HR: 0800h
AN: A21B-0853    [Abstracts]
TI: Using MISR and MODIS Data For Detection and Analysis of Smoke Plume Injection Heights Over North America During Summer 2004
AU: * Mazzoni, D
EM: Dominic.Mazzoni@jpl.nasa.gov
AF: Jet Propulsion Laboratory, California Institute of Technology, 4800 Oak Grove Drive, Pasadena, CA 91109 United States
AU: Tong, L
EM: Lingling@mail.utexas.edu
AF: Department of Computer Sciences, University of Texas at Austin, 1 University Station, Austin, TX 78712 United States
AU: Diner, D
EM: David.Diner@jpl.nasa.gov
AF: Jet Propulsion Laboratory, California Institute of Technology, 4800 Oak Grove Drive, Pasadena, CA 91109 United States
AU: Li, Q
EM: Qinbin.Li@jpl.nasa.gov
AF: Jet Propulsion Laboratory, California Institute of Technology, 4800 Oak Grove Drive, Pasadena, CA 91109 United States
AU: Logan, J
EM: jal@io.as.harvard.edu
AF: Division of Engineering and Applied Sciences, Harvard University, 29 Oxford Street, Cambridge, MA 02138 United States
AB: We aim to investigate the relationship between climate, fires, and air quality in order to predict the effect of possible future climate changes on North American air quality. To initialize and calibrate chemistry transport models such as GEOS-Chem, we are using satellite imagery to gather statistics about smoke plumes from fires over North America for a five-year period. Statistics include the location, extent, and injection height of each plume, along with local weather, topography, and surface conditions from which correlative relationships can be derived. Stereoscopic height retrievals from NASA's Multi-angle Imaging SpectroRadiometer (MISR) are our primary data for establishing the plume injection heights. Oblique imagery from MISR also helps identify plumes that are more difficult to detect in nadir views. MODerate-resolution Imaging Spectroradiometer (MODIS) data help pinpoint fire locations. We use automated data mining techniques to identify smoke plumes and limit the amount of analysis that must be done by hand. We report here on analysis of data from June - September 2004, which we are using to refine our techniques. A Support Vector Machines (SVM) pixel classifier which uses spectral, angular, and textural features from MISR identifies pixels that contain smoke. Our results show that this classifier is 75% accurate at finding smoke at the 1.1-km pixel level, and catches some smoke pixels in any given plume with nearly 100% certainty. Next, we match areas containing smoke with fire locations identified by MODIS. For candidate scenes that appear to contain both smoke and fire, we apply machine vision techniques to look for evidence of plume-like shapes. When potential plumes are found, we automatically estimate source location, orientation, and injection height, using histograms of MISR stereo data for the latter. Lastly, a human expert examines all examples found and discards any false retrievals. Exhaustive analysis of our technique applied to data from 2004 demonstrates successful identification of a majority of smoke plumes, without a low rate of false positives. Visualizations of statistical results derived from data collected during the summer of 2004 will be presented.
DE: 0305 Aerosols and particles (0345, 4801, 4906)
DE: 0555 Neural networks, fuzzy logic, machine learning
DE: 3305 Climate change and variability (1616, 1635, 3309, 4215, 4513)
SC: Atmospheric Sciences [A]
MN: Fall Meeting 2005