Multivariate approach for the retrieval of phytoplankton size
structure from measured light absorption
spectra in the Mediterranean
Sea (BOUSSOLE site)
Emanuele Organelli, Annick Bricaud, David Antoine, and Julia Uitz
Emanuele Organelli, Annick Bricaud, David Antoine, and Julia Uitz, "Multivariate approach for the retrieval of phytoplankton size structure from measured light absorption spectra in the Mediterranean Sea (BOUSSOLE site)," Appl. Opt. 52, 2257-2273 (2013)
Models based on the multivariate partial least squares (PLS) regression technique
are developed for the retrieval of phytoplankton size structure from measured
light absorption spectra (BOUSSOLE site, northwestern Mediterranean Sea).
PLS-models trained with data from the Mediterranean Sea showed good accuracy in
retrieving, over the nine-year BOUSSOLE time series, the concentrations of total
chlorophyll [Tchl ], of the sum of seven diagnostic pigments and
of pigments associated with micro, nano, and picophytoplankton size classes
separately. PLS-models trained using either total particle or phytoplankton
absorption spectra performed similarly, and both reproduced seasonal variations
of biomass and size classes derived by high performance liquid chromatography.
Satisfactory retrievals were also obtained using PLS-models trained with a data
set including various locations of the world’s oceans, with however a
lower accuracy. These results open the way to an application of this method to
absorption spectra derived from hyperspectral and field satellite radiance
measurements.
You do not have subscription access to this journal. Cited by links are available to subscribers only. You may subscribe either as an Optica member, or as an authorized user of your institution.
You do not have subscription access to this journal. Figure files are available to subscribers only. You may subscribe either as an Optica member, or as an authorized user of your institution.
You do not have subscription access to this journal. Article tables are available to subscribers only. You may subscribe either as an Optica member, or as an authorized user of your institution.
You do not have subscription access to this journal. Equations are available to subscribers only. You may subscribe either as an Optica member, or as an authorized user of your institution.
Cruises, Location, Sampling Period, Number of Samples
() and [Tchl ] range
for the First Optical
Depth, for the Data Used to Train Models
Cruise
Location
Period
[Tchl ] Range
EUMELI 3
Tropical North Atlantic
Oct. 1991
5
0.073–0.340
FLUPAC
Equatorial and subequatorial Pacific
Sep.–Oct. 1994
11
0.039–0.236
OLIPAC
Equatorial and subequatorial Pacific
Nov. 1994
34
0.072–0.291
MINOS
Eastern and western Mediterranean Sea
May 1996
24
0.028–0.070
ALMOFRONT II
Alboran Sea (Mediterranean Sea)
Dec. 1997–Jan. 1998
59
0.202–1.185
PROSOPE (upw)
Morocco upwelling
Sep. 1999
10
2.03–4.04
PROSOPE (Med)
Eastern and western Mediterranean Sea
Sep.–Oct. 1999
102
0.020–0.221
POMME 1
North Atlantic
Feb.–March 2001
116
0.105–0.933
POMME 2
North Atlantic
March–May 2001
125
0.254–1.44
POMME 3
North Atlantic
Aug.–Oct. 2001
125
0.039–0.395
AOPEX
Tyrrhenian Sea (Mediterranean Sea)
Aug. 2004
43
0.047–0.092
BIOSOPE
South Pacific
Nov.–Dec. 2004
62
0.017–1.481
Table 2.
PLS Parameters of -and -Models Trained Using HPLC Pigment
Measurements and Absorption Spectral Values Included in the MedCAL
Data Set (), from Left to Right: Number of
Components (), RMSEP (), Explained Variance (%) for
Independent [ (%)] and Dependent
[ (%)] Variablesa
LOO Prediction
RMSEP
(%)
(%)
Models
Tchl
4
0.1038
96.12
98.63
0.97
0.99
0.005
DP
3
0.0879
95.40
97.99
0.97
0.98
0.006
Micro
4
0.1031
96.10
95.20
0.85
0.90
0.014
Nano
4
0.0789
95.50
94.88
0.84
0.87
0.012
Pico
6
0.0221
97.64
95.76
0.87
0.88
0.006
Models
Tchl
3
0.1086
95.63
98.56
0.96
1.00
0.004
DP
2
0.0857
95.18
97.12
0.97
0.98
0.007
Micro
4
0.1085
96.31
96.56
0.84
0.91
0.010
Nano
4
0.0832
96.24
95.06
0.82
0.86
0.010
Pico
5
0.0207
97.29
95.62
0.88
0.88
0.010
Statistical parameters for linear regressions between
leave-one-out (LOO) predicted and measured pigment
concentrations: determination coefficient
(), regression slope
() and -intercept
().
Table 3.
Statistical Parameters of Comparison between the HPLC Measured and
PLS Pigment Concentrations Predicted by the
- and -Models Trained with the MedCAL Data
Set and Tested on the BOUSSOLE Time Series
()a
BOUSSOLE Prediction
RMSEP
BIAS
Models
Tchl
0.91
0.98
0.06
0.1690
0.0518
DP
0.91
1.03
0.04
0.1383
0.0510
Micro
0.75
0.91
0.06
0.1389
0.0477
Nano
0.66
0.98
0.04
0.1234
0.0378
Pico
0.54
0.94
0.01
0.0460
0.0039
Models
Tchl
0.91
0.98
0.06
0.1681
0.0540
DP
0.91
1.02
0.05
0.1393
0.0550
Micro
0.75
0.90
0.04
0.1322
0.0297
Nano
0.65
0.97
0.04
0.1250
0.0355
Pico
0.52
0.93
0.01
0.0470
0.0030
The various parameters are, from left to right: determination
coefficient (), regression slope
(), -intercept
(), RMSEP
() and systematic error (BIAS, in
).
Table 4.
PLS Parameters of -Models Trained Using HPLC Pigment
Measurements and Absorption Spectral Values Included in the GLOCAL Data
Set (), from Left to Right: Number of
Components (), RMSEP (), Explained Variance (%) for
Independent
[ (%)] and Dependent
[ (%)] Variablesa
LOO
Prediction
Models
RMSEP
(%)
(%)
Tchl
4
0.1145
88.88
94.96
0.94
0.94
0.02
DP
4
0.1025
89.32
95.14
0.94
0.95
0.02
Micro
7
0.0813
93.59
94.73
0.93
0.94
0.01
Nano
8
0.0618
94.85
91.74
0.89
0.91
0.01
Pico
8
0.0306
95.08
80.15
0.76
0.77
0.02
Statistical parameters for linear regressions between leave-one-out
(LOO) predicted and measured pigment concentrations: determination
coefficient (), regression slope
() and -intercept
().
Table 5.
Statistical Parameters of Comparison between the HPLC Measured and PLS
Pigment Concentrations Predicted by the
-Models Trained with the GLOCAL Data Set
and Tested on the BOUSSOLE Time Series ()a
BOUSSOLE Prediction
Models
RMSEP
BIAS
Tchl
0.91
1.01
0.05
0.1669
0.0565
DP
0.93
1.08
0.04
0.1402
0.0660
Micro
0.70
1.18
0.12
0.2353
0.1367
Nano
0.48
0.44
0.04
0.1266
Pico
0.42
0.60
0.01
0.0440
The various parameters are, from left to right: determination
coefficient (), regression slope
(), -intercept
(), RMSEP () and systematic error (BIAS, in
).
Tables (5)
Table 1.
Cruises, Location, Sampling Period, Number of Samples
() and [Tchl ] range
for the First Optical
Depth, for the Data Used to Train Models
Cruise
Location
Period
[Tchl ] Range
EUMELI 3
Tropical North Atlantic
Oct. 1991
5
0.073–0.340
FLUPAC
Equatorial and subequatorial Pacific
Sep.–Oct. 1994
11
0.039–0.236
OLIPAC
Equatorial and subequatorial Pacific
Nov. 1994
34
0.072–0.291
MINOS
Eastern and western Mediterranean Sea
May 1996
24
0.028–0.070
ALMOFRONT II
Alboran Sea (Mediterranean Sea)
Dec. 1997–Jan. 1998
59
0.202–1.185
PROSOPE (upw)
Morocco upwelling
Sep. 1999
10
2.03–4.04
PROSOPE (Med)
Eastern and western Mediterranean Sea
Sep.–Oct. 1999
102
0.020–0.221
POMME 1
North Atlantic
Feb.–March 2001
116
0.105–0.933
POMME 2
North Atlantic
March–May 2001
125
0.254–1.44
POMME 3
North Atlantic
Aug.–Oct. 2001
125
0.039–0.395
AOPEX
Tyrrhenian Sea (Mediterranean Sea)
Aug. 2004
43
0.047–0.092
BIOSOPE
South Pacific
Nov.–Dec. 2004
62
0.017–1.481
Table 2.
PLS Parameters of -and -Models Trained Using HPLC Pigment
Measurements and Absorption Spectral Values Included in the MedCAL
Data Set (), from Left to Right: Number of
Components (), RMSEP (), Explained Variance (%) for
Independent [ (%)] and Dependent
[ (%)] Variablesa
LOO Prediction
RMSEP
(%)
(%)
Models
Tchl
4
0.1038
96.12
98.63
0.97
0.99
0.005
DP
3
0.0879
95.40
97.99
0.97
0.98
0.006
Micro
4
0.1031
96.10
95.20
0.85
0.90
0.014
Nano
4
0.0789
95.50
94.88
0.84
0.87
0.012
Pico
6
0.0221
97.64
95.76
0.87
0.88
0.006
Models
Tchl
3
0.1086
95.63
98.56
0.96
1.00
0.004
DP
2
0.0857
95.18
97.12
0.97
0.98
0.007
Micro
4
0.1085
96.31
96.56
0.84
0.91
0.010
Nano
4
0.0832
96.24
95.06
0.82
0.86
0.010
Pico
5
0.0207
97.29
95.62
0.88
0.88
0.010
Statistical parameters for linear regressions between
leave-one-out (LOO) predicted and measured pigment
concentrations: determination coefficient
(), regression slope
() and -intercept
().
Table 3.
Statistical Parameters of Comparison between the HPLC Measured and
PLS Pigment Concentrations Predicted by the
- and -Models Trained with the MedCAL Data
Set and Tested on the BOUSSOLE Time Series
()a
BOUSSOLE Prediction
RMSEP
BIAS
Models
Tchl
0.91
0.98
0.06
0.1690
0.0518
DP
0.91
1.03
0.04
0.1383
0.0510
Micro
0.75
0.91
0.06
0.1389
0.0477
Nano
0.66
0.98
0.04
0.1234
0.0378
Pico
0.54
0.94
0.01
0.0460
0.0039
Models
Tchl
0.91
0.98
0.06
0.1681
0.0540
DP
0.91
1.02
0.05
0.1393
0.0550
Micro
0.75
0.90
0.04
0.1322
0.0297
Nano
0.65
0.97
0.04
0.1250
0.0355
Pico
0.52
0.93
0.01
0.0470
0.0030
The various parameters are, from left to right: determination
coefficient (), regression slope
(), -intercept
(), RMSEP
() and systematic error (BIAS, in
).
Table 4.
PLS Parameters of -Models Trained Using HPLC Pigment
Measurements and Absorption Spectral Values Included in the GLOCAL Data
Set (), from Left to Right: Number of
Components (), RMSEP (), Explained Variance (%) for
Independent
[ (%)] and Dependent
[ (%)] Variablesa
LOO
Prediction
Models
RMSEP
(%)
(%)
Tchl
4
0.1145
88.88
94.96
0.94
0.94
0.02
DP
4
0.1025
89.32
95.14
0.94
0.95
0.02
Micro
7
0.0813
93.59
94.73
0.93
0.94
0.01
Nano
8
0.0618
94.85
91.74
0.89
0.91
0.01
Pico
8
0.0306
95.08
80.15
0.76
0.77
0.02
Statistical parameters for linear regressions between leave-one-out
(LOO) predicted and measured pigment concentrations: determination
coefficient (), regression slope
() and -intercept
().
Table 5.
Statistical Parameters of Comparison between the HPLC Measured and PLS
Pigment Concentrations Predicted by the
-Models Trained with the GLOCAL Data Set
and Tested on the BOUSSOLE Time Series ()a
BOUSSOLE Prediction
Models
RMSEP
BIAS
Tchl
0.91
1.01
0.05
0.1669
0.0565
DP
0.93
1.08
0.04
0.1402
0.0660
Micro
0.70
1.18
0.12
0.2353
0.1367
Nano
0.48
0.44
0.04
0.1266
Pico
0.42
0.60
0.01
0.0440
The various parameters are, from left to right: determination
coefficient (), regression slope
(), -intercept
(), RMSEP () and systematic error (BIAS, in
).
Add to BibSonomy