Algorithm Status: 1. Core Algorithm is developed and performs well: - uses very elaborated aerosol and RT models; - based on rigorous statistical optimization; - performs well in numerical test (Dubovik et al. 2011, Kokhanovsky et al. 2010); - has a lot of flexibility for constraining retrieval: both for single-pixel and/or multi-pixel scenarios) 2. Issues: - too long - 10 sec per 1 pixel!!! - needs to be optimally set for operational processing - cloud – screening – need to be improved !!! Described in Dubovik et al., AMT, 2011 Main Objective: to make algorithm practical Réunion Parasol-Calcul-Tosca au CNES, PARIS , 10 février, Paris
15
Embed
Algorithm Status: 1.Core Algorithm is developed and performs well: - uses very elaborated aerosol and RT models; - based on rigorous statistical optimization;
CORE INVERSION BLOCK of PARASOL level 1 cloud-screened DATA N x ✕ N x ✕ N t Climatology of surface reflectance Results in pixels – “ neighbours ” (if available) INITIAL GUESS Strength of a priori constraints: single-pixel and multi-pixel “Science” Algorithm ESA, Frascati, Italy, June 26, 2012
Welcome message from author
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
Algorithm Status:
1. Core Algorithm is developed and performs well: - uses very elaborated aerosol and RT models; - based on rigorous statistical optimization; - performs well in numerical test (Dubovik et al. 2011, Kokhanovsky et al. 2010); - has a lot of flexibility for constraining retrieval: both for single-pixel and/or multi-pixel scenarios)
2. Issues: - too long - 10 sec per 1 pixel!!!
- needs to be optimally set for operational processing - cloud – screening – need to be improved !!!
Described in Dubovik et al., AMT, 2011
Main Objective: to make algorithm practical
Réunion Parasol-Calcul-Tosca au CNES, PARIS , 10 février, Paris
Pressing needs:
1. Development of the FRAMEWORK
2. Accelerating the performance of the core inversion algorithm
!!!
Algorithm developments
Réunion Parasol-Calcul-Tosca au CNES, PARIS , 10 février, Paris
CORE INVERSION
BLOCK ofPARASOL level 1 cloud-screened DATA Nx ✕ Nx ✕ Nt
Climatology of surface reflectance
Results in pixels – “neighbours” (if available)
INITIAL GUESS
Strength of a priori constraints:
single-pixel and
multi-pixel
“Science” Algorithm
ESA , Frascati, Italy, June 26, 2012
Decomposition of the processing area
Parallel processing of PARASOL data
o The zones should have the same number of pixels;
o The zones should be coherently arranged in time and space;
o The climatology and retrieval data storage should have similar zone structure
zone of Nx× Ny× Nt pixels
o The zones can be treated independently; o A special treatment will be needed for the edges of zones.
ESA , Frascati, Italy, June 26, 2012
Sub-zones Nx× Ny are treated independently using “core
inversion”
t
CORE INVERSION
Nx× Ny = 100 x 100 ?
Nt >>1 (for simplicity)
Nt = 1? (for simplicity)
ESA , Frascati, Italy, June 26, 2012
Sequential processing block-by-block of Nx x Ny x Nt zones
Climatology
Storage of the retrieval resultsblock p block p+1
Nx x Ny zone
Accelerating the performance of core inversion algorithm:
Optimizing the initial conditions (FRAMEWORK);
Optimizing sparse matrix operation and inversion;
Benefitting from parallel programming;
Using common memory arrays, etc. (FRAMEWORK ?)
Optimizing the performance of radiative transfer (OS), that includes:
- finding trade-off between speed and accuracy; - avoiding re-calculation of the same variables;
- optimizing the land surface reflectance modeling;
- parallelizing some integrations in the code
ESA , Frascati, Italy, June 26, 2012
« AERONET like » statistically optimized « no look-up tables » inversion