Ramseyer_C4ISRConf_Presentation - SUNYIT Computer ...

radiographerfictionData Management

Oct 31, 2013 (3 years and 8 months ago)

89 views

Open Architecture


for

Large Imaging Systems

of Hyperspectral Imaging Algorithms


Dr. George Ramseyer

Dr. Richard Linderman

AFRL/IF


Dr. Scott Spetka

ITT


18 October 2004

The Air Force Vision:

Battlespace Dominance through Information
Superiority

“Command and control
systems

based on
information and
communications
technology

and
precision
-
guided
munitions are critical to
all stages of the
Pentagon’s efforts to
transform itself to deal
with 21st century
threats.”




Paul Wolfowitz, DSECDEF

at AIAA lunch, 19 Feb 02

“The area with the greatest potential payoff... is in
C4ISR
... [to] ensure our commanders
have the
best information for rapid battlefield decision
-
making


Gen Richard B. Meyers, CJCS, SASC testimony, 5 Feb 02

Information Directorate

Mission

The advancement and application
of Information Systems Science and
Technology to meet Air Force
unique requirements for
Information Dominance and its
transition to air and space systems
to meet warfighter needs.

Vision:

Information Dominance for

Air and Space Superiority

Information Directorate S&T Program

Thrusts

Global Awareness



Acquires, exploits, fuses, and reasons over data/information


Provides tailored, consistent, superior situational knowledge


Sufficient precision to enable the decision process at all levels of
command

Dynamic Planning & Execution


Rapidly exploits superior, consistent knowledge of the battlespace


Faster, better informed, and more accurate decisions in complex
uncertain environments
-

Air, Space, Surface, Cyber


Shape and control the pace and phasing of engagements

Global Information Enterprise


Moves, processes, manages, and protects information supporting
GA and DP&E throughout the global information grid


Multiple military and commercial transmission media

Image Exploitation

Technical Goals


Enable

processing

resources

for

very

large

image

exploitation


Utilized

in

information
-
based

C
4
ISR

systems


Minimum

latency


High

speed

communications



Remote

parallel

processing

computers


An open systems architecture


User selects


Data sources


Exploitation time intervals


Parallelized exploitation method for execution



Resultant

imagery

products


Efficiently

disseminated

to

decision

makers


Stream

back

results

to

the

requestor

in

typical

web

prioritized

fashion
.


Framework


Code Function:



Integrated parallelized
imaging codes to
information systems


Approach
:


Integrate Framework with
BROADSWORD


Integrate Framework the
Joint Battlespace
Infosphere




Interfaced with

Hades Huinalu SKY

Rapidly process

new as well as
previously acquired raw imagery
data, so that a diverse and
distributed community of
intelligence analysts and battlefield
decision makers can take
appropriate actions, based upon
these analysis,

in near real
-
time
.



Joint Battlespace Infosphere

Information Management System



Publish


Subscribe


Query

Imaging Code Parallelization





Moderate Resolution Transmittance (MODTRAN)


Evolutionary Linear Mixing Algorithm (EVOLVE)


Pairwise Adaptive Liner Matched (PALM) Filter


Fast Line
-
of
-
sight Atmospheric Analysis of Spectral
Hypercubes (FLAASH)


Stochastic Expectation Maximization (SEM)


Reed
-
Xiaoli Anomaly Detector (R
-
X)


Automatic Target Detection (ATD) Principle
Component Analysis



Optical Identification System (ORASIS)




MODTRAN

(MODerate Resolution TRANSmittance)

Code Function:



Calculates Atmospheric
Transmittance & Radiance
for Frequencies from 0 to
50,000 cm
-
1

at Moderate
Spectral Resolution


Approach
:


Parallelization of
MODTRAN 4
-

Leveraging
previous JPL effort of
MODTRAN 3



Intensity (W/m2
-
Steradians
-
Wavenumber) vs. Wavenumber

Alpha Scalability


Runtime Improvement:
25.8

sec on 1 processor to
3.1

sec
on
16

processors


Ported to:


Hades



Huinalu

Evolutionary

Code Function:



Constrained Linear Spectral
Mixing Model


Modeled input spectral signature
as a linear combination of the
end
-
member constituents of
abundances (a
1,
a
2
,a
3
,…,a
N
),
sensor parameters (G
b
) and
atmospheric perturbations (O
b
)
.


Approach
:


Parallelization
-

Leveraged
Summer Faculty Program


Intensity vs. Band Number

Alpha Scalability


Runtime Improvement: 326 sec on 1 processor to 29 sec
on 16 processors


Ported to:


Hades


Huinalu

R
-
X

(Reed
-
Xiaoli Anomaly Detection)

Code Function:



Anomaly detection statistics


Based on Generalized Likelihood
Ratio approach


Detection theoretic basis results
in robust detector


Used as baseline for many
spectral applications.

Approach
:


Parallelization
-

Leverage
DARPA ASRP experience in
data
-
parallel implementation.

Alpha Scalability


Runtime Improvement: 2456 sec on 1 processor to 21 sec
on 128 processors


Ported to:

Longview


Hades


Huinalu

X
c

X
t

SEM

(Stochastic Expectation Maximization
-
Based Anomaly
Detection)

Code Function:



Iterative adaptive spectral
clustering


Characterize non
-
homogeneous
clutter based on a Gaussian
mixture model


Resulting classification information
is used to generate class map and
to compute anomaly detection
likelihood, based on Gaussian
mixture probabilities.

Approach
:


Global, iterative nature of
algorithm calls for multi
-
faceted
parallel approach.



Combines hierarchy of data
-
parallel and functional allocation
across nodes.

Alpha Scalability


Runtime Improvement: 1023 sec on 1 processor to 121
sec on
16

processors


Ported to:



Longview


Hades


Huinalu

Gaussian mixture model

Radiance Reflectance Reflectance


(no adjacency) (with adjacency)


Atmospheric Correction: FLAASH
-
C


(Fast Line
-
of
-
sight Atmospheric Analysis of Spectral
Hypercubes)

Code Function:



Determines/Removes Atmospheric
Contamination of Hyperspectral
Images


Returning Reflectance
-
at
-
Pixel


Approach
:


Complete revision of IDL
-
based
algorithm


Converted to C
++


Leveraging existing MODTRAN
-
based Look
-
Up Tables




Alpha Scalability by Module


Runtime Improvement:


20 sec on 2 processor to 4.4 sec on 32 processors (SKY)


Ported to:


SKY



Hades


Huinalu

SKY Timing Metrics

0.001

0.01

0.1

1

10

100

1000

1

10

100

Nodes (#)

Time (sec)

Misc

Sensor slit function

MOD GLUT Interp

MOD Spectra Conv

CUBE: Dist./Store

Band Merging

Water Retrieval

Cloud Masking

Adjacency Kernel

Adjacency Smooth

CUBE: Transpose

Reflectance Inv.

CUBE: Gather

Total

30% scaling

total ideal

Total (w/o mpi comm)

2

16

32

With enhanced HAZE


3

13

Basic Endmember Analyses

J. A. Gardner,
et al
., "Considerations In Atmospheric Compensation Of
Spectral Image Data," Proc. SPIE, 3756
-
38, 1999


7

ORASIS

(Optical Real time Analysis and Spectral

Identification System)


Supports data communication
and extraction of useful
information from large data sets.


Performs all of the following in
real time:


Retains infrequent spectra
(not statistically based).


Provides significant
compression with little
information loss.


Performs anomaly
detection and terrain
categorization.


Input

Spectrum

Spectral Libraries



Prescreener

Hyperspectral

Data Cube



+

+

+

+

.

.

.

.

.

.

.

.

.

.

Endmember

Determination



Exemplars

Filter Vector

Calculation

Redundant

Spectra

Creation of Abundance Maps by Spectral Unmixing.

Can unmix either full data set
OR

exemplars.

RGB

ORASIS Fraction Planes

Automatic Target Detection
Principal Components Analysis

Function:



Target detection reports,
detection images using
Multilayer Perceptron (MLP)
neural network trained on PCA
Eigenvectors.


Approach:


Data: Simulated


Spectral


N Sensors


N Sensors
Spectral/Spatial/Polarimetric


Training Modes: Serial


Offline


Online


Alpha Scalability


Ported to:


SGI ORIGIN, Hades, Huinalu

N
p

Detection
Report

Detection
Image

IRD

MLP

DET

HSI

PALM

Pairwise Adaptive Linear Matched Filter

Code Function:



Segments image into
background classes (ISMC)


Detects targets based on
signature library


Approach
:


Parallelization of ISMC


Multi
-
class target discrimination
components


Alpha Scalability



PALM Serial Run Times


10m 42s: Serial code with 1 processor and 4 PCs


8h 16m 59s: Serial code with 1 processor and 14 PCs



PALM Parallel Run Times


3m 25s: Parallel code with 32 processors and 4 PCs




Speed
-
up of 3.1


18m 27s: Parallel code with 32 processors and 14 PCs



Speed
-
up of 26.9



Ported to: Huinalu SUN E10000



High Performance Computers




COYOTE


-

52 processors in one chassis



-

2.66 GHz



-

6 GB/dual processor node



-

Gigabit ethernet



Heterogeneous HPC


-

96 processors

-

2.2 GHz

-

4 GB/dual processor node

-

Myrinet




Huinalu HPC


-

520 processors

-

900 MHz

-

1 GB/dual processor node

-

Myrinet



Brainerd, ARL IBM P3


-

512 processors


-

375 MHz


-

16 GB/ 16 processors


-

Colony switch


DREN

Huinalu

Cluster

Server

Hetero

HPC

Server

Rack 2 Kerberized

JBI Pub/Sub Broker

ARL MSRC

Brainerd

IBM P3

Rack 3

Kerberized

Framework

RAID

(Oracle

Database)

Rome

Users

Outside

Users

Coyote

HPC

Server

Huinalu

MHPCC

Test Framework

Rack 1 Kerberized

JBI Repository

1

2

3

N

Parallelized HIE

Applications

HIE Server

JOB Submittal

Result Retrieval

JBI Query

Cubes

JBI HPC Status

Subscriber

Parallel Hyperspectral Application Server

HPC Job

Launcher

Results

Manager

Cube

Publisher &

Subscriber

Status

Publisher

khttps

kftp

XML

kftp

NITF

NFS Mounted

Shared Disk

JBI

NITF

Cube on

Local Disk

Framework

App Interface

Program

MPI

Node

0

Disk I/O

CUBE

Chopper

Scatterer

Gatherer

Generator

NITF Output

App Specific

App Specific

Disk I/O

NITF

Parser

NITF Viewers

HPC/JBI/Server Interplay

Compute Time

Compute Time

Communication Time

Communication Time

Latency Time

Latency Time

Distributed Interactive HPC Testbed


Goal:

Assess the potential value and cost of
providing greater interactive access to HPC
resources to the DoD RDT&E community and its
contractors.


Means:

Provide both unclassified distributed HPC
resources to the DoD HPC community for interactive
experimentation exploring new applications and
system configurations

Technical Challenges


Low latency support for interactive and real
-
time
applications

proper HPC configuration?


Cohabitation of interactive and batch jobs?


Web
-
based access to network of HPC’s with enhanced
usability


Information management system supporting distributed
HPC applications


Demonstrating new C4ISR applications of HPC



Distributed HPC’s


Accessed by authorized users anywhere


Interactive and time critical problems


Distributed Interactive HPC Testbed

Legend


Remote Users


Networked HPC’s



Research and

Engineering Network

AFRL

Coyote


ARL

Powell

WPAFB

Mach 2

SSCSD

Seahawk


MHPCC

Koa

Cluster


High Performance Computers


Site

Computer

Memory and I/O

Online

ARL MSRC

Aberdeen,
MD

Powell: 128 node Dual 3.06MHz
Xeon Cluster

2 GB DRAM and 64 GB
disk/node, Myrinet &
GigEnet/ 100MB Backplane

Est. 10/04
w/batch;

4/05 share
with batch,

WPAFB

Dayton, OH

Mach2: 24 node Dual 2.66 GHz
Xeon, Linux


4 GB DRAM and 80 GB
disk/node , dual GigEnet


Est. 10/04



AFRL

Rome, NY

Coyote: 26 node Dual 3.06GHz
Xeon, Linux


6 GB DRAM and 400 GB
disk/node, dual GigEnet


Yes



SSCSD

San Diego,
CA

Seahawk: 16 node 1.3GHz
Itanium2, Linux

2 GB DRAM and 36 GB
disk/node, dual GigEnet


Est. 12/04


MHPCC

Maui, HI

Koa: 128 node dual Xeon, Linux
(system moves between
environments)

4 GB DRAM and 80 GB
disk/node, shared file
system, dual GigEnet

Yes

Applications and Experiments



HPC Information Management


Grid
-
Based Collaboration


Interactive parallel MATLAB


Image exploitation of large databases


Interactive electromagnetics simulation


Conclusions


Developing

and

demonstrating

support

for

the

rapid,

low

latency

exploitation

of

hyperspectral

imagery

information
.



Flexible

and

open

framework

to

service

exploitation

requests

from

C
4
I

users
.


Taps

into

large,

dynamic

databases

of

previously

processed

and

raw

hyperspectral

data

to

deliver

products

to

the

requester

with

minimal

latency
.



Web
-
based

interface

so

that

any

authorized

user

with

a

web

browser

can

input

requests
.



User

selects

data

sources,

high

performance

computer,

and

a

parallelized

exploitation

method

for

execution
.



Algorithm

parallelization

efforts

achieve

minimal

latency

before

initial

results

begin

to

stream

back

to

the

requestor

in

typical

prioritized

fashion
.