Spatial mapping of chlorophyll concentration in the Baltic Sea

CM 2003/ L:13.
Plankton Monitoring: Better Coverage by Ship-of-Opportunity
and Remote Sensing Methods.
Spatial mapping of chlorophyll concentration in the Baltic Sea through the
assimilation of satellite data with ship-of-opportunity observations
Jouni Pulliainen(1), Jenni Vepsäläinen(2), Seppo Kaitala(3), Kari Kallio(2),
Eija Rantajärvi(3) and Sampsa Koponen(1)
(1) Laboratory of Space Technology, Helsinki University of Technology, P.O. BOX
3000, FIN-02015 HUT Finland, tel +358 9 451 2373, fax : +358 9 451 2898, email: [email protected]]
(2) Finnish Environment Institute, PO BOX 140, 00251 Helsinki, Finland. [tel: +358
9 40 300661, fax: +358 9 403 00691, e-mail: [email protected]]
(3) Finnish Institute of Marine Research, P.O. BOX 33, 00931 Helsinki, Finland. [tel:
+358 9 61394417, fax:+358 9 6139 4494, e-mail: [email protected]]
Abstract
Unattended flow-through fluorometers are currently operatively employed in the
Baltic Sea to provide information on chlorophyll a concentration and distribution.
This information is collected on daily basis at merchant and passenger vessels
operating at the region (automated Alg@line sampling system). The observations can
be used for the general assessment of the status of the Baltic Sea, as chlorophyll a
(chl-a) concentration is one of the key factors of water quality. However, the spatial
coverage of these ship-of-opportunity data is restricted to transects cruised by vessels.
Space-borne optical spectrometers, such as SeaWiFS and MODIS, provide daily
remote sensing reflectance observations correlated to chl-a concentration of surface
water layer. These data are available whenever cloud conditions permit.
Spectrometers can be used for the mapping of chl-a using empirical or semi-empirical
algorithms. However, their accuracy is limited in the brackish waters of the Baltic Sea
due to the relatively high turbidity and yellow substance (CDOM) level. The optimum
method to apply satellite data is the assimilation of remote sensing observations with
in situ data. This is demonstrated in this investigation by introducing a method that
takes an advantage of the high accuracy of Alg@line transect data together with the
full spatial coverage of satellite observations. It is shown that the developed method
significantly improves the quantitative regional water quality assessment accuracy
when compared with the use of (a) only the transect data or (b) only the space-borne
data.
Introduction
This paper presents a method that optimally combines transect observations on
chlorophyll a (chl-a) concentration with spatially continuous optical space-borne
spectrometer data. The observations of flow-through fluorometers can be used for the
general assessment of the status of the Baltic Sea, as chl-a concentration determined
from the observations is one of the key factors of water quality. It is directly related to
phytoplankton biomass. However, the spatial coverage of these ship-of-opportunity
data, such as Alg@line observations (Leppänen and Rantajärvi, 1995), is restricted to
transects cruised by vessels.
Space-borne optical spectrometers, such as SeaWiFS and MODIS, provide daily
remote sensing reflectance observations correlated to chl-a concentration of surface
water layer. These data are available whenever cloud conditions permit.
Spectrometers can be used for the mapping of chl-a in the Baltic Sea using e.g.
empirical channel difference ratio algorithms (Vepsäläinen et al. 2003). However,
their accuracy is limited in the brackish waters of the Baltic Sea due to the relatively
high turbidity and yellow substance (CDOM) level. The optimum method to apply
satellite data is the assimilation of remote sensing observations with in situ data. This
is demonstrated in this investigation by introducing a method that takes an advantage
of the high accuracy of Alg@line transect data together with the full spatial coverage
of satellite observations.
Test data and methods
The study material includes Alg@line chl-a data from a ship route between Germany
and Finland (Travemünde and Helsinki) and nearly concurrent SeaWiFS satellite
spectrometer observations. Altogether six dates of SeaWiFS observations from the
spring algae bloom periods from years 1999 and 2000 are included. In these cases the
major part of the test site was cloud free enabling the detailed investigation of
assimilation technique performance.
The developed assimilation technique is a Bayesian approach in which the maximum
conditional probability of the chl-a value is searched as certain specific set of satellite
and in situ observations is obtained for the 1 km × 1 km pixel under investigation. The
satellite data are spatially continuous observations whereas the Alg@line data include
transect observations. The data assimilation technique optimally combines these two
data sources by taking into account the spatial autocorrelation function of chl-a
concentration estimated from the transect data. The kriging interpolation is used for
the extrapolation or interpolation of Alg@line observations for any location outside
the employed transect.
The assimilation technique applies linear modeling for linking satellite observations to
chl-a observations:
y i = β i1 x + β i 2 + ε i = f i ( x) + ε i
(1)
where x is the chl-a concentration in logarithmic scale, x = loge(chl-a×(1µg/l)-1), and y
is a space-borne observed TOA radiance (channel) ratio at the channel combination i
(i = 1 or i = 2). ε is the random error of the model. Vepsäläinen et al. (2003) show that
in the case of the Baltic Sea two particular SeaWiFS-observed channel difference
ratios are strongly and linearly related to the logarithmic value of chl-a concentration:
L490 nm − L765 nm

 y1 = L
670 nm − L765 nm



− L765 nm
L
 y 2 = 510 nm

L670 nm − L765 nm
(2)
In the case of linear model given by (1) the assimilated estimate of x is obtained
without iterative search directly by equation
T
(
y − β 2 ) Mβ1 + λ REF xˆ REF
xˆ =
β1T M β1 + λ REF
(3)
where y is the vector of observed channel difference ratios according to (2). M is the
inverse matrix of the modeling error covariance matrix M = C-1. Additionally, λREF =
−2
σ REF
, where σREF is the spatially varying accuracy of reference information (standard
deviation of statistical error) and x̂ REF is the estimate of x for the location under
investigation determined through the kriging interpolation of Alg@line data.
β1 refers to slope and β2 to the constant intercept of the model given by (1):
β = [β1
β
β 2 ] =  11
 β 21
β 12 
β 22 
(4)
In order to evaluate the quantitative accuracy characteristics and the benefits of the
assimilation technique, the results of data assimilation are compared with those of
other applicable approaches. The other methods used here for assessing the
momentary status of the Baltic Sea include the use of (empirical) satellite data
retrieval algorithms (as stand alone and bias-corrected algorithms), and the use of
values directly extra- or interpolated from the transect data.
Results and Discussion
The quantitative estimation errors are determined for each day of observation by
dividing the alg@line transects either into three or two parts with equal number of
samples. Thereafter, the chl-a estimates determined either for southern (S), central (C)
or northern (N) part of transect by assuming that Alg@line data does not cover this
particular test region, but instead are only available for other part(s) of the transect.
The comparison of obtained estimates with true Alg@line observations from the test
region yields the quantitative estimation error characteristics including both the
statistical and systematic error of different methods. An essential part of the algorithm
testing was the comparison against extrapolations and interpolations that are based
solely on Alg@line data.
An example of data assimilation results is depicted in Figure 1 for the 31 of May
1999. During that date the whole Baltic Sea region was almost cloud-free. An
example on improvement achieved by applying the assimilation method is presented
in Table 1. The results in Table 1 show the accuracy of chl-a estimates obtained by
using only the satellite observations as well as the accuracy of results determined by
extrapolating the Alg@line data. The comparison of the error characteristics of these
two methods with those of assimilation (also given in Table 1) demonstrates the
accuracy improvement obtained through the use of the assimilation technique. The
results of Table 1 show the average accuracy characteristics (systematic error and
pixel-wise RMSE) for the northward extrapolation in the six investigated cases. The
systematic (bias) error indicates the performance obtainable in the assessment of an
average value of a certain region, whereas the RMSE represents the error of an
individual pixel ( 1 km by 1 km region in this case).
Logarithmic scale: 10 × log10 [chl-a(µg/l) / 1]
Figure 1. Assimilation of Alg@line transect data to SeaWiFS observations for 20
May, 1999 showing the distribution chlorophyll a concentration in natural logarithmic
scale.
Table 1. Estimation errors of various methods for the northward extrapolation of six
dates indicating the chl-a retrieval accuracy improvement obatained through the use
of the developed assimilation technique
Satellite data
Bias error
25.9 %
Assimilation
Bias error
15.4 %
Extrapol. Alg@line
Bias error
24.8 %
Satellite data
RMSE
45.1 %
Assimilation
RMSE
32.9 %
Extrapol. Alg@line
RMSE
61.2 %
Conclusions
The Alg@line data is used for the quantitative assessment of the chl-a concentration
in the Baltic Sea. However, the spatial coverage of these data is restricted to the
transects cruised by ship-of-opportunity vessels. The optimum remote sensing aided
method to improve the spatial quality of Alg@line data is the assimilation of remote
sensing observations with transect data (observations of optical spectrometers are
highly correlated with chl-a). The developed method takes an advantage of the high
accuracy of Alg@line transect data together with the full spatial coverage of satellite
observations. It is shown that the introduced technique significantly improves the
quantitative regional water quality assessment accuracy when compared with the use
of (a) only the transect data or (b) only the space-borne data.
References
Leppänen, J.-M. and Rantajärvi, E. (1995), Unattended recording of phytoplankton
and supplemental parameters on board merchant ships – an alternative to the
conventional algal monitoring programmes in the Baltic Sea, In Harmful Marine
Algal blooms. Edited by P. Lassus, G. Arzul, E. Erard-Le Denn, P. Gentien and
Marcailloure-Le Baut, Lavoisier, paris, pp. 719-724.
Vepsäläinen, J., Pyhälahti, T., Rantajärvi, E., Kallio, K., Pertola, S., Stipa, T.,
Kiirikki, M., and Pulliainen, J., The combined use of optical remote sensing data and
unattended flow-through fluorometer measurements in the Baltic Sea. International
Journal of Remote Sensing (accepted 2003).