Explainable Machine Learning for Predicting Thermal Discomfort from Multipollutant Exposure on Dhaka Roadways

Technical Summary: Explainable Machine Learning for Predicting Thermal Discomfort from Multipollutant Exposure on Dhaka Roadways

Problem Statement
Urban roadways in rapidly developing megacities like Dhaka, Bangladesh, present complex exposure environments where road users (pedestrians, commuters, traffic police, and informal workers) face simultaneous stressors: heat, air pollution (particulate matter and CO₂), and traffic noise. Traditional studies often evaluate these factors in isolation, failing to capture the synergistic effects of multipollutant exposure on human thermal perception. Furthermore, standard statistical methods may struggle to model the non-linear interactions between these variables and subjective thermal comfort. This study addresses the need to quantify and predict thermal discomfort in Dhaka's roadway microenvironments by integrating multipollutant exposure data with explainable machine learning (ML) frameworks.

Methodology
The research employed a cross-sectional, field-based observational design conducted across 17 strategic monitoring sites along three major corridors (Northern, Central, and Southern) in Dhaka. Data collection occurred during the late monsoon, post-monsoon, and winter transition periods (August 2025 – January 2026).

• Data Acquisition:
  • Environmental Sensors: Calibrated intelligent air detectors (JLDC) measured PM₁, PM₂.₅, PM₁₀, and CO₂ concentrations. Mini sound level meters (UT353) recorded traffic noise levels. Meteorological data (temperature, relative humidity, wind velocity) were also captured.
  • Human Perception: Structured interviews with 1,109 participants collected subjective thermal sensation votes based on ASHRAE guidelines (ranging from Very Cool to Hot), alongside data on clothing levels, health complaints (e.g., headaches), and perceived air quality.
• Data Preprocessing: The raw dataset underwent rigorous cleaning, including the removal of sensor errors, duplicates, and missing values. The final modeling dataset comprised 1,109 observations.
• Machine Learning Framework: Three supervised classification algorithms were trained to predict thermal discomfort categories:
  1. Decision Tree (DT): Used as an interpretable baseline.
  2. Random Forest (RF): An ensemble method to reduce overfitting and improve stability.
  3. XGBoost: An advanced gradient boosting model for capturing complex non-linear relationships.
• Explainability and Validation: Model performance was evaluated using ROC-AUC, accuracy, precision, recall, F1-score, and confusion matrices. To ensure interpretability, SHapley Additive exPlanations (SHAP) analysis was applied to determine the relative contribution of each predictor. Sensitivity analysis was conducted to assess model robustness against variations in input variables.

Key Results

• Exposure Burden: The study found a high burden of multipollutant exposure. Mean PM₂.₅ concentration was 71.37 µg/m³, while PM₁ concentration averaged 58.35 µg/m³. Roadside measurements recorded a maximum PM₂.₅ concentration of 100.00 µg/m³ in the study dataset, though the text notes a specific peak of 172.2 µg/m³ in the Mirpur-10 zone. Mean noise levels reached 82.96 dB. These values significantly exceed World Health Organization (WHO) guidelines.
• Thermal Discomfort Prevalence: The field data revealed extreme thermal discomfort, with 99.46% of observations classified as "Hot," "Warm," or "Slightly Warm." Only a negligible fraction reported neutral or cool sensations.
• Model Performance: Ensemble models significantly outperformed the single Decision Tree. Both Random Forest and XGBoost achieved an accuracy of 0.96 and an F1-score of 0.96, with ROC-AUC values indicating excellent discrimination (AUC = 0.96). The models demonstrated high stability in classifying thermal discomfort across different exposure scenarios.
• Predictor Importance (SHAP Analysis): Contrary to the assumption that temperature is the sole driver, SHAP analysis identified Wind Velocity and CO₂ concentration as the most influential predictors of thermal discomfort. PM₁₀, clothing level, and humidity also contributed moderately, while PM₂.₅ and noise levels had smaller but non-negligible impacts. This suggests that microclimatic airflow and traffic-related emissions (indicating congestion and poor ventilation) are critical determinants of discomfort.
• Sensitivity: Sensitivity analysis confirmed that while PM and noise variables remained relatively stable in their impact, variations in CO₂ exposure caused more significant fluctuations in model predictions, highlighting the sensitivity of thermal comfort to traffic density and ventilation conditions.

Significance and Contributions
The paper claims several significant contributions to the fields of urban environmental health and machine learning:

1. Integrated Multipollutant Approach: Unlike previous studies that isolated heat, noise, or pollution, this research provides an evidence-based framework for understanding the combined impact of these stressors on thermal perception in a dense tropical megacity.
2. Explainable AI Application: By utilizing SHAP analysis, the study moves beyond "black box" predictions. It quantifies the specific influence of variables like wind velocity and CO₂, offering actionable insights for policymakers rather than just predictive accuracy.
3. Public Health Relevance: The findings highlight that thermal discomfort in Dhaka is not merely a meteorological issue but a public health crisis driven by the synergy of heat, particulate matter, and noise. The high prevalence of discomfort (99.46%) underscores the vulnerability of outdoor workers and commuters.
4. Policy Framework: The study proposes a multi-sectoral intervention strategy, including nature-based solutions (roadside tree plantation, shaded footpaths), traffic emission controls, and noise mitigation. It suggests that these measures should be targeted at specific high-risk roadway segments identified through spatial analysis.

The authors conclude that addressing thermal discomfort requires an integrated urban planning approach that simultaneously manages air quality, noise, and microclimate, supported by data-driven, explainable predictive models.