DeepUHI: Fine-grained Urban Heat Island Effect Forecasting

This repository provides the official PyTorch implementation of our paper in KDD 2025 Research Track:

"Fine-grained Urban Heat Island Effect Forecasting: A Context-aware Thermodynamic Modeling Framework"

Overview

DeepUHI is a data-driven context-aware thermodynamic modeling framework designed to forecast urban heat island (UHI) effects at a fine spatial and temporal granularity. Our approach leverages deep learning and domain knowledge to provide accurate, interpretable predictions for urban climate research and policy-making.

Features

• Fine-grained UHI forecasting using deep neural networks
• Context-aware thermodynamic modeling
• Decomposition of urban thermal mechanics within the modeling design
• Efficency and interpretablity

Requirements

• Python >= 3.8
• PyTorch >= 1.4.4
• (Recommended) Linux server with CUDA 12.2 and 1x A6000 GPUs

To set up the environment:

conda create -n DeepUHI python==3.8
conda activate DeepUHI
pip install -r requirements.txt

SeoulTemp Dataset

We collect and introduce \textit{SeoulTemp}, the first fine-grained urban temperature dataset that includes field environment data across multiple modalities. This dataset encompasses a total of 947 temperature stations covering 605 $km^{2}$ of land in Seoul from 2021 to 2024 (or later), specifically targeting spatio-temporal UHI effect forecasting at the street level in urban areas.

Note: SeoulTemp is continuously updated by us to include the latest urban temperature records.

To access the dataset for experiments, download it from:

• Google Drive

After downloading, place the data in the ./dataset directory.

Quick Start

To train and test the model, run:

python run.py
bash ./run_test.sh

Web Demo

An offline demo of the SeoulUHI system is available at: SeoUHI Web Platform

Citation

If you use this code or dataset, please cite our paper:

@inproceedings{zou2025fine,
  title={Fine-grained Urban Heat Island Effect Forecasting: A Context-aware Thermodynamic Modeling Framework},
  author={Zou, Xingchen and Ruan, Weilin and Zhong, Siru and Hu, Yuehong and Liang, Yuxuan},
  booktitle={Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2},
  pages={4226--4237},
  year={2025}
}

Acknowledgments

We thank the Seoul City Government for the S-Dot (Seoul Data of Things) project and the provision of raw temperature data via the S-Dot and Seoul's IoT city data platform.

The SeoulTemp dataset and SDot data platform are freely available for non-commercial academic use.

Disclaimer: This dataset is a personal research product and does not represent the official position of the Seoul City Government.

Contact

For questions or collaborations, please contact: [xczou@connect.hku.hk]