This repo is the official implementation for IJCAI 2025 paper: OpenCarbon: A Contrastive Learning-based Cross-Modality Neural Approach for High-Resolution Carbon Emission Prediction Using Open Data
This work aims to construct a prediction framework that predicts high-resolution carbon emissions with open data of satellite images and POI.
We conduct experiments on cities that span both developed and developing countries, including London, Beijing, and Yinchuan. Summary of the datasets are presented:
| Region | Great London | Beijing | Yinchuan |
|---|---|---|---|
| Area | 778 km² | 1381 km² | 475 km² |
| GDP pc ($) | 71k | 27k | 12k |
| POI Source | SafeGraph | Map Service | Map Service |
| Target Year | 2018 | 2018 | 2019 |
Table 1: The summary statistics of our datasets.
Due to the size limit of github, we have stored the data in an anonymous google drive link: https://drive.google.com/drive/folders/1_HHa5X6nLiB4mHfEIn42jb5fwc64nf0v?usp=sharing. [Due to cloud storage limit, we only update the Beijing dataset. All other datasets are available through email requests.] Please download them and place them inside /data.
- Tested OS: Linux
- Python >= 3.8
- torch == 2.2.1
- Tensorboard
Configs for performance reproductions on all datasets.
python main_contrastive.py --city london --batch_size 32 --lr 1e-3 --epochs 3000 --contrastive 1 --alpha 0.001 --neighbor_size 1
python main_contrastive.py --city beijing --batch_size 32 --lr 1e-3 --epochs 3000 --contrastive 1 --alpha 0.001 --neighbor_size 1
python main_contrastive.py --city yinchuan --batch_size 32 --lr 5e-4 --epochs 3000 --contrastive 1 --alpha 0.01 --neighbor_size 1