Essential components and technical details of regional ocean forecasting systems configured from the Regional Ocean Modeling System are discussed with the goal of bridging the gap between user and ocean modeling communities. Recent development of these systems and applications are also surveyed. Design considerations of such a system for the South China Sea are discussed, based on regional dynamic characteristics and potential applications.
During the past ten years or so, regional ocean forecasting systems were developed for various regions around the world. These systems tend to be high-resolution (order 1 km horizontal grid spacing), assimilate data, include tidal forcing, and provide a web interface for user applications (Chao et al., 2009; Farrara et al., 2013; Schmidt and Gangopadhyay, 2013; Warner et al., 2010). Various applications, supporting scientific field experiments, search and rescue operation, gas and oil industry, and scientific research were also developed for these systems.
This article summarizes the development of various ocean data assimilation systems along the U.S. coast, mainly based on the work of the Regional Ocean Modeling System Group at the Jet Propulsion Laboratory, California Institute of Technology and the University of California at Los Angeles. With the intention to bridge the gap between societal needs and products from regional ocean forecasting systems, the purpose of this review is fourfold. First, we provide a basic description of a regional ocean forecasting system for the benefit of general users whose field is not physical oceanography or atmospheric sciences. Second, we describe applications that use products from this type of system. Third, we discuss design considerations of such a system for the South China Sea. Fourth, we comment on some recent developments in the field and future directions that will raise the bar for the next generation of regional ocean forecasting systems.
Regional ocean forecasting systems
Several regional ocean forecasting systems were set up for various regions along the U.S. coast. These regions include the Monterey Bay (Chao et al., 2009; Wang et al., 2009), the Southern California Bight (Li et al., 2015), the Prince William Sound, Alaska (Farrara et al., 2013), and the Gulf of Mexico (Farrara et al., 2012). Despite the different physical environment that these ocean forecasting systems were developed for, all of these systems share some common characteristics. Figure 1 is a schematic diagram showing the components of a regional ocean forecasting system and their functions. In general, a regional ocean forecasting system includes six components: (1) Acquisition of atmospheric forcing fields, (2) acquisition of lateral boundary conditions, (3) acquisition of ocean observations, (4) data assimilation, (5) forecasting and (6) result monitoring and visualization. The data assimilation and forecasting components were put in one box because of the close connection between these two components. The development of an ocean forecasting system requires detailed planning about different aspects of the components, data exchange, coordination of computing time, contingency plans if there are failures in one or more components, etc. The operational stage of such a system requires monitoring of each and every component and diagnosis of issues if one component does not perform as expected.
Regional Ocean Modeling System
The central part of a regional ocean forecasting system is a general circulation model. The Regional Ocean Modeling System (ROMS) is a community modeling tool, which is widely used in regional applications. ROMS solves the Navier-Stokes equations under hydrostatic and Boussinesq assumptions in order to describe the evolution of velocity, temperature, and salinity fields. The computation kernel of ROMS is free-surface, split-explicit, and terrain-following (Shchepetkin and McWilliams, 2005, 2009). ROMS explicitly describes the time evolution of the free surface, as opposed to the approach of Bryan (1969) who used a rigid-lid assumption and solved a two-dimensional Poisson equation for surface pressure and hence surface height. The evolution of the free surface is governed by the depth-integrated momentum equation (often called the barotropic mode), which is governed by fast processes, such as surface gravity waves, and can only be integrated using small time steps satisfying the Courant–Friedrichs–Lewy (CFL) condition. The residual part (often called the baroclinic modes) is governed by processes of longer time scales, for example, Rossby waves and subgridscale mixing, and can be integrated using longer time steps. The separation of barotropic and baroclinic modes in ROMS allows separation of boundary conditions so that tides can easily be implemented through lateral boundary conditions when the domain involved is relatively small.
In vertical direction, ROMS uses a stretched sigma coordinate (s-coordinate) discretization that follows bathymetry. The vertical discretization transforms the Cartesian height to a new coordinate. For instance, ocean free surface corresponds to in the new coordinate system and ocean bottom corresponds to . Compared with a traditional sigma-coordinate discretization, the coordinate system used in ROMS provides more flexibility in choosing vertical levels in specific vertical domains, such as the surface boundary layer and the bottom boundary layer (Song and Haidvogel, 1994). Horizontally, a curvilinear grid is used to specify the domain of interest. Submodels, for example, biological models of different complexity or a sea ice model, can be coupled with the ocean circulation model, as needed by each application. A suite of subgridscale parameterizations are provided in ROMS in order to represent processes that have spatial scales smaller than the model grid size, e.g. the K-Profile Parameterization (KPP) of Large et al. (1994) parameterizations that represent turbulent mixing of momentum and mass (Haidvogel et al., 2008).
Nested domains and boundary conditions
To allow high spatial resolution in regions of interest, ROMS allows online nested domains. Online nesting allows the enclosing and enclosed domains to run simultaneously and exchange boundary conditions at every time step. By way of contrast, offline nesting integrates each domain one after the other. To illustrate domain nesting, Figure 2 shows the nested model domains used for the Prince William Sound, Alaska studies of Wang et al. (2013a) and Colas et al. (2013). In this case the outermost model domain (L0, the region that has bathymetry in color) has the coarsest resolution and covers the U.S. and Canadian western coastal region from 42.61N to 61.31N. The intermediate model domain covers the northern part of Gulf of Alaska (L1, the black line). The finest-resolution model domain (L2, the small triangle [red line]) zooms in on the Prince William Sound. The nesting of the model domains is realized through the Adaptive Grid Refinement in FORTRAN (AGRIF) package, which is based on the use of pointers (Blayo and Debreu, 1999). During the development stage of the Prince William Sound system, the nesting between model domains (between L0 and L1, and between L1 and L2) was one-way. The coarse-resolution domain provides boundary conditions for the fine-resolution domain and the solution of the fine-resolution domain does not feed back to the coarse-resolution domain. The updated version of ROMS allows two-way nesting (Debreu et al., 2012).
Because of the split-explicit scheme for the barotropic and baroclinic modes, different boundary conditions can be used for these two modes. The Flather (1976) boundary condition is used along the western, southern, and eastern open boundaries for the normal barotropic velocity to allow the propagation of dynamic signals into the model domain. The Chapman (1985) boundary condition is used for sea surface height along the open boundaries. For the Chapman boundary condition, it is assumed that the dominant wave packet approaching the boundary is non-dispersive gravity waves and the phase speed is estimated based on depth. For baroclinic velocity, temperature, and salinity, we used adaptive open boundary conditions (Marchesiello et al., 2001). The adaptive open boundary conditions treat inward and outward wave packets differently. When the phase speed estimated from the interior solution is outward, these wave packets can propagate out of the model domain (Orlanski, 1976; Raymond and Kuo, 1984). When the phase speed is inward, the interior solution at the boundary is relaxed to the exterior solution with a specified time scale of three days.
When predicting regional ocean circulation, the lateral boundary conditions of the region are a necessary input and exert influence on the circulation within the domain of interest. The boundary condition algorithm was extensively tested (Marchesiello et al., 2001). Later on, boundary conditions of barotropic normal velocity that can include tides were also introduced (Wang et al., 2009). The baroclinic lateral boundary conditions for the largest domain of the Prince William Sound and Monterey Bay systems were taken from a monthly-mean climatology (Da Silva et al., 1994). Since, in general, the time span of ocean forecasting of these systems was 3 days or less and data was heavily used in the finest model domain, the use of climatological boundary conditions for the largest domain (L0) does not have significant impact on the innermost domain (L2). For longer or more accurate forecasts, the output from global operational ocean forecasting systems (Metzger et al. 2014) can be used to provide lateral boundary conditions for regional forecasting systems. For example, Liu et al. (2009) used the Navy Coastal Ocean Model output (Shulman et al., 2004) as lateral boundary conditions for a high-resolution model along the Oregon coast. For the Gulf of Mexico ocean forecasting system (Farrara et al., 2012), ocean prediction fields from the Navy Oceanographic Office were used (Metzger et al., 2014). Before such lateral boundary conditions are used, however, systematic evaluation of hindcast needs to be conducted to analyze their compatibility with and influence on the nested regional simulation.
ROMS is compatible with both OpenMP and MPI parallel computing paradigms. For real-time regional forecasting systems with dedicated computers and a forecasting range of 2–3 days, a one-hour wall clock time for integration of one day is commonly used, which allows adequate time for preprocessing of model input, data assimilation, and post processing of model output.
Atmospheric forcing fields
The generation of accurate regional atmospheric forcing fields for regional ocean forecasting systems requires the assimilation of atmospheric data in high-resolution regional models that can resolve complex boundary layer dynamics and local features of the coastline (He et al., 2004). Because of their complexity and interdisciplinary nature, oceanic forecasting systems are usually developed through group effort. That is, the various atmospheric and oceanic data assimilation tasks are typically distributed between multiple groups at different institutions. In such situations, it is important to carefully coordinate data formats, variable units, etc. Ocean model forcing is often derived using bulk formulae based on the following surface atmospheric fields: 10-m vector wind, relative humidity, air temperature, short and long wave radiation, and precipitation. Surface fluxes of momentum, heat, and freshwater are subsequently computed using the formulae of Large and Pond (1981) and Kondo (1975). The bulk formulae allow one-way ocean-atmosphere coupling through the feedback of sea surface temperature on stability, longwave radiation, and surface fluxes. Downwelling shortwave radiation that is not reflected at the ocean surface is allowed to penetrate the ocean model below the surface level with an exponential decay coefficient.
Different data assimilation methods have been used in the meteorology and oceanography communities with different levels of sophistication ranging from the relatively simple uni- and multi-variate statistical (optimal) interpolations, Kalman Filter, smoothers, and variational data assimilation methods implemented for both three-dimensional (3DVAR) and four-dimensional (4DVAR) problems (for a review, see Daley, 1991; Li and Navon, 2001). More sophisticated algorithms usually generate better analyses and forecasts, although with a higher computational cost. For the real-time nowcast and forecast systems described here, the end-to-end computing time has to be less than a day. This time requirement restricts utilization of full 4DVAR, Kalman Filter, and smoother methodologies, and instead requires the development of less expensive approaches, i.e. optimal interpolation and approximate filters (e.g. the ensemble Kalman Filter; Evensen, 2003). The systems that we describe herein are based on a 3DVAR algorithm. 3DVAR was introduced at the major meteorological centers in the late 1980s and early 1990s and is still widely used in many operational applications. The algorithm has been described in detail in two companion papers (Li et al., 2008a,b). Here we provide a brief overview of the approach, as implemented in our regional ocean forecasting systems.
The solution depends on the specifications of and . A systematic method to estimate is discussed in Li et al. (2008b). needs to be specified based on measurement and model representation errors: an example of the former is instrument noise and an example of the latter is temporal and spatial scales in the observations that do not match the variability that can be resolved by the model. The minimization algorithm used to find is limited-memory quasi-Newton method (Liu and Nocedal, 1989), because of its extensive use in solving nonlinear problems and its computational efficiency and reliability.
The field of regional ocean forecasting became much more mature after several intensive field experiments in the Monterey Bay, California, the Prince William Sound, Alaska, the Mid-Atlantic Bight, and many other regions. Some recent developments of regional ocean forecasting systems include ensemble ocean forecasting and forecasting from coupled systems.
With the increased number of regional ocean forecasting systems, occasionally there can be several systems that provide forecasting for the same region. During the period of October and November 2009, a field experiment was conducted in the Mid-Atlantic Bight to test the concept of a coastal ocean observatory. A multi-model ensemble forecasting system was developed that could combine forecasts from four individual models using a Bayesian model averaging method (Raftery et al., 2005; Wang et al., 2013b). Figure 3 compares the forecasting skill from these four individual models and ensemble forecasting from two ensemble methods. It is clear that even a straightforward average of four model forecasts can improve the forecasting skill (Figure 3e). When forecasts from four models were combined optimally (Figure 3f), the ensemble forecast provided the most accurate prediction compared with the four individual models (Figures 3a–d).
When computation power permits, ensemble forecasting from a single model by perturbing the initial condition or the boundary condition can also improve the forecasting skill and quantify the uncertainties involved in the forecasting. Using a forecasting system for the Gulf of Mexico, the results from Farrara et al. (2012) indicate that single-model ensemble forecasting is generally better than individual forecasting. The forecasting system of the Gulf of Mexico has sufficient skill to predict loop current eddy shedding events several weeks in advance.
Ocean forecasting from coupled models have also started to appear. A Coupled Ocean Atmosphere Wave Sediment model was used to provide ocean forecasts for a broad region from the Cape Hatteras to the north of the Cape Cod (Warner et al., 2010). This reflects the recent trend to provide environmental information for potential users from a single system instead of from separate atmospheric, wave, ocean, and other models. A coupled system can better represent complex interactions among the different physical components. Data assimilation methods for such coupled systems, however, are at an early stage. For example, no data was directly assimilated in the above coupled system except for temperature and salinity, which were restored to the output of a real-time ocean forecasting system from the Naval Research Lab with a relaxation time scale of 4 days.
Support field experiment
Different applications motivated the development of the regional ocean forecasting systems discussed herein. Most of these systems were developed to support field experiments. Retrospectively, both observationalists and modelers benefited from these field experiments. The practice also reflects one issue with regional ocean forecasting. For most of the world's coastal oceans, for periods without intensive observational field experiments, the data is too sparse to make skillful prediction on kilometer spatial scales. The Monterey Bay forecasting systems was developed for the Adaptive Ocean Sampling Network of 2003 (Chao et al., 2009). During the experiment, three models were used to predict the oceanic condition around the Monterey Bay. These predictions were used to guide the sampling strategy of the gliders. This concept was pushed one step further during the Ocean Observatories Initiative experiment in November of 2009. During the experiment, the forecasts from four individual models and two ensemble methods were provided to the glider planning system, which was used to produce an optimal path for a glider to reach its targeted destination points. The resulting path was then communicated to a glider during its periodic surfacing times (Wang et al., 2013b).
Search and rescue
An important application of regional ocean forecasting systems is surface drifter trajectory computations based on ocean current predictions, which can help search and rescue missions. The web user interface of all these systems provides a surface drifter trajectory tool based on ocean current prediction. Farrara et al. (2013) describe an example of surface drifter trajectory prediction based on surface current. For the Prince William Sound ocean prediction system, the model has demonstrable skill in reproducing the trajectory of real surface drifters. The prediction of drifter trajectories, however, is sensitive to small perturbations of initial location. A slight perturbation of the initial location can cause large changes in its final location. By providing ensemble drifter trajectory forecasts, the uncertainty of the drifter trajectory forecast can be assessed. The uncertainty information can be very useful in prioritizing the search area of search and rescue missions.
Oil and gas industry
The oil and gas industry can greatly benefit from regional ocean forecasting products both for its normal operation and in case of emergency. In the Gulf of Mexico loop-current eddy-shedding region, the movement of eddies westward poses risks for the operation of oil platforms because of strong current and strong vertical shear of the current associated with the eddy. If the location and the movement of these eddies can be predicted with skill, the information would be very useful for daily operations. Farrara et al. (2012) discuss an example Gulf of Mexico prediction system during the period of November to December 2010, during which an eddy shedding event occurred. The study indicates that their Gulf of Mexico forecasting system has skill to predict eddy shedding events several weeks in advance. In case of emergency, for example, the Deepwater Horizon oil spill event of 2010, trajectory forecasting can be made based on output from regional ocean forecasting systems for the treatment of spilled oil (Liu et al., 2011).
Design considerations for an ocean forecasting system of the South China Sea
The South China Sea (SCS) is the largest semi-enclosed sea in the western tropical Pacific Ocean. The circulation of SCS is influenced by many factors, the complex geometry, the seasonal cycle of monsoon, exchange with adjacent seas, the intrusion of western boundary current from the Pacific subtropical gyre, and freshwater discharge from rivers (Hu et al., 2000; Su, 2004). As an example, Figure 4 compares flow patterns from January and July of 2004 based on a data assimilation product from the Estimating the Circulation and Climate of the Ocean, Phase 2 project (ECCO2; Menemenlis et al., 2008). It is clear that during wintertime the southern part of SCS is controlled by a cyclonic gyre and during summertime by an anticyclonic gyre. These features are generally consistent with previous work in the region (Hu et al., 2000; Su, 2004; Liu et al., 2008). Another prominent feature is that during wintertime the intrusion of western boundary current to SCS region is more evident. The eddy activity is also very strong at the northern part of SCS regardless of season.
An ocean forecasting system for a large marginal sea such as SCS with kilometer-resolution is still a challenging task computationally. Nevertheless, a few such systems with different levels of maturity are starting to appear. Peng et al. (2015) configured a system that can provide atmospheric, wave, and ocean forecasting for the region. Currently the evaluation is for the prediction of tropical cyclone trajectories. Wang et al. (2015) developed a forecasting system for SCS with capability to assimilate sea surface temperature and sea surface height anomaly observations.
Development of an ocean forecasting system for SCS, based on current computational capabilities and experience from previous similar systems, needs to include following design considerations. The nesting technique still needs to be used with fine resolution (kilometer level) for regions of interest and coarse resolution (several kilometers) for a large domain that includes part of the Western Pacific and Eastern Indian Ocean. Since the potential application may include gas and oil industry (Zhang et al., 2010), the generation and propagation of internal tides should be included. The freshwater discharge from major rivers should also be included in the forecasting system. Since the biological activity may be part of its application, a biological module of the ROMS should be included and modified to fit the needs of applications. In situ observations are sparse and are not available on a regular basis, with no prospects of improvement of this situation in the near future. Satellite observation, such as altimeter sea surface height and sea surface temperature, will still be the main source of data to be used in a potential forecasting system for SCS. A fine-resolution atmosphere model with data assimilation capability that can provide high-quality atmospheric forcing fields will also be needed before a coupled system is used to provide forecasting for the region. The initial and boundary conditions for the atmospheric model of SCS can be from operational global meteorological forecasting products from institutions such as NCEP or other agencies based on evaluation. The lateral boundary condition for the ocean model can come from operational global ocean forecasting products (Metzger et al., 2014). For hindcast evaluation and SCS ocean reanalysis data products, ocean retrospective analyses, such as those provided by the ECCO project, can be used.
Discussions and conclusions
Over the past ten years, the field of regional ocean forecasting has experienced great success, and developed from a supportive activity to a routine practice. Though the individual components (e.g. ocean modeling, data assimilation, and visualization) have already been developed for other applications, the integration of these different pieces into an operational system to meet the needs of stakeholders remains a challenging task. When a new system is developed for a region, new issues always come out and demand new strategies. For instance, the first version of the Monterey Bay forecasting system had no tides. Because of the interaction of barotropic tides and complex bathymetry there, internal tide generation and propagation plays an important role in regional dynamics. To meet this challenge, tidal boundary conditions were added resulting in a system where the error of barotropic tides is less than 5% in the open ocean and less than 10% in the coastal region (Wang et al., 2009). For the Prince William Sound, Alaska, large amount of freshwater from snow melting and precipitation reaches the ocean in the form of ungauged rivers and streams. To account for the dynamic effect of freshwater input, a hydrodynamic model is used to estimate freshwater discharge at the river mouth, which in turn is used as lateral forcing for the ocean forecasting system (Wang et al., 2013a). These unique regional challenges will continue to need our attention when developing forecasting systems for a new region. Another trend is that user needs and applications are hard to meet with a single component model, be it atmospheric, oceanic, or wave. An interdisciplinary approach is needed and coupled forecasting systems are needed to assess the environment conditions and make forecasts. The development of an oceanic forecasting system for the South China Sea can benefit from previous experience and lessons learned in developing similar systems for other regions. Still, new and unexpected challenges are expected, which will require new developments and solutions.
Xiaochun Wang would like to thank Dr. Dongxiao Wang for his invitation to attend the 7th International Workshop on Tropical Marine Environmental Changes in November of 2013. Comments from two reviewers greatly helped the authors in the revision process, by pointing out new references and related work, which had not been included in the first version. This work is contribution 060 of ESMC of Nanjing University of Information Science and Technology.
Research was supported by the Chinese National Science Foundation (41328006), Nanjing University of Information Science and Technology Faculty Start-up Fund (S8113046001), the Priority Academic Program Development of Jiangsu Higher Education Institutions (PAPD) and Program for Innovation Research and Entrepreneurship Team in Jiangsu Province.