This project connects computing with research to advance community management of environments and support environmental sustainability. The project team has investigated several ideas to enable more accurate estimation and prediction of pollutant concentrations at finer spatial and temporal granularities than previously possible. The work has taken two main paths: developing new, fine-grained pollution sensing methods and developing methods to intelligently integrate data from disparate sources.

The project has developed new methods of determining airborne pollution concentrations at high spatial resolution. The key idea is to use detailed physical models of the effects of pollutants on light scattering and absorption. Using normal photographs, it is possible to determine the types and concentrations of pollutants by observing their influence on visible haze. This makes it possible to produce spatially fine-grained pollution maps from single images, instead of producing a single pollution estimate per image as was the case for prior work. By accounting for wavelength-dependent scattering and absorption properties of specific pollutants (e.g., compared to black carbon particles, important ozone precursors preferentially scatter blue light), concentrations can be determined even when multiple pollutants are present, for the first time. Combined, these approaches also improve pollution estimation accuracy by 22% compared with the best prior work.

The project team has also developed tools and techniques to collect, process, and analyze data from online resources and in-field studies, as well as fine-grained chemical transport model simulations. It studied correlations among surface ozone concentrations and other relevant features to develop a new prediction model. This model uses Long Short-Term Memory to train and predict dynamic traffic and meteorological features, which are combined with static local environmental features and fed through a Support Vector Machine based regression model to predict community-level ozone concentration in the next three days. The approach uses multi-source data, especially high-frequency grid-based weather data, to model air pollutant dynamics. Convolution operators on grid weather data capture the impacts of various weather parameters on air pollutant variations. Cross-domain features are automatically grouped based on their correlations. Finally, a multi-group Encoder-Decoder networks is used to fuse multiple feature groups, enabling 9%-16% improvement in pollution prediction accuracy relative to the best existing work.

The project has broadened and advanced vertical research and education integration across information technologies and environmental science and engineering. Key research results have been published at relevant scientific venues and integrated into the courses taught by the PIs. Through the Air Quality Inquiry Program, students have been trained to use air quality sensors to answer questions of personal interest. The students then take the training and the sensors to rural schools in Colorado to work with environmental science high school classes. That program has impacted over 500 high school students.

Last Modified: 12/13/2019
Modified by: Robert P Dick