Pollution due to atmospheric particulates is a major cause of respiratory illnesses such as asthma particularly in California’s Central Valley which contains five of the ten most polluted cities in the United States. Real-time monitoring of the size and density of airborne particles is important for providing health advisories. Traditional monitoring techniques provide only sparse, localized measurements. This proposal seeks seed funding to investigate terrestrial remote sensing as a novel low-cost wide-coverage alternate to current approaches for monitoring atmospheric particulates. In particular, hyperspectral imaging will be used to estimate the size and density of ground-level particulates based on light scattering principles. The technology could also find application in measuring the solar irradiance of widely dispersed photovoltaic energy sources.

2009 Update:

This project investigates the feasibility of using digital cameras for monitoring atmospheric particulates. California’s Central Valley suffers from some of the worst air pollution in the nation and yet there are only handful of monitoring sites that provide real-time measurements. Digital cameras of static scenes containing objects at varying distances (such as the Sierra Nevada mountains) represent a cost-effective alternate to specialized monitoring equipment. The key technical challenge is in developing image processing techniques for extracting quantitative features for estimating pollution levels, specifically particulate concentrations.

The guiding premise is that increased particulate pollution results in decreased visibility due to atmospheric scattering. We obtained one year of images taken every 15 minutes of two scenes of the Phoenix region from the Arizona Department of Environmental Quality visibility web camera project. We also obtained as ground truth data standard measurements of visibility taken every hour during the same period using transmissometers and nephelometers. We developed two image contrast measures that were shown to have a simple log-linear correlation with the ground truth data (as is expected from a simple model of atmospheric scattering). The first contrast measure is based on how visible the horizon is and is computed as the average difference between pixels in bands above and below the horizon. We explored the effect of different sized bands as well as different spectral channels (red, green, blue, or intensity). The second contrast measure is based on the amount of “detail” visible in objects at different distances and was computed using low-, band-, and high-pass filtering. As expected, this contrast measure showed higher correlation with the ground truth data for regions that were further from the camera. This work is being extended using more complex regression models and a semi-supervised learning framework (only one fourth of the Phoenix images have visibility measurements).

This seed project has allowed the PI to initiate a new research direction with potential for direct societal impact. The work is continuing along several fronts. We are analyzing images of the Merced region for which we have accompanying particulate concentration measurements. This includes several years of low-resolution images acquired from the UC Merced CatCam system as well as several months of images acquired using our own high-resolution capture systems. A modular static scene image processing pipeline (SSIPP) is under development that will allow different feature extractors and regression models to be applied in a unified framework. We are investigating hyperspectral imaging for increased spectral analysis. We plan to continue interacting with air pollution researchers at UC Davis as well as investigate deploying the technology with the aid of the California Air Resources Board.