Thermal Comfort in Sight: Thermal Affordance and Its Visual Assessment (VATA)

Urban Analytics Lab | Ideas Lab | Singapore-ETH Centre

• 2026-08-16: Code 1–5 training notebooks, survey tables, and the two-stage MTNNL weights from this repo.
• 2025-04-16: Added UrbanCode integration for thermal comfort prediction.
• 2025-04-15: 1st version of Github page updated.
• 2025-01-22: Paper available online in Building and Environment journal.
• 2024-10-12: 1st version preprint available on arXiv.

├── code_1_svi_perception_survey_results.ipynb   # Pairwise survey → TrueSkill → training tables
├── code_2_svi_image_features.ipynb              # 59 image features (or load released Features.csv)
├── code_3_dataset_splits.ipynb                  # Frozen 60/20/20 and 5-fold splits
├── code_4_mtnnl_train_predict.ipynb             # Train the two-stage network
├── code_5_enrm_inference.ipynb                  # Elastic-net inference
├── scripts/                                     # Shared helpers imported by the notebooks
├── examples/test_svi_comfort_prediction.ipynb   # Score a folder of SVIs
├── data/
│   ├── perception_survey/raw/                   # Pairwise comparisons (Code 1)
│   ├── perception_survey/ranked/                # TrueSkill μ, σ, 0–5 Score
│   ├── training_features/                       # Features.csv, svi_scores.csv, splits
│   └── images/survey/                           # 500 SVIs — Hugging Face, not in git
└── model_outputs/                               # best_model.pth + training metrics

Perception dimensions: thermal_comfort (VATA) plus 20 visual-perceptual indicators (comfort, temp_intensity, sun_intensity, …).

Survey images and the Code 4 checkpoint are on Hugging Face (dataset, model). Code 2 / Code 4 download them into data/images/survey/ and model_outputs/best_model.pth (the .pth is not in git).

1. Pairwise survey → scores (Code 1). Each of the 21 dimensions is a set of image-pair choices. TrueSkill turns the full set of pairs into a 0–5 score per image (ranked/). Those scores are joined onto the 59 image features to make svi_scores.csv.
2. Image features (Code 2). Semantic segmentation, object counts, colour/texture, and scene probabilities (59 columns). The released Features.csv can be used as-is.
3. Splits (Code 3). Freeze 60% train / 20% validation / 20% test (seed 42), plus a 5-fold index file.
4. MTNNL (Code 4). ResNet-50 reads the street-view image. Its embedding is concatenated with the 59 image features to predict the 20 VPIs, then image + IF + VPI predict VATA. Training is 40 epochs, batch size 8, Adam 1e-4, combined VPI + VATA MSE. The best validation checkpoint is model_outputs/best_model.pth.
5. ENRM (Code 5). Elastic-net on the same table for IF → VATA, VPI → VATA, IF+VPI → VATA, and IF → each VPI.

Each notebook can be run independently. For a full training run, go 1 → 5:

pip install -r requirements.txt

Survey images and best_model.pth are pulled from Hugging Face the first time you run Code 2 or Code 4.

from scripts.download import ensure_checkpoint, ensure_survey_images
from scripts.mtnnl import evaluate_checkpoint

ensure_survey_images()
ckpt = ensure_checkpoint()
print(evaluate_checkpoint(ckpt))

Set TRAIN = True in code_4_mtnnl_train_predict.ipynb to train the same two-stage network from svi_scores.csv (40 epochs, batch size 8). A shorter walkthrough is in examples/test_svi_comfort_prediction.ipynb.

• Release survey pairwise files, TrueSkill rankings, image-feature tables, and Code 1–5.
• Host the Code 4 checkpoint on Hugging Face.
• Update predicted datasets for 10 southeast asia cities.
• Update predicted datasets for 10 global cities.
• Update Tutorial on how to download SVIs with ZenSVI and predict thermal affordance with our study.

The intensifying Urban Heat Island (UHI) effect poses significant challenges to cities worldwide. It drives extreme climate changes, increases energy consumption for cooling, and degrades public health by worsening air quality and reducing outdoor thermal comfort (OTC). OTC, defined as one's subjective satisfaction with urban thermal conditions, is critical for urban livability. Poor OTC can lead to heat-related illnesses and even affect mental health.

Traditional methods for evaluating OTC rely on field surveys and environmental measurements, which are often costly, resource-intensive, and limited in spatial scale and precision. While indices like PET, PMV, and UTCI exist, applying them effectively in complex urban environments remains challenging.

Recent advances in using Street View Imagery (SVI) combined with computer vision offer promising avenues for large-scale streetscape analysis. SVI can capture both objective image features (IF) like geometry and greenery, and subjective visual-perceptual indicators (VPI) like comfort and enclosure through surveys. Studies have linked SVI features to thermal environments, but significant research gaps remain.

There's no clear theoretical framework linking the objective properties of the built environment to thermal comfort potential in streetscape design.

Our Response: We introduce the concept of Thermal Affordance, inspired by Gibson's theory, to describe the inherent capability of a streetscape's configuration to influence thermal comfort.

Human visual assessment of streetscapes' thermal properties using SVI is understudied.

Our Response: We propose the Visual Assessment of Thermal Affordance (VATA) framework, integrating SVI-derived IFs and survey-based VPIs to quantify thermal affordance.

There's a lack of a robust, replicable workflow integrating SVI and perception surveys for urban-scale thermal-affordance modelling.

Our Response: The VATA framework provides a data-driven workflow from image features and pairwise surveys to a two-stage prediction model and an interpretable inference model.

Figure 1: Conceptual illustration of Thermal Affordance

Thermal Affordance refers to the inherent capability of an environment (like a streetscape) to impact thermal comfort, integrating various environmental factors. It aims to be:

• Unified: Encompassing all relevant fixed environmental variables.
• Objective: Focused on environmental properties, not just subjective feelings.
• Heuristic: Inspiring understanding and analysis of thermal comfort.
• Spatially Dependent: Emphasising differences between environments.
• Interpretable: Linking environmental attributes to thermal comfort potential.
• Expandable: Allowing incorporation of additional variables over time.

Figure 2: The VATA Framework for assessing thermal affordance

The VATA framework addresses the challenges of traditional OTC evaluation. It leverages the connection between visual data (SVI), human perception, and thermal comfort.

• Image Features (IF): Extracted from SVI using computer vision (e.g., semantic segmentation for greenery/buildings, object detection for cars/people, pixel features for colour/texture, scene recognition).
• Visual-Perceptual Indicators (VPI): Gathered through online surveys where participants compare pairs of SVIs based on 19 indicators (e.g., perceived temperature, greenery rate, enclosure, safety, beauty) plus VATA itself.

• A Multi-Task Neural Network Learning (MTNNL) model predicts VATA scores. It uses a two-stage approach: IFs predict VPIs, and then IFs + VPIs predict VATA.
• An Elastic Net Regression Model (ENRM) is used for inference, revealing the interpretable relationships between specific IFs, VPIs, and the final VATA score.

Figure 3: Research framework and methodology

• Study Area: Singapore (chosen for its diverse urban forms and consistent tropical climate).
• SVI Survey: 500 representative SVIs (selected via k-means clustering of 92,233 images) were evaluated by 176 participants in an online survey. Participants made pairwise comparisons for VATA and 19 VPIs. TrueSkill algorithm converted comparisons into scores (0-5 scale).
• IF Extraction: 59 image features (segmentation, objects, colour/texture, scene) from each SVI.
• Model Training: MTNNL trained on IFs, the street-view image, and VPI scores (60% train, 20% validation, 20% test).
• Inference Modelling: ENRM on the same table to read feature weights and relationships.

The two-stage MTNNL trained in Code 4 reaches test R² 0.724 (MAE 0.395, RMSE 0.504) on the frozen 20% holdout, ahead of the IF-only tabular baselines in the same notebook.

Figure 10: High-resolution VATA map of Singapore

We generated a high-resolution map of VATA across Singapore, aggregated into hexagonal units. This map visually identifies areas with high (e.g., parks like Windsor Nature Park, East Coast Park) and low (e.g., parts of Choa Chu Kang) thermal affordance, guiding potential interventions.

The ENRM model (IF+VPI Adj R² 0.732 on this training table) revealed key factors:

Positive Contributors to VATA:

• Vegetation
• Terrain
• Sky proportion
• Shading area
• Human-scale design elements
• Perceived humidity/wind
• Perceived beauty/safety

Negative Contributors to VATA:

• Traffic elements
• Complex/construction scenes
• High pixel detail/contrast
• Perceived temperature/sunlight intensity
• Perceived traffic flow/complexity/boredom

Figure 13: Discussion of the VATA framework's significance and applications

The VATA framework provides a scalable, cost-effective method for assessing urban thermal affordance.

High-resolution VATA maps help planners identify areas needing improvement and prioritise interventions like adding greenery or shading.

Inference models reveal which specific streetscape features (IFs) and perceptual qualities (VPIs) most influence thermal affordance, informing evidence-based design decisions.

The framework supports a continuous monitoring cycle, using updated SVI data to refine models and track the impact of interventions over time.

While developed for Singapore, the methodology can be adapted to other cities by conducting local surveys.

1. Temporal Variance: Did not account for time-of-day, weather, or seasonal changes in SVI.
2. Data Sources: Could be enhanced by integrating satellite imagery, LST data, or detailed microclimate simulations.
3. Survey Limitations: Online surveys have inherent limitations; incorporating expert assessments could be beneficial.
4. Cross-city use: Testing the framework in different climatic and urban contexts is crucial.

This research introduces the novel concept of thermal affordance and the VATA framework for its assessment. By integrating SVI, human perception surveys, and machine learning, VATA offers a tool to evaluate thermal comfort potential in urban streetscapes and to inform sustainable urban planning and design.

Yang S, Chong A, Liu P, Biljecki F (2025): Thermal comfort in sight: Thermal affordance and its visual assessment for sustainable streetscape design. Building and Environment 271: 112569. https://doi.org/10.1016/j.buildenv.2025.112569

@article{yang2025thermal,
  author  = {Yang, Sijie and Chong, Adrian and Liu, Pengyuan and Biljecki, Filip},
  title   = {Thermal comfort in sight: Thermal affordance and its visual assessment for sustainable streetscape design},
  journal = {Building and Environment},
  volume  = {271},
  pages   = {112569},
  year    = {2025},
  doi     = {10.1016/j.buildenv.2025.112569}
}

This work is licensed under a Creative Commons Attribution 4.0 International License.

This means you are free to:

• Share — copy and redistribute the material in any medium or format
• Adapt — remix, transform, and build upon the material for any purpose, even commercially

Under the following terms:

• Attribution — You must give appropriate credit, provide a link to the license, and indicate if changes were made. You may do so in any reasonable manner, but not in any way that suggests the licensor endorses you or your use.