📄 earth_science

UHIQ: An Interactive Explainable Framework for Urban Heat Island Prediction Using Physics-Informed Regression and XGBoost

This study introduces UHIQ, an interactive and explainable framework that integrates physics-informed regression with XGBoost to predict urban heat island intensity, offering researchers and policymakers a transparent tool for analyzing environmental drivers and evaluating heat mitigation strategies.

Original authors: Tara Balaji

Published 2026-09-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tara Balaji

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Cities are often warmer than the countryside surrounding them, a phenomenon known as the urban heat island effect. This happens because concrete, asphalt, and buildings absorb and hold heat much better than soil and plants, while the natural cooling provided by trees and grass is lost. As global temperatures rise and more people move into cities, this extra heat becomes a serious problem, straining power grids, worsening air quality, and putting the health of vulnerable residents at risk. To fight this, city planners need to know exactly how different parts of a city will heat up and what changes, like planting more trees, might do to lower the temperature. However, predicting these temperatures is difficult. Traditional computer simulations that try to model every physical detail of a city are incredibly complex and require powerful supercomputers, while simpler statistical methods often miss the messy, non-linear ways that different environmental factors interact.

A new framework called UHIQ, developed by researcher Tara Balaji, offers a different approach to solving this puzzle. Instead of relying on just one method, the system combines two distinct ways of looking at the data: a straightforward, physics-based model that is easy to understand, and a sophisticated machine learning model that is very good at finding hidden patterns. The goal was not just to predict temperatures accurately, but to create a tool that city officials and researchers could actually use to test ideas in real time. By feeding the system data about vegetation cover, changes in plant life over time, and elevation differences, the framework generates predictions for how hot a specific urban area will be during the day. What makes this work unique is that it doesn't just spit out a number; it includes a visual dashboard that lets users slide controls to change the environment and immediately see how those changes affect the predicted heat, all while showing the reasoning behind the numbers.

When the researchers tested their system, they found that the machine learning model, which uses a technique called XGBoost to learn from data, was the more accurate predictor. It estimated the temperature with an average error of just over one degree Celsius, whereas the simpler, physics-based model had a slightly larger average error. The machine learning model also explained nearly half of the variations in heat intensity across the different cities in the study, compared to about 29 percent for the simpler model. This suggests that the complex, non-linear relationships between the environment and heat are better captured by the advanced algorithm. However, the researchers did not discard the simpler model. They kept it because it acts as a clear, transparent baseline that shows the general rules of how heat works, proving that even without the complex math, the basic principles of cooling through vegetation remain valid.

The most important discovery from this work is that vegetation is the single most powerful factor in controlling urban heat. The study showed that as the amount of green cover increases, the predicted temperature drops. The machine learning model identified that not just the current amount of greenery, but also how much that greenery has changed over the past year, plays a critical role in determining the heat. The system was able to show that if a city increases its vegetation, the heat island effect weakens, confirming what scientists have long suspected but now with a tool that can quantify the impact. The researchers also found that changes in elevation matter, though they are less influential than the presence of plants. By analyzing which variables caused the biggest changes in the predictions, the system confirmed that managing plant life is the most effective lever for cooling a city.

To make these findings useful for real-world planning, the team built an interactive dashboard that anyone can use without needing to know how to code or understand complex statistics. A user can select a specific city from a database and then adjust sliders to represent different scenarios, such as increasing the amount of trees or changing the landscape. As the user moves the sliders, the dashboard instantly updates the temperature prediction, showing the results from both the simple model and the advanced machine learning model side by side. This allows a planner to see, for example, that adding more green space might lower the temperature by a specific amount, and to compare how the two different modeling approaches agree or disagree on that outcome. The tool essentially turns a static scientific study into a dynamic conversation about what a city could look like if it were greener.

While the framework is a significant step forward, the researchers acknowledge that it is not a complete solution for every climate challenge. The current version relies on a specific set of data from urban areas in the United States and focuses on daytime temperatures, which means it might not work exactly the same way in different climates or at night. The model also uses a limited set of environmental factors, leaving out other potential influences like building density or proximity to water. Despite these limitations, the study demonstrates that it is possible to build a system that is both highly accurate and easy to understand. By combining the predictive power of modern machine learning with the clarity of traditional physics-based models, UHIQ provides a practical way for decision-makers to visualize the benefits of green infrastructure before they ever plant a single tree.

Technical Summary: UHIQ Framework for Urban Heat Island Prediction

Problem Statement

Urban Heat Islands (UHIs) represent a critical environmental consequence of rapid urbanization and climate change, leading to elevated temperatures that negatively impact public health, energy consumption, and urban sustainability. While various prediction approaches exist, they often present a trade-off between accuracy and interpretability. Traditional physics-based numerical simulations (e.g., ENVI-met, SOLWEIG) offer realistic microclimate simulations but require substantial computational resources and specialized expertise, limiting their utility for rapid decision support. Conversely, while machine learning (ML) models like XGBoost and deep learning have improved predictive accuracy by capturing complex nonlinear relationships, they often function as "black boxes," offering limited insight into the environmental drivers behind predictions. Furthermore, many existing studies conclude with offline evaluations, lacking accessible tools that allow planners to compare modeling approaches or explore environmental scenarios in real time. There is a demonstrated need for a unified framework that integrates predictive accuracy, model transparency, explainability, and interactivity to support evidence-based urban planning.

Methodology

The study introduces UHIQ, an interactive, explainable framework that combines a physics-informed linear regression model with an Extreme Gradient Boosting (XGBoost) machine learning model. The methodology utilizes the United States Surface Urban Heat Island Database (Version 3), focusing on annual daytime UHI intensity across multiple US urban areas.

Data and Preprocessing

The framework utilizes a subset of variables from the source dataset:

Target: Annual daytime urban heat island intensity ( $UHI_{annual\_day}$ ).
Predictors: Urban Normalized Difference Vegetation Index ( $NDVI_{urb\_CT\_act}$ ), annual vegetation change ( $\Delta NDVI$ ), and elevation difference ( $\Delta DEM$ ).
Preprocessing: Missing values were removed. To ensure numerical stability, $\Delta DEM$ was normalized by dividing by 100.

Dual-Model Architecture

The framework employs two parallel modeling approaches using the same environmental inputs but different feature formulations:

Physics-Informed Linear Regression:
- Designed for interpretability, this model uses a linear equation: $UHI = \beta_0 + \beta_1(1-NDVI) + \beta_2(\Delta NDVI) + \beta_3(\Delta DEM_{100})$ .
- The vegetation metric was transformed to a "vegetation deficit" ( $1-NDVI$ ) to intuitively align higher predictor values with lower vegetation cover and higher expected heat.
- This model serves as a transparent baseline to quantify the influence of individual environmental factors.
XGBoost Machine Learning Model:
- Designed to capture nonlinear interactions, this model utilizes the same predictors plus an engineered interaction feature: $NDVI \times \Delta NDVI$ .
- Configuration: 1,200 boosting estimators, learning rate of 0.05, maximum tree depth of 7, with L1 (0.01) and L2 (1.0) regularization.
- This model aims to improve predictive performance by learning complex relationships between existing vegetation conditions and long-term vegetation dynamics.

Explainability and Evaluation

Performance Metrics: Both models were evaluated on a held-out test set (80/20 split) using Mean Absolute Error (MAE) and the coefficient of determination ( $R^2$ ).
Sensitivity Analysis: The framework systematically varied NDVI from 0.00 to 0.95 while holding other variables constant to observe model responses to vegetation changes.
Permutation Feature Importance: Applied to the XGBoost model to quantify the relative contribution of each feature by measuring the increase in prediction error when a feature is randomly permuted.
Interactive Dashboard: The entire workflow is encapsulated in a Streamlit-based web application, allowing users to select urban scenarios, modify environmental variables via sliders, and compare real-time predictions from both models against observed values.

Key Results

Predictive Performance: The XGBoost model outperformed the physics-informed linear regression model.
- XGBoost: MAE of 1.03°C and $R^2$ of 0.485.
- Linear Regression: MAE of 1.27°C and $R^2$ of 0.290.
- The results indicate that the machine learning model provided more accurate predictions and explained a greater proportion of the variance in UHI intensity.
Sensitivity Analysis:
- The linear regression model exhibited a consistent, nearly linear inverse relationship between NDVI and UHI intensity, showing a reduction of approximately 0.43°C as NDVI increased from 0.00 to 0.95.
- The XGBoost model produced a nonlinear sensitivity profile, with predicted UHI values ranging from -0.70°C to 2.40°C, capturing localized increases and decreases rather than a single monotonic trend.
Feature Importance: Permutation importance analysis identified $\Delta NDVI$ (annual vegetation change) as the most influential predictor (score: 0.638), followed by the interaction feature $NDVI \times \Delta NDVI$ (score: 0.557). Baseline NDVI and $\Delta DEM$ showed moderate and low importance, respectively. Both models consistently identified vegetation-related variables as dominant drivers.

Significance and Claims

The paper claims that UHIQ addresses the gap between high-accuracy predictive modeling and the need for transparent, accessible decision-support tools in urban planning. Its primary contributions include:

Unified Framework: By integrating a physics-informed baseline with a nonlinear ML model, UHIQ allows users to directly compare the trade-offs between interpretability and predictive performance within a single workflow.
Enhanced Explainability: The inclusion of permutation feature importance and sensitivity analysis moves beyond static numerical outputs, providing insights into why predictions are made and how specific environmental changes (e.g., vegetation increase) influence UHI intensity.
Practical Accessibility: The interactive dashboard translates complex modeling into a user-friendly tool for researchers, planners, and policymakers. It enables scenario-based exploration, allowing users to visualize the potential impacts of green infrastructure strategies before implementation.
Balanced Approach: The study demonstrates that while ML models offer superior accuracy for complex urban thermal environments, interpretable statistical models remain valuable for understanding general environmental trends. UHIQ bridges these approaches, supporting evidence-based decision-making for climate-resilient cities without relying solely on "black box" algorithms or computationally intensive simulations.

The author notes limitations, including the use of only three predictors (which excludes factors like building density or impervious surfaces) and the focus on US daytime data, suggesting future work could expand the framework to include more variables, international datasets, and advanced explainable AI techniques like SHAP values. However, the current framework is presented as a practical step toward transparent, data-driven urban heat mitigation planning.