Your query was:

cervone

HR: 11:20h
AN: IN22A-05
TI: Combining Machine Learning and Mesoscale Modeling for Atmospheric Releases Hazard Assessment
AU: * Cervone, G
EM: gcervone@gmu.edu
AF: George Mason University, 4400 University Dr MS 6C3, Fairfax, VA 22030, United States
AU: Franzese, P
EM: pfranzes@gmu.edu
AF: George Mason University, 4400 University Dr MS 6C3, Fairfax, VA 22030, United States
AU: Ezber, Y
EM: ezber@itu.edu.tr
AF: George Mason University, 4400 University Dr MS 6C3, Fairfax, VA 22030, United States
AU: Ezber, Y
EM: ezber@itu.edu.tr
AF: Eurasia Institute of Earth Sciences, Istanbul Technical University, Istanbul, 34469, Turkey
AU: Boybeyi, Z
EM: zboybeyi@gmu.edu
AF: George Mason University, 4400 University Dr MS 6C3, Fairfax, VA 22030, United States
AB: In applications such as homeland security and hazards response, it is necessary to know in real time which areas are most at risk from a potentially harmful atmospheric pollutant. Using high resolution remote sensing measurements and atmospheric mesoscale numerical models, it is possible to detect and study the transport and dispersion of particles with great accuracy, and to determine the ground concentrations which might pose a threat to people and properties. Satellite observations from different sensors must be fused together to compensate for different spatial, temporal and spectral resolutions and data availability. Such observations are used to initialize and validate atmospheric mesoscale models, which can provide accurate estimates of ground concentrations. Such numerical models are, however, usually slow due to the complex nature of the computations, and do not provide real time answers. We will define probability maps of risks by running several atmospheric mesoscale and T&D simulations spanning the climatological input conditions of an entire year, observed using high resolution remote sensing instruments. Such maps provide an immediate risk assessment area associated with a given source location. If a release indeed occurs, the computed risk maps can be used for first assessment and rapid response. We analyze the output of the mesoscale model runs using machine learning algorithms to find characteristic patterns which relate potential risk areas with atmospheric parameters which can be observed using remote sensing instruments and ground measurements. Therefore, when a release occurs, it is possible to give a quick hazard assessment without running a time consuming model, but by comparing the current atmospheric conditions with those associated with each identified risk area. The offline learning provides knowledge that can later be used to protect people and properties.
DE: 0305 Aerosols and particles (0345, 4801, 4906)
DE: 0555 Neural networks, fuzzy logic, machine learning
DE: 1622 Earth system modeling (1225)
DE: 3235 Persistence, memory, correlations, clustering (3265, 7857)
SC: Earth and Space Science Informatics [IN]
MN: 2007 Fall Meeting