Next-generation OMI SO 2 Retrieval Algorithm based on Principal Component Analysis Can Li 1,2 , Joanna Joiner 2 , Nick Krotkov 2 , P. K. Bhartia 2 1 ESSIC, University of Maryland College Park 2 NASA GSFC 18 th OMI Science Team Meeting KNMI, De Bilt, The Netherlands March 13, 2014 Email: [email protected]
20
Embed
Next-generation OMI SO 2 Retrieval Algorithm based on Principal Component Analysis
Next-generation OMI SO 2 Retrieval Algorithm based on Principal Component Analysis. Can Li 1,2 , Joanna Joiner 2 , Nick Krotkov 2 , P. K. Bhartia 2 1 ESSIC, University of Maryland College Park 2 NASA GSFC 18 th OMI Science Team Meeting KNMI, De Bilt , The Netherlands March 13, 2014. - PowerPoint PPT Presentation
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Next-generation OMI SO2 Retrieval Algorithm based on Principal
Component AnalysisCan Li1,2, Joanna Joiner2, Nick Krotkov2, P. K. Bhartia2
1ESSIC, University of Maryland College Park2NASA GSFC
18th OMI Science Team MeetingKNMI, De Bilt, The Netherlands
Outline• Background and Motivation• Methodology (Framework)• Application to OMI– Results (Planetary Boundary Layer SO2)
– Results (Volcanic SO2)
• Data Continuity: Comparison of OMI and OMPS
• Next Steps and Conclusions
Background and Motivation
•Motivation: Band Residual Difference (BRD) algorithm fast and sensitive, but large noise and artifacts (only 3 pairs of wavelengths)•Objective: develop an innovative approach to utilize the full spectral content from OMI while maintaining computational efficiency
Operational OMI SO2 (Sept. 2004 – Feb. 2008)
Basis – Spectral fitting algorithms
First look at the DOAS Equation:Measured sun-normalized radiances
Various gas absorbers (O3, SO2 etc.)
Rayleigh and Mie scattering, surface reflectance etc.
The Ring effect
Plus additional measurement artifacts terms (e.g., wavelength shift, stray light, etc.) and/or radiance data
correction schemes
Utilization of the full spectral content, but some terms are difficult to model (e.g., RRS)
Methodology (Framework): PCAInstead of explicit modeling of ozone, RRS, and other
instrumental features, we use a data-driven approach based on principal component analysis (PCA) with spectral fitting
Measured N-value spectrum
PCs from SO2-free regions, (O3 absorption, surface reflectance, RRS, measurement artifacts etc.) other than SO2 absorption
Pre-calculated SO2 Jacobians (assuming O3
profiles, albedo, etc.)
SO2 column amount
Fitting of the right hand side to the spectrum on the left hand side -> SO2 column amount and coefficients of PCs
(See Guanter et al., 2012; Joiner et al., 2013; Li et al., 2013)
Application to OMI•Spectral window: 310.5-340 nm – avoid stray light at shorter wavelengths
•Each row (scan position) processed individually – different characteristics between different rows of the 2-D CCD
•Each swath processed individually – account for orbit-varying dark current
•# of PCs determined dynamically – exclude SO2-related PCs and avoid overfitting by checking the correlation between PCs and SO2 Jacobians
•1st step: Simple Jacobians similar to those used in operational BRD algorithm for straight-forward comparison
Step 1 Ps
See Li et al., [GRL, 2013] for details
Principle Components and Residuals
(a-c) First few PCsBlue line: scaled reference Ring spectrum
(d) Least squares fitting residuals for a pixel near Hawaii
(Var.% 99.8492)
(Var.% 0.1264)
Example PCs from entire row # 11, Orbit 10990
(Var.% 0.0217)
(Var.% 5.32E-5)
(Var.% 4.79E-5)
PC #1: Mean spectrum
PC #2: O3 absorptionPC #3: Surface reflectance (also Ring signature)
•PCA algorithm reduces retrieval noise by a factor of two as compared with the BRD algorithm•SO2 Jacobians for PCA algorithm calculated with the same assumptions as in the BRD algorithm
August, 2006
OMI operational BRD
Results: Boundary layer pollution SO2
eastern U.S., August 2006
PCA Operational BRD
PCA algorithm reveals major SO2 point sources (circles), with much reduced noise and artifacts.
Ex. Sudbury, Canada (~220 kt in 2006) Ex. analyzed with pixel averaging (super sampling) reveals
details of emission sources [e.g., Fioletov et al., 2011]
PCA, 2006 BRD, 2006 only BRD, 2004-2012
•One year’s worth of PCA retrievals yield results similar to that from 3-5 years worth of BRD data. •Global survey shows that PCA SO2 removes most artifacts in BRD data without significantly altering signals from real sources (Fioletov and McLinden, personal communication)
Largely hidden by artifacts
Volcanic SO2: Kasatochi eruption August 7-8, 2008
For volcanic SO2, nonlinearity due to saturation at shorter wavelengths• Iteration of SO2 Jacobians (pre-calculated assuming loadings of 1 - 500 DU)• Shift of spectral fitting window to longer wavelengths
PCA closest to estimated released SO2 mass of ~2200 kt based on observed decay of SO2 [Krotkov et al., 2010]
Transport of the plume
Ln(SO2)
August 10, 2008
August 11, 2008
August 12, 2008
•Our algorithm eliminates the need for explicit instrument-specific radiance correction schemes
•Test on OMPS: minimal changes to algorithm –biggest is the use of OMPS slit function for Jacobians–spectral window, etc., same as in OMI
•Reduces the chance of introducing artifacts/biases between different instruments
Comparison with OMPS
OMI and OMPS comparison
OMPS and OMI PCA SO2 retrievals show good agreement despite somewhat different sampling
October, 2013
OMI and OMPS comparison
Both OMI and OMPS PCA SO2 retrievals show enhanced SO2 loading over northern China in January 2013, when severe pollution attracted media and public attention.
OMPS, Jan. 2013 OMI, Jan. 2013
OMI and OMPS comparison
OMI and OMPS PCA SO2 data show similar seasonal patterns and SO2 signals over eastern India (several coal-fired power plants built in recent years) [Lu et al., 2013].
Next Steps• Expanded table for SO2 Jacobians to more accurately account for
• Addition of scattering weight to output to allow convenient adjustment of SO2 column amount based on user-provided profile
• Inclusion of error estimates
• Operational implementation, public release – >1 year processed and currently under evaluation– Initial release for boundary layer pollution this year– Improved Jacobians and volcanic data to follow soon– On 12 CPUs, 1-2 days to process a year of OMI data
Conclusions• Significant improvements in retrieval quality – PCA algorithm
uses full spectral content from OMI and similar instruments offering increased temporal resolution and source detection
• Computation efficiency – over an order of magnitude faster than comparable spectral fitting algorithms; increasingly important given the greater data volumes expected from future missions (e.g., TROPOMI, TEMPO)
• Maximal data continuity between instruments – no need to develop instrument-specific radiance data correction schemes
• Flexibility – fitting window can be easily adjusted to optimize sensitivity under different conditions
Backups
Results: Daily boundary layer SO2August 13, 2006 August 14, 2006