Sections below: Overview | Using CARS | Software | Bottom water | Access | Updates | Making CARS | References | Sample plots


This webpage gives an overview of the CSIRO Atlas of Regional Seas, and describes how to access and use it. Individual sections can be accessed by the shortcuts above.

CSIRO Atlas of Regional Seas (CARS)

CARS is a digital climatology, or atlas of seasonal ocean water properties. It comprises gridded fields of mean ocean properties over the period of modern ocean measurement, and average seasonal cycles for that period. It is derived from a quality-controlled archive of all available historical subsurface ocean property measurements - primarily research vessel instrument profiles and autonomous profiling buoys. As data availability has enormously increased in recent years, the CARS mean values are inevitably biased towards the recent ocean state.

A number of global ocean climatologies are presently available, such as NODC's World Ocean Atlas. CARS is different as it employs extra stages of in-house quality control of input data, and uses an adaptive-lengthscale loess mapper to maximise resolution in data-rich regions, and the mapper's "BAR" algorithm takes account of topographic barriers. The result is excellent definition of oceanic structures and accuracy of point values.

  Dynamic height 1000/2000m
revealing global deep
circulation, derived from

Mapped Properties
Water propertiesVersionsUnits Derived propertiesUnits Versions
temperature2009, Argo-only DegC (detail) Bottom water . 2009
salinity 2009, Argo-only PSU (detail) Mixed Layer Depth m2009, Argo-only
oxygen2009 ml/l (detail) Dynamic Height wrt 2000m m 2009, Argo-only
nitrate 2009 umol/l (detail) . . .
silicate 2009 umol/l (detail) . . .
phosphate 2009 umol/l (detail) . . .


CARS2009 covers the full global oceans on a 1/2 degree grid, but until June 2011 only included temperature and salinity fields. The T and S fields were created in July 2009 and were was based on World Ocean Database 2005 (WOD05) [July 2008 Update], surface-pressure-corrected Argo global archives to May 2009, WOCE Global Hydrographic Program (v3.0), and many other datasets available up to 2008. See the updates section below for history of occasional sub-version releases. The nutrient fields created in June 2011 were based on WOCE and WOD09 (March 2011 download).


An alternative version of CARS is produced every few months which uses only Argo data. All available Argo data is used (both Real-Time and Delayed-Mode) and it is subjected to extra local screening before use.


CARS2006 covers the southern hemisphere and tropical north, to 24N (Indian and Pacific) and 10N (Atlantic), on a 1/2 degree grid. The quantities mapped are 4 nutrients as well as temperature and salinity. See also the CARS2006 page.

Other versions

A temperature-only version based on the CSIRO Quality Controlled Ocean Temperature Archive ( QuOTA) is also available. This uses XBTs and hence a much richer temperature dataset than the standard CARS. It is available in a monthly-values netCDF file spanning the Indian and South-West Pacific Oceans, at QuOTA Data.

The earliest version of CARS is still accessible at CARS2000 .

Information for a given year

CARS does not provide information for any given year. CARS is created by averaging/interpolating all available oceanographic profile data, most of which was collected in the last 50 years. Especially when trying to provide an estimate at every location and every depth in the world's oceans, there is not enough data to resolve any one year, so we ignore the year of collection of each observation and retain only the day-of-year - and then fit a mean and mean-seasonal-cycle at each point.

In the Western Equatorial Pacific and the Gulf of Carpentaria we apply corrections for interannual signals, but in general any such signals will effect the maps, both by these signals being aliased into spatial structure or seasonal cycles, and by biasing towards the interannual anomaly of data-rich periods.

Density inversions

Many CARS T-S profiles are not dynamically stable. Two of many reasons for this are that input data is often not stable (and not all T and S points come from the same measured profile), and each depth level is mapped essentially independantly of every other.

The CARS T and S mean fields could be adjusted so that every vertical profile is stable. Further, CARS could be adjusted so that any computed seasonal profile is stable. However, the stabilization routines available may not be suitable for every season in every water body. They may adjust the wrong parts of the profiles, or adjust S instead of T or vise versa. They may adjust well mapped values to accommodate poorly gap-filled sub-bottom values, or high variability surface waters.

Adjustment so that all seasonal profiles are stable requires the choice of whether to achieve this by allowing adjustment of the mean or not. If adjust the mean, then may introduce errors from poor seasonal harmonics. If do not adjust the mean then can only stabilize by reducing the magnitude of seasonal cycles, which we have taken great efforts to obtain at realistic magnitudes.

CARS is as computed by the loess smoother. We leave it to the user to apply dynamic stabilisation as required (and the distribution of modifications required is itself informative to the user.) One method is described in "A Conserved Minimal Adjustment Scheme for Stabilization of Hydrographic Profiles", Peter C. Chu and Chenwu Fan, JAOT, 2010. If stable seasonal profiles are required it may be preferable to first evaluate the profiles then stabilize, rather than adjusting the CARS seasonal harmonics.

T inversions

Temperature inversions occur in many parts of the world oceans, but are not observed within some depth bands in many other regions. Temperature as mapped by the CARS methods has inversions in the mean field where they are not known to occur naturally, and many more inversions in seasonal computed values. We have created a database describing broadly where T inversions are not observed in the worlds oceans and modified mean and seasonal harmonics so that inversions in computed seasonal profiles are minimised.

Using CARS

CARS2009 is stored and available online in netCDF files. The CARS2006 T and S fields are also available via ftp in a collection of GIS-suitable ASCII files.

CARS2009 Argo-only Mapped without Argo T & S Data Bottom water

It is mapped on version 3 CSL (CSIRO standard depth levels), on a .5X.5 degree grid covering the region 0E - 360E, 75S - 90N. Seasonal cycles are estimated in the upper ocean. The following variables are likely to be most useful. The last few variables are more obscure and may not be present in all versions.

Name Description
lat,lon grid point locations
depth depths of the 79 mapping levels (in metres)
depth_ann depths of the levels for which annual cycles are estimated
depth_semiann depths of the levels for which semiannual cycles are estimated
mean estimate of mean value
an_cos cosine of annual cycle
an_sin sine of annual cycle
sa_cos cosine of semiannual cycle
sa_sin sine of semiannual cycle
nq number of data points used in mapping. Also, values <=1 indicate there was insufficient data to map (and any value at that point is a result of postmapping gap filling by vertical extrapolation of other desperate measures, because modellers want a value at every wet point.)
sa_sin sine of semiannual cycle
std_dev standard deviation of observations (locally-weighted standard deviation of the "data-grab" for each grid point)
map_error estimate of mapping Standard Error of the Mean
RMSspatialresid RMS of residuals w.r.t the spatial mean, ie RMS of difference between data and mapped mean field at data locations.
RMSresid RMS of residuals w.r.t the full mapping, ie RMS of difference between data and mapped seasonal field at data locations and day-of-year.

Note that the seasonal coefficients can be conveniently treated as complex numbers, eg: an = an_cos + i*an_sin. The Matlab access software uses this approach.


To construct the temperature map for mid-February at 200m depth:
Extract variable "depth" and find that 200m is at level 25.
Extract at level 25, and in the region required:
Evaluate at day-of-year 45 (mid February)
t = 2pi x 45/366
feb = mean + an_cos*cos(t) + an_sin*sin(t) + sa_cos*cos(2*t) + sa_sin*sin(2*t)

Impossible values

Unrealistic values can arise from fitting sinusoidal seasonal cycles. The most obvious example occurs when nutrients are seasonally depleted so that the "true" seasonal curve would have a peak or two and intervening zero flatspots (see red curve below). A best fit sinusoid will undershoot, creating negative values (black dashes). The user will therefore obtain a more realistic representation by setting any such negative values to zero. [click on picture to enlarge]

Software (a limited package of access routines)

Many popular software products have interfaces or library routines to interrogate and extract data from netCDF files.

A small package of unsupported Matlab access routines can be downloaded from the CARS ftp site (see access details). They will require local installation of the Matlab-netcdf "toolbox" [note: the CARS team do not maintain or support this or any other netcdf interfaces.]

Matlab functions for CARS access
Name Use
getchunk extracts a 3D chunk
getmap extracts a single depth layer or horizontal slice
get_clim_casts extract vertical profiles at the lats/longs and optionally time of year (and so can be used to create sections, for example.)
get_clim alternative to get_clim_casts
atday evaluate the mean and temporal harmonics at a particular day-of-year.
atdaypos as for atday, but also interpolates to desired locations
dep_csl, csl_dep convert between depth (m) and CSIRO standard depth levels (CSL)

Bottom water properties

Seawater properties at the ocean floor are required for some analyses. CARS is derived from oceanographic profile data, and such profiles typically do not approach closer than 5 to 10m from the ocean floor (to prevent damage to the instruments.) However we can estimate the seafloor properties because in many places there is nearby deeper water which has been sampled. In making CARS, at the "coastline" for each depth level we extrapolate landwards by one gridpoint. This means we often have a CARS value just below, as well as above, the bottom. The seafloor maps are computed by vertically interpolating those near-bottom CARS values to the ocean depth at each gridpoint. Where we cannot interpolate because there is no below-bottom value, we use the value immediately above the bottom. However there are still a portion of grid points for which there is no value immediately above the bottom, due to absence of nearby deep observations (as particularly occurs in hollows in the seafloor.) The bottom water fields necessarily have gaps in these locations.

As well as mean values, seasonal values, seasonal range, and standard deviation values may also be provided for the seafloor, and these are all derived from CARS in the same way. These are presently only available for CARS2006.

Accessing CARS

If publications arise from work that makes use of CARS, please send us a copy, via email
or to:
Jeff Dunn
CSIRO Marine Laboratories
GPO Box 1538
Hobart, TAS, 7000, Australia

Before retrieving data please read the conditions below and acknowledge that you accept them. Acceptance of the conditions will activate the download addresses page.


The User acknowledges that the Product was developed by CSIRO for its own research purposes. The CSIRO will not therefore be liable for interpretation of or inconsistencies, discrepancies, errors or omissions in any or all of the Product as supplied.

Any use of or reliance by the User on the Product or any part thereof is at the User's own risk and CSIRO shall not be liable for any loss or damage howsoever arising as a result of such use.

The User agrees that whenever the Product or imagery/data derived from the Product are published by the User, the CSIRO Marine Laboratories shall be acknowledged as the source of the Product.

The User agrees to indemnify and hold harmless CSIRO in respect of any loss or damage (including any rights arising from negligence or infringement of third party intellectual property rights) suffered by CSIRO as a result of User's use of or reliance on the Data.

If you accept these conditions then you may download CARS by any of the following methods. Note that CARS is stored in one large (up to 400MB) netCDF file per property.

What is available where Reliability
NetCDF files for several versions of CARS
also some Matlab software Has worked for years - complain if it doesn't
CARS2009 and latest Argo-only netcdf files ftp from :
in pub/dunn/cars2009/
web-based select-subset-and-download tools for CARS2009 AODN Thredds Server you might be lucky
web-based select-subset-and-download tools for CARS2009 AODN Portal no harm in trying


Large or small revisions may be released from time to time, to correct errors or incorporate new data. These updates will be recorded in the table below and detailed in the CARS release notes.

Date Level Comment cars_version
Sep 09Major First release 2009.A.1.0
27/10/09 minor Correct all values at 360E 2009.A.1.0
6 Jul 2010 minor Repair deep gap filling, Baltic salinity 2009.A.1.1
30 May 2011 major Released cars2009 global nutrient fields, replacing limited-domain cars2006 fields 2011.1.0
22 Mar 2012newMake Argo-only version available -

About making CARS


CARS2009 has been developed by CSIRO Marine and Atmospheric Research with the support of the following projects:
Bluelink logo
BlueLink - Ocean Forecasting Australia
IMOS logos
Integrated Marine Observing System (IMOS)
CSIRO Wealth from Oceans Flagship

Data sources

The atlas is based on the BOA (BLUElink Ocean Archive), which in turn is based on a number of datasets including:
Major input datasets URL
World Ocean Database 2005 - including July 2008 update (WOD05) NODC home page
WOCE WHP3.0 WOCE Data resource page
CSIRO data holdings CMAR home page
Argo floats Argo home page
TAO array TAO home page

Data was screened for duplicates and bad positions, outliers to globally mapped t-s relations, and outliers of residuals to intermediate mappings.

Mapping method

The mapping algorithm is adapted from the weighted least-squares quadratic smoother, known as a "loess" smoother. Quadratics were fitted in the horizontal plane, with bathymetry-influenced weighting. {Local profile shape is used to project next-level values to the mapping depth to fill any gaps in input profiles, but no other vertical fitting is used.} Annual and semiannual harmonics were simultaneously fitted, and these are damped at deeper levels until first semiannual and then annual fit is extinguished.

For every mapped point, a (variably zonally stretched) radius was calculated that provided 400 data points at that depth. Hence, in ocean of uniform depth, the data source region forms an ellipse. However, the BAR and TAR bathymetry-influence systems (Dunn & Ridgway, 2002) distorts this ellipse, for example extending it along the shelf where the grid point is on the shelf, or truncating it at topographic barriers such as the subsurface Chatham Rise, or the Central American isthmus. An important characteristic of this type of mapping is that length scales are automatically adapted to data density, providing maximum resolution in areas of high sample density.

A recent alternative to BAR is called DLU (Distance LookUp). It provides topography-adjusted water-path distances at a large number of depth levels (whereas BAR has been computed for only 0m, 100m and 1000m.) DLU is potentially useful for a range of applications and we are happy to share both the software and pre-computed lookup tables. See the DLU page.

A value is provided everywhere the ocean is deep enough, and one gridpoint landwards of each depth "shoreline" (this allows interpolation between gridpoints to locations near the shorelines).

Profile shape, inversions

CARS is essentially mapped on depth planes. There is a very small cross-influence between adjacent depth layers, but no explicit attempt at preserving profile-shape fidelity - the focus is on getting the best estimate of mean and seasonal cycle at each depth level. So, in places where different subsets of the population of profiles are used at different depth levels (because run out of water depth or profiles stop short of bottom or have gaps), and especially where the adjacent-depth data-grab is sampling different water masses to different degrees, then a poor representation of profile shape is expected.

Despite this, because CARS is generally very true to the data, profile shape is usually good, but widespread small T inversions do occur in the mean, and a very few isolated locations are quite pathological. There is a much greater potential for T inversions in seasonal profiles. The T fields have had a minimalist treatment to reduce seasonal and mean T inversions, mostly by damping upper level seasonal cycles. However T inversions occur naturally in vast regions of the worlds oceans, so we have coarsely mapped the regions where T inversions do not naturally occur and only apply inversion suppression in those areas.

Dynamic instability is not uncommon in CARS, especially in the seasonal fields. Density inversions are widespread in the world's hydrographic data. Our Quality Control seeks to minimise this (complete elimination can be dangerous though - you could throw away most of the valid data in the Japan Sea if a brute force approach is used.) Even if the input data had no density inversions, some would still occur in CARS, as explained in the paragraphs above. We do not attempt to correct these.

CARS is designed to provide an estimate of water properties, at each location in the grid, that is faithful to the data. We have not attempted to fudge the results of averaging observational data in an attempt to create a physically consistent model of the oceans.


Mapping nutrients is more problematic than mapping temperature and salinity because measurement accuracy is poor and has varied over time and between countries, and data distribution in space and time is terrible, and the real variability scales are pretty enormous.

Many minor data sources are not worth using because data is just too suspect, so main sources are NODC/World Ocean Database and WOCE. WOD have over time more rigourously screened nitrate and other nutrients: with each edition of WOD they have thrown out more of the data (yes, actually getting less and less data in some areas.) There are some clear cases where good but unusual data has been rejected, which is inevitable when forced to use automated screening systems, but ignoring WOD screening flags subjects you to all the subtle biased data they have identified over the years. So, we use their screening and also apply our own on top of that, then struggle to fill in the huge gaps!

Sometimes I have deliberately rejected real but unhelpful data. For example, a CSIRO cruise in 2004 specifically targetted cold core eddies off WA. This dense sampling of very high nutrient water will distort the maps. Removing it gives a much better representation of typical ocean condition, but does mean the high variability is less well depicted.

In this product you will see structure in the fields that is more related to the data distibution, local concentrations of observations with systemic biases, and the mapping system itself. This structure is particulary found in the seasonal signal and variability fields. However, the representation of the mean property values throughout the ocean is, we believe, reasonable. The knowledgable user will in most cases be able to discern the structure arising from imperfect data and that which truly represents the state of the oceans. This gives insights which would be lost if more aggressive screening and smoothing had been applied.

Jeff Dunn, June 2011


- primary CARS citation:

Ridgway K.R., J.R. Dunn, and J.L. Wilkin, Ocean interpolation by four-dimensional least squares -Application to the waters around Australia, J. Atmos. Ocean. Tech., Vol 19, No 9, 1357-1375, 2002

- algorithm details:

Dunn J.R., and K.R. Ridgway, Mapping ocean properties in regions of complex topography, Deep Sea Research I : Oceanographic Research, 49 (3) (2002) pp. 591-604

- CARS seasonal fields and MLD:

Scott A. Condie and Jeff R. Dunn (2006) Seasonal characteristics of the surface mixed layer in the Australasian region: implications for primary production regimes and biogeography. Marine and Freshwater Research, 2006, 57, 1-22.


CARS2009 metadata record: MarLIN record: 8539, Anzlic identifier: ANZCW0306008539

This webpage ( ) is itself the authoritative reference for CARS2009.

Sample plots - Selected images from CARS2009

CARS2009 Figures

Restricted access to CSIRO Marine and Atmospheric Research staff only

Access to CARS from within the Hobart CSIRO network

Production details and progress and development records of creating CARS2009


Legal Notice and Disclaimer | Copyright | Privacy | Last modified: 10/4/2014 | Page author: Jeff Dunn (now retired) | Contact: