Autonomous Ocean Sampling Network (AOSN)
Autonomous Ocean Sampling Network (AOSN) project seeks to bring
together sophisticated new robotic vehicles with advanced ocean
models to improve our ability to observe and predict the ocean.
Key to our effort is the development of control strategies to command
our mobile vehicles to places where their data will be most useful.
We call this 'adaptive sampling.' The ability to predict physical
properties of the ocean, such as temperature and current, and their
biological (ecosystem productivity) and chemical (nutrient fertilization)
are central to our program. The operational system includes data
collection by smart and adaptive platforms and sensors that relay
information to a shore in near real-time (hours) where it is assimilated
into numerical models, that help visualize the four dimensional
fields and predict future conditions. The success of the AOSN program
depends on the collaboration of a network of institutions spread
across the United States."
Excerpt from the AOSN Home page at the Monterey Bay Aquarium Research Institute (MBARI)
During August 2003, the AOSN-II field trial
was held in Monterey Bay. A staggering number of different observational
platforms were deployed in and around the bay during that month. Two numerical
models were used to predict the state of Monterey Bay, for the first time,
assimilating much of the in-situ data that was collected. The main physical
process of interest during AOSN-II was coastal upwelling, although other
features such as eddy circulations were investigated.
My own research in the AOSN project has two goals:
(1) To understand how atmospheric processes affect the ocean behavior in and around Monterey Bay
(2) To develop a reliable ensemble-based adaptive sampling strategy to select
the time and type of autonomous observations to be collected, to improve
the analysis or forecast of a physical process of interest (e.g. upwelling,
The AOSN-II field trial gave an opportunity to develop ensembles of Regional
Ocean Modeling System (ROMS) forecasts, based on (a) Initial perturbations
using the breeding technique, and (b) Phase-shifted perturbations of COAMPS
model wind stress fields. Once the ensembles were stable and running, the
Ensemble Transform Kalman Filter (ETKF) adaptive sampling strategy was developed.
Yi Chao, Zhijin Li, Jei-Kook Choi (NASA Jet Propulsion Laboratory). ROMS Ocean Model.
Craig Bishop (NRL Monterey). Atmospheric forcing, adaptive sampling, ensemble generation.
Pierre Lermusiaux (DEAS, Harvard University). Adaptive sampling, ensemble generation.
(date = ensemble initialization time)
ANIMATED GIF OF 3-KM COAMPS SURFACE WINDS DURING AOSN (WARNING: 25MB!!)
There are hundreds of plots for each day (green=available, red=not available)
COAMPS Wind Stress at Ocean Surface ("CO" in tables)
The COAMPS control forecast, issued by the Naval Research Laboratory, Monterey is given for 0h, 24h, 48h and 72h.
A phase-shift method based on Autoregressive functions is used to perturb
the COAMPS wind stress fields in space and time. The phase shift functions
"lx" and "ly" are shown in the top panels for each time, for each of the
8 ensemble members (there is a similar phase shift "tau" in time, not shown)
The wind stress fields u_s and v_s are perturbed in the following manner:
u_s (x,y,t) --> u_s (x+lx,y+ly,t+tau)
v_s (x,y,t) --> v_s (x+lx,y+ly,t+tau)
Using this technique, it is expected that realistic atmospheric features
are preserved, and that ensemble variance is largest in areas where the gradient
is highest. The ensemble generation technique still need to be refined considerably,
but the first attempt is encouraging.
ROMS Ocean Model
The ROMS model (labeled "ctrl")
was run each day at JPL, with all available and quality-checked AOSN-II data
being assimilated using a 3d-Var scheme. The model consists of 3 nested grids,
with resolution 15, 5 and 1.67 km. The innermost domain covers Monterey Bay.
Each of the 8 ROMS ensemble members was run to 2-3 days on the JPL Supercomputer,
with the corresponding perturbed COAMPS atmospheric forcing, and initial
condition derived from breeding. The forecast ocean currents, temperature
and salinity fields are shown for depths of 0m, 10m, 30m, 50m and 200m in
the Monterey Bay region.
ETKF Summary Maps
Based on the ROMS ensembles, the ETKF adaptive sampling (or "targeting")
strategy was tested near the end of AOSN-II. Some maps are linked from the
lower frame. The "summary map" gives the predicted reduction in forecast
error variance as a function of the location in which a targeted observation
is taken. Hence, the best location to take an observation is where the value
is highest (red shading). Unlike our atmospheric targeting to date, these
summary maps are 3-dimensional fields. Only horizontal and vertical cross-sections
are shown here. In principle, one can draw a hypothetical glider track through
the maxima of the ETKF result. Alternatively, the ETKF calculations can be
computed for any combination of prescribed observational paths, over a series
of times. This is the direction in which we'd like to head, but first the
ETKF needs to be evaluated before taking this ambitious step. Currently,
I don't have tremendous faith in the ETKF results since the number of ensemble
members (8 at most) is far too small, and the ensemble perturbation technique
above needs more development. Improved computing power over the next few
years will alleviate the small-ensemble problem considerably.
For more details on the adaptive sampling using the ETKF, click here.
The Near Future
1. The Error Subspace Statistical Estimation approach of Lermusiaux and Robinson
(1999) has some theoretical similarities with the ETKF. A theoretical inter-comparison
between the two strategies in data assimilation, ensemble generation, and
adaptive sampling will be performed in the very near future. The relative
strengths of the two techniques can be exploited to produce an improved ensemble-based
strategy for adaptive sampling.
2. The breeding technique for ensemble initialization will be replaced by
the ETKF technique of Wang and Bishop (2003), once the observation operators
and error variance estimates for the myriad of observation types are available.
3. One of the primary mechanisms for producing an upwelling event
is strong sustained winds, that last for several days. Typically, these winds
prevail from the north-west during the California summer. The sensitivity
of the ROMS ocean model to changes in atmospheric wind fields needs investigation.
The development of a fully coupled atmosphere-ocean model requires considerable
research and is probably some time away, but the simple sensitivity studies
using prescribed forcings and fluxes may still shed valuable light on how
the atmospheric fields are affecting the oceanic state.
4. The ETKF targeting strategy needs to be evaluated. First, the
ability of the ETKF to predict the variance of "signals" will be assessed.
A signal represents the difference in two ocean state estimates; one of which
has assimilated the targeted observations, whereas the other has ignored
the targeted observations. The techniques of Majumdar et al. (2001) and Bishop
et al. (2003) will be used to test whether the ETKF is making reliable predictions.
As mentioned above, there are several different applications of the ETKF
that can be developed and evaluated. However, large, reliable ensembles are
necessary (although by no means sufficient) for the ETKF to make trustworthy
predictions of the effects of targeted observations.
A reliable adaptive sampling strategy can in principle be used to deploy
autonomous vehicles along any path, at any time, to improve forecast skill.
The development of such a strategy that is consistent with a state-of-the-art
data assimilation scheme that assimilates observations nearly continuously
in time is the ultimate goal. The intelligent deployment of autonomous vehicles
in the ocean (and atmosphere) is likely to lead to improved forecasts and
understanding of key physical processes of interest.
Photos (click to enlarge)
Last Updated: May 2 2004