Retrieval of phytoplankton size classes from light absorption spectra

Report
THE
45TH INTERNATIONAL LIÈGE COLLOQUIUM
17TH MAY 2013
Retrieval of phytoplankton size classes from
light absorption spectra using a multivariate
approach
Emanuele ORGANELLI, Annick BRICAUD, David ANTOINE and Julia UITZ
Laboratoire d’Océanographie de Villefranche, UMR 7093, CNRS and Université
Pierre et Marie Curie, Paris 6, 06238 Villefranche sur Mer, FRANCE
[email protected]
Motivations
 To assess Total Primary Production in the
oceans, new approaches (Uitz et al.,
2008,
2010,
2012)
concern
the
estimation of PHYTOPLANKTON CLASSSPECIFIC contributions.
 Combination of ocean color-based PP
models with algorithms for retrieving
Phytoplankton Size Classes (PSC) from
optical properties (IOPs and AOPs).
Uitz et al. (2012), Glob. Biogeochem.
Cycles, GB2024
Classification of current approaches by Brewin et al. (2011)
1. Spectral Response-based approaches
(based on differences in optical signatures of phytoplankton groups)
2. Abundance-based approaches
(rely with the trophic status of the environment and the type of
phytoplankton)
3. Ecological-based approaches
(based on the knowledge of physical and biological regime to identify
different types of phytoplankton)
Uncertainties and sources of errors!
Brewin et al. (2011). Remote Sens. Environ., 115, 325-339
Objective
To develop and test a new model for the retrieval of PSC using the
multivariate Partial Least Squares regression (PLS) technique.
 Scarcely applied in oceanography but with satisfactory results
(Moberg et al., 2002; Stæhr and Cullen, 2003; Seppäla and Olli, 2008;
Martinez-Guijarro et al., 2009).
 PLS is a spectral response approach which uses light absorption
properties.
0.12
Diatoms
Prymnesiophytes
0.1
a* p (m2 mg TChl a -1)
Prasinophytes
Cyanobacteria
0.08
Prochlorococcus sp.
0.06
0.04
0.02
0
400
450
500
550
wavelength (nm)
Bricaud et al. (2004), J. Geophys. Res., 109, C11010
600
650
700
PLS: INPUT and OUTPUT
Multivariate technique that relates, by regression, a data matrix of
PREDICTOR variables to a data matrix of RESPONSE variables.
INPUT VARIABLES
Fourth-derivative of
PARTICLE (ap(λ)) or
PHYTOPLANKTON (aphy(λ))
light absorption spectra
(400-700 nm, by 1 nm)
OUTPUT VARIABLES
(in mg m-3)
[Tchl a]
[DP] ([Micro]+[Nano]+[Pico])
[Micro] (1.41*[Fuco]+1.41*[Perid])a
[Nano]
a
Coefficients by Uitz et al. (2006). J. Geophys. Res., 111, C08005
(1.27*[19’-HF]+0.35*[19’-BF]+0.60*[Allo])a
[Pico]
(1.01*[TChl b]+0.86*[Zea])a
Plan of the work
1. INPUT and
OUTPUT
2. TRAINING
3. TEST
REGIONAL data set for PLS training
Data: HPLC pigment and light absorption (ap(λ) and aphy(λ)) measurements
from the first optical depth.
MedCAL data set (n=239): data from the Mediterranean Sea only
MedCAL-trained models
5.0
[Tchl a] predicted
6.0
R2=0.97
RMSE=0.10
1:
1
4.0
3.0
2.0
1.0
4.0
3.0
2.0
1.0
[Tchl a] measured
R2=0.90
RMSE=0.10
3.0
2.5
2.0
1.5
1.0
0.5
0.0
0.0 0.5 1.0 1.5 2.0 2.5 3.0
[Micro] measured
[Micro] measured
(e)
[Nano] predicted
variable
1.5
2.0
1.0
0.5
0.0
including
(LOO)
trained
leave-one-out
cross-validation
technique
0.5
1.0
1.5
0.5
2.0
0.6
0.5
R2=0.88
RMSE=0.02
1:
1
were
R2=0.86
RMSE=0.08
1.0
0.0
[Nano] measured
(g)
[Pico] predicted
 Models
1.5
0.0
0.0
0.4
0.3
0.2
0.1
0.0
0.5
1.0
1.5
2.0
[Nano] measured
0.6
[Pico] predicted
output
R2=0.87
RMSE=0.08
1:
1
each
[Nano] predicted
model
2.0
R2=0.91
RMSE=0.11
1:
1
3.0
2.5
2.0
1.5
1.0
0.5
0.0
1:
1
[Tchl a] measured
(a)
[Micro] predicted
0.0 1.0 2.0 3.0 4.0 5.0 6.0
1:
1
0.0 1.0 2.0 3.0 4.0 5.0 6.0
0.0 0.5 1.0 1.5 2.0 2.5 3.0
 1
R2=0.96
RMSE=0.11
0.0
0.0
[Micro] predicted
5.0
0.5
R2=0.88
RMSE=0.02
1:
1
[Tchl a] predicted
MedCAL aphy(λ)-models
1:
1
MedCAL ap(λ)-models
6.0
0.4
0.3
0.2
0.1
0.0
0.0 0.1 0.2 0.3 0.4 0.5 0.6
0.0 0.1 0.2 0.3 0.4 0.5 0.6
[Pico] measured
(i)
[Pico] measured
MedCAL-trained models: TESTING
MedCAL aphy(λ)-models
0.1
0.01
0.01
0 .1
1
0 .0
R2=0.75
RMSE=0.14
0.1
0.01
0.01
0.001
0.001
HPLC
pigment
and
light
absorption measurements at the first optical
depth in the period January 2003-May 2011
R2=0.66
RMSE=0.12
0.1
0.01
0.01
0.001
0.001
trained models
1
[Pico] predicted
 Similar performances of ap(λ) and aphy(λ)
1
0.1
(n=484).
showed), Micro, Nano and Pico
0 .0
[Micro] measured
(e)
0 .0
 Good retrievals of Tchl a, DP (not
1
1:
1
monthly
1
0 .1
0 .1
0 1 0 .0 1
1
[Nano] measured
(g)
1
R2=0.54
RMSE=0.046
0.1
0.1
0.01
0.01
0.001
0 .0
01
0.001
0 .0
1
0 .1
[Pico] measured
01 0.01
1
0.1
1
[Micro] measured
R2=0.65
RMSE=0.12
0 .0
1:
1
Sea):
[Nano] predicted
BOUSSOLE time-series (NW Mediterranean
0 1 0 .0 1
1:
1
0.1
0 .0
1
R2=0.75
RMSE=0.13
1
1:
1
[Micro] predicted
1
0 .1
1
[Tchl a] measured
[Tchl a] measured
(a)
1:
1
1
1:
1
1
0.1
0 .0
R2=0.91
RMSE=0.17
0 1 0 .0 1
0 .1
1
[Nano] measured
R2=0.52
RMSE=0.047
1:
1
1
R2=0.91
RMSE=0.17
1:
1
[Tchl a] predicted
MedCAL ap(λ)-models
0 .0
01
0 .0
1
0 .1
[Pico] measured
1
Boussole time-series from MedCAL-trained models
Tchl a
Micro
Nano
Pico
Seasonal dynamics of algal size structure at BOUSSOLE
Tchl a
Max in Spring bloom (from mid-March to mid-April)
Low concentrations from June to October
Increase in Winter
Micro-phytoplankton
Max in Spring bloom (from mid-March to mid-April)
Low concentrations during the rest of the year
Nano- and Pico-phytoplankton
Recurrent maximal abundance in late Winter and
early Spring
Increase
December
in
Summer
and
from
October
to
If PLS models are trained with a global dataset...
GLOCAL data set (n=716): HPLC pigment and phytoplankton light
absorption measurements (aphy(λ)) from various locations of the world’s
oceans (Mediterranean Sea included).
2.0
2.0
1.0
1.0
0.0
0.0
0.0
0.01.0
1.02.0
2.03.0
3.04.0
4.05.0
5.06.0
6.0
[Tchl
[Tchla]
a]measured
measured
2.0
1.0
0.0
0.0 1.0 2.0 3.0 4.0 5.0
[DP] measured
0.5
2.0
1.0
0.0
1.5
R2=0.89
RMSE=0.06
1:
1
1:
1
2.0
[Pico] predicted
3.0
3.0
3.0
3.0
R2=0.93
RMSE=0.08
[Nano] predicted
4.0
4.0
4.0
4.0
R2=0.94
RMSE=0.10
[Micro] predicted
5.0
1:
1
R2=0.94
RMSE=0.11
[DP] predicted
5.0
5.0
11::1
1
[Tchl a]
a] predicted
predicted
[Tchl
6.0
6.0
1.0
0.5
0.0
0.0
1.0
2.0
3.0
[Micro] measured
4.0
0.4
R2=0.76
RMSE=0.03
1:
1
GLOCAL aphy(λ) Trained -models
0.3
0.2
0.1
0.0
-0.1
0.0 0.5 1.0 1.5
[Nano] measured
2.0
-0.1 0.0 0.1 0.2 0.3 0.4 0.5
[Pico] measured
...but when we test the models...
0.1
0.01
R2=0.93
RMSE=0.14
1
0.1
0.01
1:
1
[DP] predicted
R2=0.91
RMSE=0.17
1
1:
1
[Tchl a] predicted
GLOCAL aphy(λ)-models
0.001
0.0
01 0.01
0.1
1
0 .0
and DP
0.1
0.01
0.001
0 .0
0 1 0 .0 1
0 .1
 Underestimation of Nano
and Pico
0.1
0.01
0.001
0.0
01 0.01
R2=0.42
RMSE=0.044
0.1
0.01
0.001
0.0
0.1
1
[Nano] measured
[Micro] measured
[Pico] predicted
 Overestimation of Micro
R2=0.48
RMSE=0.13
1
1
1
1
1:
1
[Nano] predicted
R2=0.70
RMSE=0.23
0 .1
[DP] measured
1:
1
 Good retrievals of Tchl a
1
1:
1
[Micro] predicted
[Tchl a] measured
1
01
0.0
1
0.1
[Pico] measured
1
How to explain differences?
Amplitude
wavelength
bands
in
and
of
center
absorption
the
fourth–
derivative spectra at the
BOUSSOLE site are:
 Similar to those of the
other Mediterranean areas.
 Different to those of the
Atlantic and Pacific Oceans.
Summary and Conclusions
 Retrieval of algal biomass and size structure from in vivo hyper-spectral absorption
measurements can be achieved by PLS:
 High prediction accuracy when PLS models are trained and tested with a
REGIONAL dataset (MedCAL and BOUSSOLE);
 The dataset assembled from various locations in the World’s oceans (GLOCAL)
gives satisfactory predictions of Tchl a and DP only.
 The PLS approach gives access to the analysis of SEASONAL DYNAMICS of algal
community size structure using optical measurements (absorption).
 Main advantage of PLS approach is the INSENSITIVITY of the fourth-derivative
to NAP and CDOM (new analyses reveal it!) absorption properties that means:
 Prediction ability is very similar for ap(λ) and aphy(λ) PLS trained models
 This opens the way to a PLS application to total absorption spectra derived
from inversion of field or satellite hyperspectral radiance measurements (this is
currently being tested over the BOUSSOLE time series!)
Citation: Organelli E., Bricaud A., Antoine D., Uitz J. (2013). Multivariate approach for the
retrieval of phytoplankton size structure from measured light absorption spectra in the
Mediterranean Sea (BOUSSOLE site). Applied Optics, 52(11), 2257-2273.
Acknowledgements: This study is a contribution to the BIOCAREX (funded by ANR) and
BOUSSOLE (funded by ESA, NASA, CNES, CNRS, INSU, UPMC, OOV) projects.
Many thanks to the
BOUSSOLE team!
Thank you for the
attention!

similar documents