TL;DR
This paper introduces a novel data-driven approach using multiresolution dynamic mode decomposition (mrDMD) to identify optimal locations for PM2.5 air quality sensors across the contiguous United States. The findings reveal a significant shift in high-variability regions towards the West, suggesting current EPA monitoring networks may be suboptimal for capturing emerging pollution patterns, particularly those driven by wildfires.
What Problem Does This Paper Address?
Measuring ambient air pollution, specifically fine particulate matter (PM2.5), is crucial for public health, but current monitoring networks are expensive to maintain and may not be optimally placed to capture the full spectrum of pollution variability. Existing studies on sensor placement are often limited to small urban areas or short time scales. This research addresses the challenge of determining the most effective placement of a minimal number of sensors to maximize spatiotemporal coverage of PM2.5 pollution across the entire contiguous United States, considering its complex multiscale dynamics.
Methods
The core of this study is the application of Multiresolution Dynamic Mode Decomposition (mrDMD). This technique is an extension of Dynamic Mode Decomposition (DMD) that recursively decomposes a dataset into low-rank spatial modes and their associated temporal dynamics. Unlike standard DMD or Principal Component Analysis (PCA), mrDMD is designed to capture both short-term (daily) and long-term (decadal) variability, and importantly, it can incorporate transient phenomena like wildfires and temperature inversions that might otherwise be averaged out or discarded.
The mrDMD algorithm works by progressively removing low-frequency signals, allowing for the separation of background pollution levels from high-pollution episodes. This process generates a library of modes that represent distinct spatiotemporal patterns of PM2.5 concentration.
To identify optimal sensor locations from these modes, the researchers employed QR pivoting. This is a greedy algorithm that efficiently selects a subset of spatial locations (sensors) that best represent the dominant modes identified by mrDMD. Essentially, it identifies locations with the highest modal frequencies, indicating areas of greatest variability and significant pollution events.
Datasets
The study utilizes a daily PM2.5 concentration dataset for the contiguous United States from January 2000 to December 2016. This dataset was generated through a data fusion method combining surface monitoring measurements, satellite aerosol optical depth, land-use data, and chemical transport model results. The original resolution of 1 km x 1 km was coarsened to 10 km x 10 km for this analysis.
Key Insights and Ideas
- Wildfires are a major driver of PM2.5 variability, especially in the West: The mrDMD analysis strongly highlights wildfire events as significant contributors to PM2.5 pollution modes, particularly in the western United States.
- A geographical shift in pollution drivers: The study suggests that while historical EPA monitoring networks might have been designed for eastern pollution patterns prevalent in the 1980s-1990s, current PM2.5 variability is increasingly driven by western phenomena like wildfires and unique meteorological conditions in rugged terrain.
- Suboptimal current sensor distribution: The mrDMD-identified optimal sensor locations are significantly denser in the western US compared to the current EPA network, indicating potential gaps in monitoring high-variability regions.
- The San Joaquin Valley as a critical monitoring area: This region consistently emerges as a dominant area for high PM2.5 variability, driven by agricultural activities, wildfires, and temperature inversions, yet it is underserved by current EPA monitoring.
Key Results
- The mrDMD framework identified 1369 optimal sensor locations across the contiguous United States.
- 53% of the proposed optimal sensor locations are west of the 100th meridian, compared to only 32% of the current EPA network.
- The analysis identified 72 significant pollution modes in the first time window (Jan 2000-Mar 2011) and 77 in the second (Sep 2005-Dec 2016).
- Wildfire-related modes constituted a substantial portion: 38% of modes in the first window and 60% in the second.
- The study identified specific areas needing more sensors, including the San Joaquin Valley, northern California, and the Pacific Northwest.
Limitations and Future Work
- Assumption of representativeness: The study assumes that PM2.5 spatiotemporal patterns from 2000-2016 are representative of future conditions. While the authors argue this is reasonable due to the expected continued impact of wildfires, this is a key assumption.
- Daily timescale limitations: The mrDMD algorithm, with its chosen tolerance, did not identify significant pollution modes on timescales less than a week. Tightening this tolerance could reveal daily patterns but might also introduce more noise.
- Future Work: The authors suggest that their framework can be extended to regional and urban scales using higher-resolution datasets. They also propose incorporating different sensor types (high-cost and low-cost) and cost-constraining functions based on factors like topography, population density, and environmental justice metrics. The framework could also be used to inform stakeholders about the benefits and hazards of prescribed fires.
Summary updated