Contrastive Learning for PM2.5 Prediction

Link to the paper

This repository presents a contrastive learning framework for predicting ground-level PM2.5 from high-resolution micro-satellite imagery. Two contrastive learning frameworks, SimCLR and SimSiam, are tested and then extended to formulate a new framework called Spatiotemporal Contrastive Learning (SCL). The satellite imagery and PM2.5 data is obtained from 2 cities: Delhi and Beijing. The original framework is implemented in Pytorch Lightning Bolts repository: https://github.com/PyTorchLightning/lightning-bolts.

Announcement

Due to the continuous update of Pytorch Lightning and pl-bolts dependency, there might be some code conflicts when running the contrastive learning scripts. Therefore, if you just want to replicate the experimental results in the paper without dealing with those conflicts, please contact [email protected] for pre-trained model weights.

Structure of the Repository

The structure of this repository is given below:

• contrastive_learning: This module contains scripts for unsupervised pre-training with unlabeled satellite images by using regular contrastive learning (SimCLR and SimSiam) frameworks and SCL frameworks.
• contrastive_models: This module contains the backbone architecture of original SimCLR and SimSiam frameworks as well as the corresponding data augmentation functions.
• data: This module should contain the satellite imagery, PM2.5 data and pre-trained weights for both contrastive and supervised learning tasks.
• model_utils: This module contains all the models and the utility functions for both contrastive and supervised learning tasks.
• rt_rf_checkpoint: This module contains the pre-trained parameters for Random Forest and Totally Random Trees Embedding models, which are used for predicting PM2.5 with multi-modal data (e.g. satellite images, meteorology, AOD, etc.). This is considered as future work and is beyond the scope of this paper.
• supervised_learning: This module contains scripts for training and testing the pre-trained model with satellite images and corresponding PM2.5 labels.

Related Work

Contrastive Learning

• T. Chen, S. Kornblith, M. Norouzi, et al., "A simple framework for contrastive learning of visual representations," in International conference on machine learning, PMLR, 2020, pp. 1597-1607.
• T. Chen, S. Kornblith, K. Swersky, et al., "Big self-supervised models are strong semi-supervised learners," arXiv preprint arXiv:2006.10029, 2020. https://github.com/google-research/simclr
• X. Chen and K. He, "Exploring simple siamese representation learning," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 15750-15758. https://github.com/facebookresearch/simsiam
• W. Falcon et al., Pytorch lightning, https://github.com/PyTorchLightning/pytorch-lightning, 2019.

PM2.5 Prediction from Satellite Imagery

• T. Zheng, M. H. Bergin, S. Hu, et al., "Estimating ground-level pm2.5 using micro-satellite images by a convolutional neural network and random forest approach," Atmospheric Environment, vol. 230, 2020. doi: 10.1016/j.atmosenv.2020.117451.
• T. Zheng, M. Bergin, G. Wang, et al., "Local pm2.5 hotspot detector at 300 m resolution: A random forest-convolutional neural network joint model jointly trained on satellite images and meteorology," Remote Sensing, vol. 13, 2021. doi: 10.3390/rs13071356.

Conducting Experiments

Before conducting experiments, please first prepare the data in the format as specified in the file requirements.md under the data directory, or contact [email protected] if you want to replicate the experiments in the paper. Next, run the scripts named in the xxx_pretrain_SimCLR_xxx.py or xxx_pretrain_SimSiam_xxx.py format under the contrastive_learning directory. After that, the pre-trained model weights will be saved to the data directory. Finally, run the scripts named in the xxx_supervised_xxx.py format under the supervised_learning directory. This will generate the visualizations of PM2.5 predictions and the corresponding statistics.

Citation

If you publish any materials using this repository, please include the following Bibtex citation:

@article{jiang2022improving,
  title={Improving spatial variation of ground-level PM2. 5 prediction with contrastive learning from satellite imagery},
  author={Jiang, Ziyang and Zheng, Tongshu and Bergin, Mike and Carlson, David},
  journal={Science of Remote Sensing},
  pages={100052},
  year={2022},
  publisher={Elsevier}
}