IEEE Robotics & Automation Magazine - September 2013 - 47

Depth

20
19
18
44.2
44
9.2

9.4
9.6
Longitude

9.8

43.8
43.6 Latitude

17
16
15

0
-10
-20
-30
-40
-50
-60
-70
-80

21
20

Depth

0
-10
-20
-30
-40
-50
-60
-70
-80
9

19
18
44.2
44
9.2

9.4
9.6
Longitude

9.8

(a)

43.8
43.6 Latitude

Depth

0
-1
44.2

-2

44
9.4
9.6
Longitude

9.8

43.8
43.6 Latitude

-3

2
0
-10
-20
-30
-40
-50
-60
-70
-80

1
0

Depth

9.2

(a)
3

0
-10
-20
-30
-40
-50
-60
-70
-80

-1
44.2

-2

44
9.2

9.4
9.6
Longitude

9.8

(b)

43.8
43.6 Latitude

-3

(b)

Figure 7. (a) Temperature field estimations and (b) associated
errors in °C obtained from the data-driven environmental
assessment at the locations sampled by the ScanFish. The
coastline is represented by the black line.

Figure 8. (a) Temperature field estimations and (b) associated
errors in °C obtained from the model-driven environmental
assessment at the locations sampled by the ScanFish. The
coastline is represented by the black line.

score R 2 is a dimensionless coefficient used to compare the variance explained by the estimated fields with the total variance of
the observations considered as validation data set [34]. Values of
R 2 close to one indicate good performance, while values equal
to or less than zero are suggestive of limited skill.
Figure 7 displays the vertical profiles estimated from the
data-driven assessment method and the associated error. The
estimated temperature field shows a well-defined layered vertical structure with an appropriate range of temperature values
at depth. However, the estimated field resembles a smoothed
version of the real data. The data-driven assessment method
fails to represent the small-scale variability observed in the
thermocline records from the ScanFish. This is evidenced by
the uneven vertical distribution of errors in Figure 7(b). In
fact, errors are rather low at all depths, except in the layer
between 20- and 30-m depth. Temperature gradients associated with the thermocline lead to significant inaccuracies if the
depth and gradient characteristics of the thermocline are not

perfectly reproduced in the estimation process. The overall
performance of the data-driven assessment method is given by
a value of 0.51 for R 2, indicating that half of the initial variance
in the data is properly explained.
Profile estimations and corresponding errors from the
model-driven strategy are displayed in Figure 8(a) and (b),
respectively. Temperatures estimated in the surface layer are
cooler than the ScanFish observations. The model predicts
an eastward negative tilt in the thermocline depth, with
warmer temperatures onshore. This corresponds to the
thermal structure observed by the ScanFish in the two
northernmost transects, but it is reversed when compared
with the structure found along the southern transect. Compared with results from the data-driven assessment
method, errors are larger throughout the water column,
and more particularly near the thermocline. The R 2 is 0.14,
which indicates an overall lower skill compared with the
other method.
september 2013

IEEE ROBOTICS & AUTOMATION MAGAZINE

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - September 2013