TL;DR
This paper introduces a novel data-driven approach, mrDMDcc, to optimize the placement of PM$_{2.5}$ air quality sensors. By incorporating socioeconomic data (race and income) alongside air pollution variability, the method demonstrably shifts sensor deployment towards historically underserved and polluted communities, addressing environmental injustices.
What Problem Does This Paper Address?
Air pollution, particularly fine particulate matter (PM$_{2.5}$), poses a significant public health risk. While regulatory monitoring networks exist, they often have spatial gaps. Low-cost sensor networks have emerged to fill these gaps, but they are frequently deployed in higher-income, predominantly White neighborhoods. This uneven distribution can lead to inaccurate exposure assessments and exacerbate existing health disparities, particularly for minority and low-income communities who often face higher pollution burdens. The paper aims to solve this by developing a method to strategically place sensors to ensure equitable monitoring coverage in urban areas with a history of racial and income segregation.
Methods
The core of the paper's methodology is the multiresolution dynamic mode decomposition with cost constraints (mrDMDcc). This approach builds upon the established mrDMD technique, which decomposes complex spatiotemporal data (like PM$_{2.5}$ concentrations) into dominant modes representing different spatial patterns and their temporal dynamics across various timescales (from days to years).
The novelty lies in the integration of a cost function into the mrDMD optimization process. This cost function penalizes sensor placements that are too far from neighborhoods with high proportions of nonwhite residents or low median annual incomes. The algorithm then uses a greedy QR pivoting approach to identify sensor locations that balance capturing high PM$_{2.5}$ variability with addressing these environmental justice metrics. This is a departure from traditional sensor placement methods that solely focus on air pollution data or random distribution.
Datasets
- PM$_{2.5}$ Data: Modeled daily PM$_{2.5}$ concentrations at a 1 km x 1 km resolution for the contiguous United States from January 2000 to December 2016. This dataset was generated through data fusion of surface monitoring, satellite data, land-use information, and chemical transport models.
- Census Data: Sociodemographic features from the 2020 American Community Survey, specifically focusing on the proportion of nonwhite individuals and median annual household income at the census tract level. This data was interpolated to match the 1 km x 1 km grid of the PM$_{2.5}$ data.
Key Insights and Ideas
- Socioeconomic Data is Crucial for Equitable Monitoring: Simply optimizing for PM$_{2.5}$ variability alone leads to sensor distributions that mirror existing biases, often neglecting areas with the greatest need.
- mrDMDcc Effectively Shifts Sensor Placement: The mrDMDcc algorithm demonstrably directs more sensors to historically marginalized communities compared to methods that only consider air pollution data.
- Income vs. Race Optimization Nuances: In some cities, optimizing for income captured more nonwhite and low-income neighborhoods than optimizing for race alone, suggesting that income data might be more tractable for the algorithm in certain segregation patterns.
- Legacy of Segregation Matters: The "shape" of segregation within a city significantly impacts how the algorithm distributes sensors, highlighting the complex interplay of historical policies and current environmental burdens.
- A Roadmap for Environmental Justice: The study provides a practical, data-driven framework for urban planners and environmental agencies to design more equitable air quality monitoring networks.
Key Results
- In St. Louis, the mrDMDcc algorithm, when considering race, shifted sensors to the historically Black East St. Louis neighborhood, an area with significant environmental justice issues. The income-optimized mrDMDcc also directed more sensors to low-income northern suburbs.
- Houston's race- and income-optimized networks focused on the Ship Channel region and majority Black neighborhoods, areas disproportionately affected by industrial pollution.
- Buffalo showed similar sensor distributions across all methods, but the mrDMDcc highlighted the Niagara Falls and Love Canal neighborhoods, which have long histories of environmental injustice.
- Boston's income-optimized mrDMDcc captured more nonwhite and low-income neighborhoods than race-optimized or pollution-only methods, though the race optimization also directed sensors to South Boston.
- Cumulative distribution plots confirmed that mrDMDcc optimizations captured a higher density of nonwhite and low-income residents compared to the mrDMD approach alone.
Limitations and Future Work
- Sensitivity to Cost Function Parameters: The authors note that the mrDMDcc algorithm is sensitive to the chosen values for the cost function (γ) and the number of sensors. Determining the optimal values for these parameters may require further investigation.
- Data Binning: For cities with complex segregation patterns (like Boston), binning census data into coarser statistical groups might improve optimization, though this could risk averaging out the most socioeconomically disparate communities.
- Incorporating Sensor Accuracy: Future work could extend the framework to incorporate sensors with varying levels of accuracy (high-cost vs. low-cost) and explore other relevant metrics like population density or land-use information.
- Systematic Approach for Parameter Selection: A more systematic approach is needed to determine the appropriate γ value and the ideal number of sampling sites.
- Application to Existing Networks: The framework could be applied to existing EPA and PurpleAir networks to identify optimal locations for new sensor placements.
Summary updated