Lujia Zhu

Project 1

Visual determinants of outdoor thermal comfort

Understanding how colors, scene composition, and greenery shape perceived thermal sensation outdoors, using computer vision + perceptual surveys + explainable machine learning.

Computer Vision Perception Survey Explainable AI (SHAP) Urban Climate
Hero media

Overview

Problem

Thermal comfort is often modeled using physical climate variables, but people also react to what they see.

Idea

Extract visual features from street-level images and connect them to thermal sensation votes.

Outcome

Quantify which visual cues matter most, and make the model explainable for design implications.

This project explores the visual determinants of outdoor thermal comfort in tropical urban contexts. Beyond air temperature, people’s thermal sensation can be influenced by colors, material appearance, and perceived shade or openness in the visual scene.

By combining perceptual data (thermal sensation votes) with computer-vision-derived features and explainable ML, the study provides interpretable evidence on how “what we see” shapes “what we feel.”

Method

1) Data

  • Street-level / on-site images across diverse outdoor settings
  • Perceptual survey collecting thermal sensation votes (TSV)
  • Optional: microclimate measurements (e.g., temperature range, peak hours)

2) Visual features

  • Color metrics (hue composition, key color proportions)
  • Scene composition (segmentation-based class proportions)
  • Greenery / sky / building / pavement ratios

3) Modeling

  • Random Forest (baseline) + alternative models (optional)
  • Train/validation split and robustness checks
  • Explainability using SHAP to rank visual predictors

4) Interpretation

  • Feature importance & partial dependence behavior
  • Non-linear / diminishing effects (e.g., greenery threshold)
  • Design implications for climate-responsive urban spaces

Tip: replace the bullets above with your exact dataset sizes (e.g., number of images, participants) once you’re ready — this template is intentionally “publication-safe” and not overly specific.

Key findings

Colors matter

Certain color cues strongly correlate with perceived coolness/warmth.

Greenery isn’t always linear

The effect of greenery can show diminishing returns beyond a threshold.

Explainability helps design

SHAP outputs translate ML findings into interpretable design insights.

These findings challenge a purely “physical metrics only” approach and suggest designers can consider visual/psychological cues alongside shading, ventilation, and materials.

Figures & artifacts

Add your figures as images in /assets/images/ and replace the placeholders below.

Figure 1 — Pipeline / framework diagram
Figure
Workflow overview

Computer vision → features → ML model → SHAP interpretation.

Figure 2 — SHAP summary / key results
Figure
Explainability results

Visual features ranked by contribution to perceived thermal sensation.