High Resolution Air Pollution Mapping for Louisiana
Alberto Lopez Rueda, Alexandra Hurst,
Amanda Murray, Kory Kirshenbaum
Problem Statement
Air Pollution is a Major Concern in Louisiana
“You gotta live here to try to breathe the air, drink the water, see the children so sick and watch people die.”
-Geraldine Watkins
Concerned Citizens of St. John the Baptist Parish
Source: ProPublica 2019[1] (adapted)
Source: CNN 2017 [2]
Lack of Sensors (and Trust)
Source: LA Illuminator 2023 [3]
“We have enough problems without running industry out”
-State Senator Bodi White
In response to a proposed bill that would require fence line air monitors
Our Project
Data?
Project Goal:
Support local research and advocacy efforts to understand and reduce impacts of air pollution in Louisiana.
Research Question:
Creating The Dataset
Advancing Research
Supporting Advocacy
Pollution Estimate Dashboard
Our Contribution:
Sensor Placement Algorithm
First regional analysis of disparities in individual pollutant exposure
High-resolution weekly ground level NO2
estimates for Louisiana for 2019 - 2023
Creating
The Dataset
Supporting
Advocacy
Advancing Research
Creating
The Dataset
The Gap: No up-to-date high-resolution NO2 data
Our Objective: Use machine learning to develop an open-source NO2 dataset at a resolution of 1km2 using the state-of-the-art Sentinel-5P satellite
Data Sources
X values
Sentinel-5P Satellite Data
17 Additional Covariates
Y values
NO2 Ground-Sensors
~50 sensors across
Louisiana and Texas
RESEARCH
THE DATASET
ADVOCACY
Tropospheric NO2 Concentrations
(3.5 x 5.5 km resolution)
Data Pipeline
Satellite Data
X
Y
Additional Covariates
Sensor Values
RESEARCH
THE DATASET
ADVOCACY
Aggregate Weekly
Aggregate/ Rescale to Weekly
Downscale
1km x 1km
Regrid to satellite coordinates
Biharmonic Inpainting
R = 0.91
Filter Validated Values
Create Raster
Aggregate
weekly
Match to inputs to sensor with inverse distance weighting
Convert
to tabular
Feature Engineering
Ready for Modeling
Modeling Overview
ML & Deep Learning Techniques
FFNN
1D CNN
XGBoost
TabNet�tabular transformer
Final Ensemble�XGBoost + FFNN + TabNet
RESEARCH
THE DATASET
ADVOCACY
Minimized RMSE using 10-fold
Sensor Cross-validation
Full Coverage Ground-Level �Predictions
Data Pipeline
Satellite Data
X
Y
Additional Covariates
Sensor Values
Accurate and Reliable Predictions
| Final Ensemble (LA & TX) | Ghahremanloo �et al. (2021) (TX) |
RMSE | 2.69 | 3.41 |
RMSE (%)* | 43.6% | 47.0% |
MAB (%)* | 31.5% | 33.2% |
RESEARCH
THE DATASET
ADVOCACY
* RMSE divided by mean observed NO2 values
Better results than most similar published research
Highly accurate predictions in new, unseen locations
Prediction
Actual
Sample sensor from test set
Predictions and Residuals
RESEARCH
THE DATASET
ADVOCACY
Predicted vs. Actual NO2
Residuals by Sensor
Results Comparable to Highly-Cited Research
*Difference is metrics, methods, and calculations makes it challenging to compare directly across research papers
Center For Air, Climate, and Energy Solutions National NO2 Dataset[4]
1979-2015, National
Cross Validation R2=0.87
(Annual, Block Groups)
Cited in health and disparity studies
Anenberg et al. (2022)[5]
1990-2019, Global
Cross Validation R2=0.54
(Annual, 1km x 1km)
Cited in pediatric asthma and disparity studies
Atmospheric Composition Analysis Group (St. Louis University)[6]
2005-2019, Global
Test R2=0.53
(Annual, 2.8 km x 2.8 km)
Authors >100 publications since 2019
RESEARCH
THE DATASET
ADVOCACY
Test R2 | Cross Validation R2 |
0.70 | 0.60 |
Comparison to Expert-Recommended NO2 Datasets
Summary
RESEARCH
THE DATASET
ADVOCACY
Developed the only up-to-date, high resolution dataset for NO2 in Louisiana. Comparable to existing state-of-the art models.
Advancing
Research
Our Objective: Demonstrate how our high resolution air pollution dataset can enhance existing environmental justice research in the region.
Prior Research
RESEARCH
THE DATASET
ADVOCACY
Terrell et al. 2022[7]
Disparities in NO2 emissions by Census Tract
“We didn’t have access to a good dataset for NO2 concentrations”- K. Terrell
Our Addition
Disparities in NO2 pollution by Census Block
~170 km2
(average)
<1 km2
(average)
Community <> Community
Individual <> Individual
Block
Tract
Disparity Analysis Methodology
Assign Pollution by Census Block Centroid*
RESEARCH
THE DATASET
ADVOCACY
Aggregate and Average by Race**
5 ppb
Avg= 4.2 ppb
0.5 ppb
Avg= 1.86 ppb
5 ppb
1 ppb
4 ppb
2 ppb
5 ppb
5 ppb
5 ppb
1 ppb
5 ppb
5 ppb
*Methodology adapted from: Sun-Young Kim, Matthew Bechle, Steve Hankey, Lianne Sheppard, Adam A. Szpiro, and Julian D. Marshall. 2020. Concentrations of criteria pollutants in the contiguous U.S., 1979 – 2015: Role of prediction model parsimony in integrated empirical geographic regression. PLOS ONE, 15(2):e0228535.
**Methodology adapted from: arah E. Chambliss, Carlos P. R. Pinon, Kyle P. Messier, Brian LaFranchi, Crystal Romeo Upperman, Melissa M. Lunden, Allen L. Robinson, Julian D. Marshall, and Joshua S. Apte. 2021. Local- and regional-scale racial and ethnic disparities in air pollution determined by long-term mobile monitoring. Proceedings of the National Academy of Sciences, 118(37).
Racial disparities in NO2 pollution exposure | |
Average Exposure: 4.60 ppb | |
White (NH)* | -8% |
Other (NH)** | -3% |
Black (NH)* | +10% |
Hispanic or Latinx | +17% |
Asian (NH)* | +20% |
* NH signifies non-hispanic or latinx
**Other includes american indian or alaskan native, native hawaiian or pacific islander, other, or two or more races
Pollution exposure is assigned based on the average amount of pollution at the centroid of the census block. “Black” and “White” are specified in the graph because these demographics make up over 80% of the population (56% and 31% respectively).
RESEARCH
THE DATASET
ADVOCACY
Summary
Identify demographic disparities in individual pollution exposure
Lay the groundwork for future research identifying drivers of disparate exposure
RESEARCH
THE DATASET
ADVOCACY
Individual Exposure Disparities
Community Exposure Disparities
“Majority black communities experience higher emissions on average”
“Black and brown individuals are exposed to higher pollution on average”
Supporting
Advocacy
Our Objective: Demonstrate two applications of our dataset for local environmental advocacy groups to support existing efforts to monitor and reduce air pollution.
Application 1: Pollutant Estimation Dashboard - WIP
ADVOCACY
THE DATASET
RESEARCH
The Problem
Our Solution
Advocacy groups lack good data on air pollution to support their efforts to combat air pollution.
Rely on modeled emissions data
High resolution pollutant dashboard that can supplement advocacy efforts to reduce health and safety risks.
ADVOCACY
THE DATASET
RESEARCH
Application 2: Sensor Placement Algorithm
ADVOCACY
THE DATASET
RESEARCH
The Problem
Our Solution
Interpretable, unbiased, and customizable algorithm to determine placement of new ground-level NO2 monitors.
Lack of trust in the placement process of ground level sensors.
Lack of updated data to substantiate locations for new sensors
Align metrics to EPA considerations for existing sensors:
ADVOCACY
THE DATASET
RESEARCH
Summary
Develop two high-impact applications of our dataset that can be used to support advocacy
ADVOCACY
THE DATASET
RESEARCH
Advocate Needs
Our Solutions
Quickly identify regions with higher pollutant levels
Determine where to place sensors for greatest effect
Next Steps
Include additional sensors from Clarity Movement Co.
Include additional Criteria Air Pollutants (SO2, O3)
Present at Louisiana Environmental Conference and Trade Fair (hopefully!)
Creating
The Dataset
Supporting
Advocacy
Data
Change
Advancing Research
"This team's machine learning approach is an important part of identifying the sources of air pollution in Louisiana, and the communities that are impacted."
- Alex Kolker, Louisiana Universities Marine Consortium
Questions?
Questions?
To learn more, visit:
bit.ly/louisianaairpollution
Acknowledgements
D. Alex Hughes and Uri Schonfeld
University of California, Berkeley
All the wonderful people who spent their time meeting with us and responding to our many emails to support us with our project!
Alex Kolker
Associate Professor at Louisiana Universities Marine Consortium
References
Original slide template and icons: https://slidesgo.com/theme/air-pollution-infographics#search-Pollution&position-28&results-87
Clipart for data pipeline from https://publicdomainvectors.org/
Logo was made with the help of ChatGPT-4