|
|
|
Massively Parallel Imagery Assimilation Using the 3D Multiscale Multicomponent Modeling Framework (MMMF)
|
|
|
| This proposal will increase by several orders of magnitude the use of EOS data in numerical weather cloud resolving models at the
synoptic and global scales by extracting EOS satellite image-derived atmospheric parameters using high-throughput parallel stream
processing algorithms implemented on inexpensive graphics processing units (GPUs). Rapid image processing will be coupled to a
massively parallel Grid-based process to assimilate image derived data into the Multiscale Multicomponent Modeling Framework
(MMMF) that combines for the first time cloud resolving models within a global circulation model for hurricane studies. The objective
of this proposal involves two components. One is to develop new computational image analysis and computer vision techniques for
satellite wind and microphysics extraction using highly parallelized implementations on inexpensive GPUs (graphics processing units)
for rapid processing of large volumes of satellite imagery. The second is to assimilate atmospheric fields such as wind and microphysics
parameters into the MMMF for hurricane studies, on NASA Grid-clusters, using a multiscale approach that embeds the fine resolution
Goddard Cumulus Ensemble (GCE) for cloud scale parameters within the coarse resolution Goddard finite volume Global Circulation
Model (fvGCM) for global coupling of dynamic quantities. Approximately 13,104 GCE models are embedded, one each per 2x2.5
degree grid element of the fvGCM and code parallelization using the Message Passing Interface (MPI) shows extremely encouraging
speedups and scalability. A massively parallel version of the 3D GCE model using MPI for the 10,240 processor Columbia
supercomputer at NASA Ames based on the current 512 CPU implementation when operational will be used in this project. A testbed
system that integrates experimental atmospheric wind and cloudiness extraction from multiple satellite imagery data with multiscale
modeling using the MMMF will be evaluated using hurricane data from the recent intense hurricane seasons. We propose new
approaches for semi-fluid cloud tracking, estimating cloud structure from multiview multispectral imagery, robust tensor-based optical
flow, adaptive data thinning and spline-based 3D scene flow motion estimation. The proposed novel computational approaches using
both inexpensive GPUs and large scale Grid computing based clusters will be utilized in this project for satellite image analysis and
data assimilation to evaluate impact on hurricane prediction and seasonal forecasts using the MMMF running on multi-resolution
spatial grids. |
|
|
