This work presents the data processing, model training, testing, and analysis for the purposes of surface ozone bias modelling with Deep Learning and Uncertainty Quantification. Air pollution is a global hazard, and as of 2023, 94% of the world’s population is exposed to unsafe pollution levels Sanchez-Triana (2023). This code implements an uncertainty-aware U-Net architecture to predict the Multi-mOdel Multi-cOnstituent Chemical data assimilation (MOMO-Chem) model’s surface ozone residuals (bias) using Bayesian and quantile regression1 methods for North America and Europe, with extensions to a Global analysis (WIP).
The satellite data products used in this study are available from Google Earth Engine. The list of datasets used to generate the model feature space are below:
Gridded Population of the World
To run the pipeline, the following command is used:
python run_pipeline.py --epochs --optimizer --classes --test-year --overfitting_test --channels --target --region --seed --model_type --data_dir --save_dir --val_percent --analysis_date --tag
The available configurable parameters are:
--epochs: Training Epochs--optimizer: U-Net optimizer--classes: Number of target classes--test_year: Year of test set data--overfitting_test: Specify if you would like to run quick experiment to ensure data can be overfit--channels: Number of channels in feature space--target: Name of target variable (currently supports bias)--region: Region of analysis. Currently supports NorthAmerica, Europe, Globe (WIP)--seed: Specify seed if deterministic experiment desired--model_type: Specify model type. Supports standard (no UQ), CQR and MC-Dropout--data_dir: Directory of stored data--save_dir: Results directory--val_percent: Desired percentage of training set to be set aside for validation--analysis_date: Test month--tag: Wandb experiment tag
The below folders host the following code:
unet: home of run_pipeline.py script and modules for loading dataset, training, testing.
unet/data_processing: all pre-processing scripts to generate feature space.
unet/analysis: scripts for post-processing results into figures and maps.