Met Office Hadley Centre observations datasets

The following is a brief description of the data set construction process. For more detail please read the papers.

## Quality control

Observations were taken from version 2.5 of ICOADS. Observations were subject to a number of quality-control checks. First the observation had to have a valid time and location. The location had to be over the ocean and have a measured sea-surface temperature. The locations and times of the reports from each ship were then checked to ensure that the ship was not travelling unrealistically fast. This check removes observations where the location or time has been misreported.

SST observations were compared to the climatological average and rejected if they were more than 8 degrees from it. Observations below the freezing point of sea water, -1.8°C, were also rejected. Observations were then buddy-checked using other nearby observations.

ICOADS is composed of a number of 'decks' which refer back to the decks of punched cards on which many of the data were originally stored. A number of observations in deck 732 were rejected before other QC checks were applied because they were obviously erroneous. It appears as if a number of 5-degree latitude and longitude areas in the deck were somehow swapped around during the digitization process so that the observations were either much warmer or cooler than the climatological average and very different to observations in neighbouring grid cells.

## Gridding

The gridding process proceeded as in HadSST2.

• For each observation that passes Quality Control:
1. Discard the observation if it cannot be gridded because the location or date is bad, or because the location is land-locked or there is no normal for that location.
2. Find the 5°x5° grid-box the observation will be in
3. Find the 1x1xpentad superob within that grid-box that the observation will be in
4. Add the observation to the list of those in that grid-box superob
• For each grid-box:
1. For each 1x1xpentad superob within that grid-box
1. Calculate the winsorised mean of the obs in that 1x1xpentad superob
2. Convert that mean to an anomaly by subtracting the superob normal
3. Calculate the area fraction of the 1x1xpentad superob that is within the grid-box
2. Calculate the area weighted winsorised mean of the 1x1xpentad superob mean anomalies
3. Count the number of obsservations in the grid-box

For a more thorough description of the process see Part 2 of the paper (1Mb).

Measurements of sea surface temperature made by Voluntary Observing Ships have been taken typically either using a bucket to collect a water sample or by measuring the temperature of the water pumped in to the ship to cool the engine (known as Engine Room Intake or ERI measurements). The material used to make these buckets has changed from wood to canvas to rubber. The material used to make the bucket affects the measurement taken, through the ability of the material to effectively insulate the water sample: canvas buckets led to measurements being generally biased cool, wooden and rubber buckets are better insulated so the biases are smaller. Measurements of engine room intake water tend to be biased relatively warm.

Drifting and moored buoys, on the whole, tend to provide fairly accurate measurements of sea surface temperature, although an individual buoy's instrumentation can be biased. Near-coincident measurements from ships and drifting buoys show that ship measurements are, on average, biased warm by between 0.1 and 0.2K relative to drifting buoy measurements.

We have attributed a known or likely measurement method to nearly all observations of sea-surface temperature made in situ, contained within the International Comprehensive Ocean-Atmosphere Data Set (ICOADS, version 2.5), except for a small fraction. Where we have no direct information about measurement method, we infer a likely method from the country of origin of the ship. This attribution of measurement method, coupled with our understanding of the relative biases between measurement methods, has allowed us to develop adjustments for each monthly, grid-box average sea-surface temperature anomaly value. However, we do not know perfectly the method used to make all measurements because some of the information is uncertain, e.g. the dates individual countries switched from using one type of bucket to another. We account for this by creating many sets of bias adjustments, varying the assumptions we make about these uncertain aspects within their likely ranges each time. The following parameters were varied in each of 100 realisations:

• Night Marine Air Temperature (NMAT) data set: 50 realisations of the bucket adjustments were generated using MOHMAT and 50 used an alternative NMAT data set based on ICOADS 2.1.
• Bucket corrections: Realisations of the bucket corrections were generated using the method described in Rayner et al. (2006). Fast ship corrections were used after 1941.
• Engine Room Intake (ERI) biases: Realisations of the ERI biases were generated by producing random AR(1) series with a lag-1 correlation of 0.99. The realisations were scaled to have a mean of 0.2K and a range drawn from a uniform distribution between 0 and 0.2K. The same value was used for all regions except the North Atlantic.
• ERI biases (North Atlantic): In the North Atlantic between 1970 and 1994, Engine Room Intake biases were generated from the means and standard errors from Kent and Kaplan 2006.
• Unknown measurements: AR(1) series with lag correlation 0.99 and range between 0 and 1 were used to assign a time varying fraction of unknown measurements to bucket. The remainder were set to be ERI measurements. The same value was used at all locations.
• ERI recorded as bucket: 30±10% of bucket observations were reassigned as Engine Room Intake measurements. One value per realisation was used and applied to all times and places after 1940.
• Canvas to rubber bucket: A linear switchover was assumed. The start point (all canvas) was chosen randomly between 1954 and 1957. The end point (all rubber) was chosen randomly between 1970 and 1980.
• Buoy-ship bias: Values were generated using the mean difference and standard error for a number of different geographical regions.

Fields of the numbers of observations associated with each measurement method and the biases estimated for each measurement method were combined to calculate the bias in the gridded temperature fields. The spread of the 100 realisations gives an estimate of the uncertainty of the bias adjustment process.

## Measurement and Sampling Uncertainty

For a more thorough description of the process see Part 1 of the paper (1Mb).

The uncertainties inherent in the bias adjustment process (described above) are most important at longer time scales and larger space scales and are most clearly expressed in long-term trends in the global average. At smaller scales, two other sources of uncertainty are also important.

1. Sampling uncertainty: arises from attempting to estimate a grid-box average temperature from a finite, and often small number of observations. The size of the sampling uncertainty depends on how well correlated sea-surface temperatures are within a grid box and their variability. It also depends on the number of observations contributing to the grid box average.
2. Measurement error uncertainty: individual measurements are prone to errors from badly sited or calibrated instruments, misreading of the thermometer, errors in transcription and transmission, and any number of effects peculiar to the circumstances of a particular observation. In HadSST3, such errors are partitioned into two components. For each ship or drifting buoy there is a constant offset (microbias) which affects all measurements made by that particular ship and a random error component that changes from one observation to the next, but has a fixed standard deviation. The random error component can be reduced by averaging together observations, but the micro- biases can only be diminished by aggregating measurements from many different ships.

## Coverage Uncertainty

When calculating area average sea-surface temperatures, such as the global average, using the gridded fields there will often be large areas for which no observations are available. Therefore, any area average will accrue an additional uncertainty from the imperfect sampling. This uncertainty can be estimated by using SST analyses that are globally complete and subsampling them at the locations where observations were available historically. Comparison between the complete fields and the subsampled fields gives an idea of the likely uncertainty accruing from imperfect sampling.

It should be noted that this component of the uncertainty depends, to a certain extent, on the statistical and physical assumptions made to produce the globally complete SST data sets used to estimate the coverage uncertainty. This possibility was explored somewhat by calculating the coverage uncertainty using three different interpolated analyses. For the regions considered in the paper, the three different analyses give similar results despite making quite different assumptions about how interpolation should be done.

## Commercial and media enquiries

You can access the Met Office Customer Centre, any time of the day or night by phone, fax or e-mail. Trained staff will help you find the information or products that are right for you.
Contact the Met Office Customer Centre