IEEE Robotics & Automation Magazine - September 2013 - 48

0
-10

Depth

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

0.5

1.5

RMSE
Figure 9. Vertical distribution of the RMSE for the data-driven
(solid line) and model-driven (dashed-dotted line)
environmental assessment methods.

The validation of both assessments is done more quantitatively in Figure 9. This figure displays the root mean square
error computed for each 5-m-thick vertical layer. Results
again indicate that both estimation procedures fail to accurately reproduce the variability observed at the thermocline
depth. The data-driven approach, based on glider and remote
sensing data, outperforms the numerical model at the surface
and in deeper layers. The similarity in the shapes of the curves
provides insights about the nature of the error in the modeldriven assessment. Except for the surfacemost layer, the thermal estimation obtained from the model is shifted toward
warmer values. Further analysis confirms that the numerical
model temperature field has a bias of 0.44 °C, whereas the
data-driven estimate is essentially unbiased.
Data- or Model-Driven Marine
Environmental Characterization?
Observational MILOC is transitioning from ship-based
capabilities to networked robotic platforms, which ensure
safe, cost effective, and discreet access to areas otherwise
denied to traditional units. Present glider technology is
mature enough to carry out persistent environmental monitoring by a fleet of robots, which allows a battlespace environmental characterization exclusively on the basis of
observational assets. This contrasts with the traditional
numerical modeling characterization. This article compared
the two approaches, with the aim to quantify the gain in
information, if any, obtained from a robotic environmental
characterization. Results show that the data-driven environmental assessment substantially improves the field estimations compared with the model-driven environmental
assessment. The environmental characterization based on
the glider and remote sensing data is accurate in the surface
layer and below the thermocline. Integration of remote sensing information is essential to improve the results at the surface, because no glider measurement is considered above
20 m to simulate battlespace safety conditions. Below the
48

IEEE ROBOTICS & AUTOMATION MAGAZINE

september 2013

thermocline, the accuracy of the data-driven estimation is
related to the low spatial variability of the thermal field in the
deep layers. The accuracy is substantially degraded in the
thermocline. As this portion of the water column is of special
interest for MILOC because of the impact of large temperature gradients on the sound propagation, devoting part of the
glider fleet to track the thermocline variability (e.g., confining glider paths between 20- and 40-m depth) should be
considered in the future.
It is certainly easier to reconstruct environmental conditions from observations than from the numerical simulation
of the full dynamical interactions. However, despite the lower
accuracy obtained in this experiment by the numerical model,
environmental simulation remains a very valuable tool as it
provides not only a synoptic description of the oceanic field
but also a predictive view of the marine environment. Numerical ocean models can now be considered mature in the simulation of processes like external tides, storm surges, river plumes,
coastal topographic waves, upwelling and alongshore currents,
and mesoscale activity [35]. Notice that the direct characterization from glider observations of part of these environmental
processes is unfeasible because of the slow motion and lack of
accurate underwater positioning of glider platforms. Indeed,
numerical models are the core of the optimized environmental
characterization discipline, focused on developing predictions
of the evolution of the marine environment. Further research
is required to improve the models' accuracy. Numerical ocean
models are inevitably limited in their spatiotemporal resolution, which necessitates use of ad hoc parameterizations to
represent the subgrid scale processes. For example, vertical
mixing (including the mixing induced by internal waves)
involves very small-scale processes (i.e., turbulence), and thus,
it is parameterized in the present model application. This is
then reflected in the simulation of the thermocline, which is
too smooth when compared with observations. Such features
can be better represented with robotic platforms such as gliders. It should be noted that no data were assimilated in the area
of interest in the numerical approach considered in this study.
Assimilation of observations from robotic fleets would likely
improve the realistic representation of marine environments
by operational ocean models. In this approach, the network
topology of the robotic fleet may be established by a continuous feedback of information between the observational nodes
and a numerical simulation engine that produces a physically
consistent analysis of the battlespace on the basis of the information received from in situ and remote sensors. The hierarchy of near future sampling strategies established by data
assimilation may be communicated to the robots through
underwater or above water communication networks. Exploiting synergism between both procedures is envisioned to
improve the assessment and predictions of marine environments. The performance of the above procedure (known as
adaptive sampling) to characterize mesoscale activity in accessrestricted marine areas, is being investigated at NURC.
To summarize, our results indicate that marine underwater characterization obtained from adequate observational

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