:
In this study, I investigate how green and blue spaces mitigate the Urban Heat Island (UHI) effect in Boston. Urban areas often experience higher temperatures than their rural counterparts, intensifying heatwaves and associated risks. Boston’s diverse mix of historical, commercial, and natural environments makes it an ideal setting to assess the cooling impacts of parks, rivers, and waterfronts. Findings from this analysis underscore the importance of integrating green and blue spaces into urban planning strategies to counteract the UHI effect.
Methodology:
- Data Sources: Landsat 8 satellite imagery and GIS datasets
- Metrics: Land Surface Temperature (LST), proximity to green and blue spaces, NDVI, NDBI, and Albedo
- Approach:
- Adapted the methodology of Park Jong-Hwa & Cho Gi-Hyoug (2016), focusing on LST as a measure of UHI intensity
- Excluded DEM data due to Boston’s relatively flat terrain
- Analyzed how water proximity influences UHI dynamics, controlling for other key factors
Results and Implications:
The size of, and proximity to, parks and water bodies show a strong inverse correlation with urban temperatures. These findings highlight the crucial role of urban nature in reducing heat stress and provide actionable insights for city planners seeking to mitigate the UHI effect through thoughtful landscape design and policy interventions.
❏ Grid Vulnerability Index (GVI) - New York and Beyond
Summary:
The Grid Vulnerability Index (GVI) assesses power grid stability by integrating real-time weather data with key infrastructure assets—power plants, transmission lines, and critical community facilities. Employing a hierarchical hexagonal grid system (H3), the GVI captures vulnerability scores across various spatial scales and rolls them into a single measure of grid resilience.
Methodology:
- Data Sources:
- U.S. Energy Information Administration (EIA) – Power plant and transmission line data
- United States Geological Survey (USGS) – Critical infrastructure data
- NOAA Weather API – Real-time weather updates
- Scoring System:
- Assigned vulnerability scores based on energy source, capacity, and transmission line length
- Integrated weather impacts (temperature, dewpoint, wind speed, humidity) into final GVI
- Analysis: Merged individual scores for each grid cell to create a composite visualization of grid risk
Results and Implications
High GVI scores were detected in densely populated areas and locations with older infrastructure. These insights are vital for resilience planning, guiding infrastructure investment and emergency preparedness. By enabling real-time updates, the GVI aids utilities and policymakers in anticipating and mitigating power grid disruptions.
❏ Spatial Statistics on Land Surface Temperature in Seoul
Summary:
This project examines how urban form, vegetation coverage, and elevation shape land surface temperature in the Seoul Metropolitan Area. By applying a variety of spatial and non-spatial statistical techniques—such as Moran’s I, Spatial Lag, and Spatial Error models—I evaluated the spatial clustering of temperature and its contributing factors.
Methodology:
- Spatial Autocorrelation: Calculated Moran’s I to confirm clustering in temperature data.
- Regression Modeling:
- Linear regression to assess basic relationships between temperature and predictors (tree coverage, built-up areas, and altitude).
- Spatial Lag and Spatial Error models to account for unobserved spatial dependencies.
- Model Evaluation: Used AIC and Log-Likelihood to compare model performance, applying Lagrange Multiplier tests to distinguish between lag and error specifications.
Results & Implications:
Spatially explicit models offered deeper insight into the drivers of urban heat, underscoring the role of green infrastructure in mitigating temperature extremes. This analysis aids city planners in designing heat-resilient strategies and supports evidence-based policy interventions in high-risk urban zones.
❏ Economic Momentum Indicator
Summary:
In this exercise, I evaluated the economic strength and potential of various locations to withstand climate-related challenges. This indicator is vital for understanding how regions can either mitigate or amplify the impacts of climate events through robust economic foundations. Key metrics integrated into this indicator include employment rates, median income, house prices, and education levels, along with dynamic measures like the income-to-house-price ratio to assess housing affordability. The indicator’s effectiveness was validated by its correlation with returns from five types of real estate, providing deep insights into local economic trends and their broader implications.
Methodology:
I conducted a spatial analysis of economic metrics (e.g., employment, income, education) across CBSAs, incorporating dynamic measures, such as changes in employment rates, median income levels, and educational attainment over time. This allowed me to capture not just the current economic health of an area but also its trajectory. To validate my approach, I modeled and correlated these metrics with real estate returns across property types, using data from the ACS, FHFA, and NCREIF. I calculated and scaled various ratios (e.g., education level, employment, and income-to-house-price) and aggregated them into composite economic momentum scores. Correlation analyses were then used to examine the relationship between these scores and real estate returns, and the geographic distribution of scores was analyzed to identify leading and lagging counties.
Results & Implications:
The analysis of Economic Momentum Values highlighted significant spatial variability, with top-performing counties demonstrating robust economic indicators. I observed distinct geographical patterns, with notable differences between southeast and northeast states. The correlations between economic indicators and returns from different property types varied, revealing the nuanced nature of economic vitality and real estate performance. The insights from the Economic Momentum Indicator can guide policymakers and investors in making informed decisions to enhance resilience and optimize investments in diverse economic landscapes.
❏ ETL and EDA for Internal Migration Analysis
Summary:
This project involved building an ETL pipeline and conducting exploratory data analysis on U.S. Internal Revenue Service (IRS) migration data to identify a potential “Internal Migration” signal. Insights from this analysis could inform residential market demand and investment strategies.
Technical Details:
- ETL Process:
- Extracted data from multiple IRS sources
- Cleaned and standardized datasets
- Loaded the final data into a unified structure for further exploration
- EDA:
- Statistical and visual approaches to detect net migration trends at the county level
- Python libraries (GeoPandas, Plotly, seaborn) for geospatial and statistical analysis
Impact:
This project demonstrates the value of combining ETL pipelines with geospatial analytics to reveal migration hotspots. The newly developed “Internal Migration” signal provides a basis for market demand forecasting and policy decision-making in diverse urban contexts.
❏ Travel Mode Choices: Large Language Models vs Traditional Decision-Making Frameworks
Abstract:
This study analyzed how Large Language Models (LLMs) make judgments in the context of the travel mode choice decision-making process and existing behavioral studies, and compared it with a traditional decision-making framework. Utilizing the Swissmetro dataset, we first applied Expected Utility Theory via a Multinomial Logistic Model to establish a benchmark for travel mode choice predictions. We then established three sets of experiments with GPT-3.5 and compared the results with the benchmark model. First, we analyzed the impact of feature choice on LLM performance and reasoning. Second, we investigated the value of role prompting in terms of human behaviors. Lastly, we surveyed students and requested reasonings on travel behaviors and utilized them to fine-tune the LLM to make predictions. We found that in general, enriching the model with a broader range of features, designing prompts that reflect behavioral tendencies, and fine-tuning the model using human-like reasoning significantly enhance the predictive capabilities of GPT-3.5. We achieved a result almost comparable to that of Expected Utility Theory trained on a large dataset without showing GPT any training examples. This provides crucial evidence and directs potential future research in operationalizing LLMs to make travel behavior decision-making predictions that could benefit transportation planners.