PosterPDF Available

Abstract

Geolocation methods have been applied to electronic tagging data to estimate locations of groundfish species. Such information can improve stock assessments and fishery management plans that account for population structure, including movements across stock boundaries. Many popular geolocation methods have limitations including low horizontal resolution, flawed land boundary treatment, and long computational time. The particle filter is a state-space approach that has been applied to localization problems and addresses the aforementioned issues. We present a geolocation method based on the particle filter that is accelerated using graphics processing units (GPUs).
Geolocation methods have been applied to electronic tagging data to estimate
locations of groundfish species. Such information can improve stock
assessments and fishery management plans that account for population
structure, including movements across stock boundaries. Many popular
geolocation methods have limitations including low horizontal resolution, flawed
land boundary treatment, and long computational time. The particle filter is a
state-space approach that has been applied to localization problems and
addresses the aforementioned issues. We present a geolocation method based
on the particle filter that is accelerated using graphics processing units (GPUs).
Background
Archival Tagging
Particle Filter Geolocation
A particle filter geolocation method with hardware acceleration for demersal fish
Chang Liu and Geoffrey Cowles
School for Marine Science and Technology (SMAST), University of Massachusetts Dartmouth, New Bedford, MA
Fig. 1: Data storage tags attached to an Atlantic cod (left) and a yellowtail flounder (right).
Acknowledgements
We thank the Center for Scientific Computing and Visualization Research (CSCVR) at UMass Dartmouth for providing an
array of hardware for benchmarking and testing. Cod tagging research in the Spring Cod Conservation Zones was supported
by the United States Fish and Wildlife Service through the Sportfish Restoration Act and the Massachusetts Marine Fisheries
Institute. Funding for the research conducted as part of this manuscript was provided by NOAA Saltonstall-Kennedy Grant
award NA15NMF4270267.
!"#$%&'&$()**&#$%+(,$-()'./##+$0-12$(,)($'&3"'#$45&$2&'+&2$"6$(&57&')(8'&$)9#$
7'&228'&$#)():$;<$=2,$%&'&$'&3)7(8'&#:$
Fig. 2: Example of depth (blue) and temperature (red) data from a DST attached to an
Atlantic cod.
GPU Benchmark
>&$#&?&@"7&#$)$ABC.)33&@&')(&#$BD(,"9$7)'43@&$=@(&'$EBFG$*&"@"3)4"9$7)3H)*&$6"'$0-1$
#)()$E,I72JKK*+(:+"K?0L;MG:$
Release Predict
Update Resample
(a) (b)
(c) (d)
Fig. 3: Demonstration of the four steps of a particle filter iteration.
N$O'+&6$%"'HP"%$"6$BF$*&"@"3)4"9$)@*"'+(,5J$
Q: R9+4)@+S&$(,&$7)'43@&2$OD$'&@&)2+9*$(,&5$+9$(,&$'&@&)2&$@"3)4"9:$
T: U""7$"?&'$&)3,$#)D$(,&$=2,$+2$)($@)'*&J$
E)G B'&#+34"9J$5"?&$(,&$7)'43@&2$,"'+S"9()@@D$82+9*$)$')9#"5$%)@H:$
EOG C7#)(&J$%&+*,($(,&$7)'43@&2$OD$+9(&'7"@)49*$(,&$"O2&'?)4"9$@+H&@+,""#$
68934"9$("$(,&$7)'43@&2V$)9#$9"'5)@+S&$(,&$%&+*,(2:$
E3G W&2)57@&J$7)'43@&2$%+(,$@"%$%&+*,(2$)'&$'&5"?&#$)9#$'&7@)3&#$OD$(,"2&$
%+(,$,+*,&'$%&+*,(2:$
X: 0&(&'5+9&$(,&$5"2($7'"O)O@&$(')3H$OD$3)@38@)49*$(,&$)33858@)(&#$%&+*,($6"'$&)3,$
7)'43@&$)9#$2&@&3($(,&$"9&$%+(,$(,&$,+*,&2($)33858@)(&#$%&+*,(:
Fig. 4: Comparison of the time percentage for each step of
PF geolocation for the serial CPU (left bars) and GPU (right
bars) versions.
R9$BFV$&)3,$7)'43@&$+2$,)9#@&#$+9#&7&9#&9(@D$+9$(,&$5)+9$
@""7V$5)H+9*$+($?+)O@&$("$27&&#$87$(,&$3"578()4"9$82+9*$
7)')@@&@+S)4"9:$>&$3"57)'&#$(,&$7&'6"'5)93&2$"6$BF$
*&"@"3)4"9$O&(%&&9$!BC$)9#$ABCV$)9#$)3'"22$#+Y&'&9($
ABC2:$W&28@(2$28**&2($(,)($ABC$+57@&5&9()4"9$7'"?+#&2$
)77'"Z+5)(&@D$;[.TLL\$27&&#87V$')9$"9$#+Y&'&9($ABC2$6'"5$
)Y"'#)O@&$("$,+*,.&9#$5"#&@2:$1,&$(,'"8*,78($"9$(,&$2&'+)@$
ABC$3"#&$+2$L:;$#K,:
NVIDIA&GPU&Model
Single&Precision&
GFLOPS
PF-throughput&(d/h)
1&2@)$]Q^^$
Q;[__
Q`^<
1+()9$a
Q^Q<`
Q^_^
1&2@)$b[^
<<_Q
L<`
A&F"'3&$A1a$Q^<^
Q`XX
;[X
1&2@)$b;^3
;T_Q
;QX
A&F"'3&$A1a$<L^$1+
QTLX
XQ<
Figure 5: Total runtime for serial CPU (left bars) and GPU (right
bars) with speedup factors as a function of particle count.
Geolocation Results
Fig. 6: Progression of the daily posterior distribution (blue color rendering) and the most
probable track (black line) for two DST-tagged cod.
1,&$ABC.)33&@&')(&#$BF$*&"@"3)4"9$5&(,"#$%)2$)77@+&#$("$N(@)943$3"#$()**&#$
%+(,$0-12:$W&28@(2$28**&2($(,)($3"#$cQX$5+*')(&#$2"8(,%)'#$)d&'$()**+9*$)9#$
'&5)+9&#$+9$(,&$'&*+"9$9&)'$!)7&$!"#$6"'$)$7'"@"9*&#$7&'+"#$"6$45&$O&6"'&$
,&)#+9*$9"'(,$("$(,&$'&3)7(8'&$@"3)4"9$EF+*:$L)GV$)9#$(,)($3"#$cQ`$5+*')(&#$
"Y2,"'&$(,&9$9"'(,%)'#$("%)'#2$(,&$'&3)7(8'&$@"3)4"9V$#8'+9*$%,+3,$(,&$
7'"O)O+@+(D$#+2('+O84"9$#&5"92(')(&#$(%"$7@)82+O@&$(')M&3("'+&2$EF+*:$LOG:
(a) Cod #13
(b) Cod #17
ResearchGate has not been able to resolve any citations for this publication.
ResearchGate has not been able to resolve any references for this publication.