Validation report of the MACC 43- year multi- sensor reanalysis of ozone columns, version 2 Period

Transcription

1 MACC- III Deliverable D38.3 Validation report of the MACC 43- year multi- sensor reanalysis of ozone columns, version 2 Period Date: March 2015 Lead Beneficiary: KNMI (#21) Nature: R Dissemination level: PU Grant agreement n

2 Work- package 38 (VAL, validation) Deliverable D38.3 Title Validation report of the MACC 43- year multi- sensor reanalysis of ozone columns, version 2 Period Nature R Dissemination PU Lead Beneficiary KNMI (#21) Date 24/03/2015 Status Final Authors Henk Eskes, Ronald van der A, Marc Allaart (KNMI) Approved by Edith Botek (BIRA- IASB) Contact info@gmes- atmosphere.eu This document has been produced in the context of the MACC- II project (Monitoring Atmospheric Composition and Climate - Interim Implementation). The research leading to these results has received funding from the European Community's Horizon Programme (FP7 THEME [SPA ]) under grant agreement n All information in this document is provided "as is" and no guarantee or warranty is given that the information is fit for any particular purpose. The user thereof uses the information at its sole risk and liability. For the avoidance of all doubts, the European Commission has no liability in respect of this document, which is merely representing the authors view. 2

3 Executive Summary / Abstract The Ozone Multi- Sensor reanalysis version 2 (MSR- 2) provides global ozone column field time series, for the period , on a grid with 0.5 x 0.5 degree resolution and with a time interval of 6 hours. This detailed data set is produced by assimilating all independent satellite column observation data sets publicly available (15 data sets in total, BUV- Nimbus4, TOMS- Nimbus7, TOMS- EP, SBUV- 7, - 9, - 11, - 14, - 16, - 17, - 18, - 19, GOME, SCIAMACHY, OMI and GOME- 2). These satellite retrievals have been calibrated against a "ground truth" consisting of ozone column measurements from the network of Brewer and Dobson instruments. In this report, we have compared the MSR- 2 fields with individual Brewer- Dobson measurements, and with individual satellite observations. Because the same data set is used to construct the MSR- 2, it should be noted that this is not a fully independent validation. When comparing the MSR- 2 ozone column fields with the full set of Brewer- Dobson groundbased observations the mean fitted offset was found to be smaller than 0.2 DU, and both the trend and seasonal variation of the bias are negligible in the period Comparisons with individual stations do not show significant regional biases or geographical patterns (similar to what was found in the satellite comparisons). Many stations show small offsets compared to MSR- 2 of about 1%. Individual stations, however, have offsets up to about 5%, but often these stations are close to stations with very small offsets, or offsets of the opposite sign, suggesting that this is a feature of the individual ground measurements and not of the MSR analysis. This consistency check shows that the MSR calibration coefficient are well tuned to reproduce the mean levels of the ozone columns observed from the ground. To evaluate the quality of the MSR- 2 data, the Observation- minus- Forecast (OmF) and Observation- minus- Analysis (OmA) statistics have been analysed. The OmA of this dataset is often much less than 1%, which is better than for the assimilation of observations of a single sensor and is improved as compared to the MSR- 1. The model bias as estimated by the difference between OmF and OmA is in general small: for periods of a couple of days with no data, the bias remains within 1 percent. As discussed, this holds also for the period with only sparse BUV observations, although model biases of several percent as a function of latitude become visible. The RMS errors are around 2-3 percent between 1979 and 2012, which is small given that the RMS errors contain contributions from the representativity errors, forecast errors and instrumental noise. For very long time periods without any data (e.g. in 1977), longer than several months, the error becomes more than 20%. These cases may be efficiently excluded from the dataset by filtering with the forecast error estimate provided in the ozone data product, which correctly indicates large model forecast errors during these periods. Somewhat larger OmF values are observed for large solar zenith angles, close to the dark, unobserved polar cap. Note that the OmF difference contains contributions from the model forecast error, the observation error and the representativity error, and that OmA is smaller than OmF. The MSR error at a given location will therefore be smaller than the OmF RMS values quoted above. 3

4 The quality of the provided error fields has been tested by investigating the distribution of (OmF)- observed versus (OmF)- modelled. Very good results are obtained in describing the errors in data- poor (BUV, 1971) and data rich periods (2010). One exception is 1995, when the error of the MSR fields seems to be overestimated. Ozone profiles in the model are stable, and compare well with ozone sondes to within 10-20%. This helps to keep the total column forecast error low. 4

5 Table of Contents 1. System summary and model background information The ozone multi- sensor reanalysis, version Evaluation of the Ozone- MSR v2 data Intro Comparison against Brewer- Dobson OmF, OmA Evaluation of the error fields Evaluation of the ozone profiles References

6 1. System summary and model background information The MACC ozone column reanalysis product is a 43- year reanalysis of ozone column amounts performed with the KNMI TMDAM data assimilation system. This product is called the Multi- Sensor Reanalysis version 2 (MSR- 2; van der A et al., 2015) and became available in the second half of It is a major update and extension of the previous 30- year MSR- 1 reanalysis (van der A et al., 2010), which was also part of the MACC catalogue. The specifics of the model and data used to generate the 30- year Multi- Sensor Reanalysis (Ozone- MSR) are given below. Within MACC- II the entire period will be reprocessed, resulting in a MSR- 2 data set. When this becomes available, the present validation report will be updated accordingly The ozone multi- sensor reanalysis, version 2 A single coherent total ozone data set (the Ozone- MSR version 2; van der A et al., 2015) has been created from all available ozone column data measured by polar orbiting satellites in the near- ultraviolet Huggins band in the last four decades. In total 15 satellite data sets were used in the assimilation run, including BUV- Nimbus4, TOMS- Nimbus7, TOMS- EP, SBUV- 7, - 9, - 11, - 14, - 16, - 17, - 18, - 19, GOME, SCIAMACHY, OMI and GOME- 2. Table 1. The satellite datasets used in this study. The columns show (1) the name of the dataset, (2) the satellite instrument, (3) the satellite, (4 and 5) the time period,(6) the maximum distance allowed in an overpass, (7) the number of ground stations (GS) and(8) the total number of overpasses for this dataset. Name Instrument Satellite From To Dist. #GS Overpasses BUV BUV Nimbus- 4 1 Apr May TOMS- N7 TOMS Nimbus Oct May TOMS- EP TOMS Earthprobe 25 Jul Dec SBUVN07 SBUV Nimbus Oct Jun SBUVN09 SBUV/2 NOAA- 9 2 Feb Feb SBUVN11 SBUV/2 NOAA Dec Mar SBUVN14 SBUV/2 NOAA Feb Sep SBUVN16 SBUV/2 NOAA Oct Dec SBUVN17 SBUV/2 NOAA Jul Dec SBUVN18 SBUV/2 NOAA Jun Dec SBUVN19 SBUV/2 NOAA Feb Dec GDP5 GOME- 1 ERS Jun July TOGOMI2 GOME- 1 ERS Jun July km SGP5 SCIAMACHY Envisat 2 Aug Apr km TOSOMI2 SCIAMACHY Envisat 2 Aug Apr km OMDOAO3 OMI Aura 1 Oct Dec km OMTO3 OMI Aura 1 Oct Dec km GOME2A GOME- 2 Metop- A 4 Jan Dec km

7 BUV TOMS- N7 TOMS- EP SBUV07 SBUV09 SBUV11 SBUV14 SBUV16 SBUV17 SBUV18 SBUV19 GOME- 1 SCIAMACHY OMI GOME Fig. 1 Data availability for each satellite instrument used in the ozone MSR- 2. The ozone MSR is produced in two steps (van der A et al., 2015). First, the latest reprocessed versions of all available ozone column satellite datasets are collected, and are corrected for biases as function of solar zenith angle, viewing angle, time (trend), and stratospheric temperature using Brewer/Dobson ground measurements from the World Ozone and Ultraviolet Radiation Data Centre (WOUDC). The list of stations can be found in van der A et al. (2015). Subsequently the debiased satellite observations are assimilated within the ozone chemistry and data assimilation model TMDAM driven by meteorological analyses of the European Centre for Medium- Range Weather Forecasts (ECMWF). The MSR- 2 reanalysis upgrade described in this report consists of an ozone record for the period The chemistry- transport model and data assimilation system have been adapted to improve the resolution, error modelling and processing speed. BUV satellite observations have been included for the period The total record is extended with 13 years compared to the first version of the ozone multi sensor reanalysis version 1, the (MSR- 1). The latest total ozone retrievals of 15 satellite instruments are used. The resolution of the TMDAM model runs, assimilation and output is increased from 2x3 degree to 1x1 degree. The analysis is driven by three- hourly meteorology from the ERA- interim reanalysis of ECMWF starting from 1979, and ERA- 40 before that date. The ozone chemistry parameterization in the stratosphere has been updated. The data assimilation method is a sub- optimal implementation of the Kalman filter technique and is based on a chemical transport model (a simplified version of the TM5 model) driven by ECMWF meteorological fields. The chemical transport model provides a detailed description of (stratospheric) transport and uses parameterisations for gas- phase and ozone hole chemistry. The vertical grid is hybrid- pressure and consists in 44 levels extending from 0.1 hpa to the surface. 7

8 The MSR data set results from a 43- year data assimilation run with the 14 corrected satellite data sets as input, and is available on a grid of degree with a sample frequency of 6h for the complete time period ( ). Fig. 2 The number of annual satellite observations used in the compilation of the MSR- 2 (black line and diamonds). Note the logarithmic scale of the y- axis. The red asterisks show the number of annual ground observations used for the MSR- 2. Apart from the ozone column fields also detailed error estimate fields are provided with the same 6h sample frequency. This is a special feature of the sub- optimal Kalman filter approach. The error modelling accounts for: A reduction of the error due to the assimilation of new satellite data in the analysis step A growth of the forecast error with time (model error term) Displacement of information (or the error field) by the wind field Correlated errors for nearby air masses An example of an instantaneous error field (available in the datasets) is shown in figure 3. All the above aspects are clearly visible in the spatial distribution of the error field. The Multi- Sensor Reanalysis can be found at 8

9 Fig 3. The instantaneous error field during the assimilation of the ozone data for the MSR- 2 at 12 UTC on 30 Oct 1970 (top panel) when satellite data is sparse and on 26 June 2006 (bottom panel) when the number of satellite observations is more or less at its peak. The location of the last assimilated measurements coincides with the lowest errors. The model error term leads to an increase of the forecast error with the time passed since the last analysis. The advection of the error is visible as distortions of the orbit shape. 9

10 Fig 4. Southern hemisphere October monthly mean ozone column amounts for a period of 44 years, showing the deepening of the ozone hole during the 80's. In the years the amount of satellite observations was very sparse, and the ozone field has been greyed out using the ozone error field as indicator of the quality of the results. 2. Evaluation of the Ozone- MSR v2 data 2.1. Intro The Ozone- MSR version 2 is generated by assimilating 15 available satellite ozone column data sets, using a three- dimensional model to describe transport and chemistry of ozone in the atmosphere. These ozone satellite data sets have been "calibrated" by means of surface Brewer and Dobson observations. Because of this calibration, the overall global biases in the product will be small by construction. However, these calibration factors are global and do not say much about the comparisons at individual locations. Similarly, the satellite observations are assimilated, resulting in a small overall bias between the model and the satellite after the analysis step, but this does not guarantee a good comparison with individual satellite observations before the analysis (comparison with the "first guess" or after the forecast between observations). In this validation report, we focus on the comparison of the assimilated product with both individual Brewer- Dobson surface measurements and with the individual satellite measurements before these are assimilated: the observation- minus- forecast (OmF) differences. Furthermore, we conducted a comparison with ozone sondes to check if the vertical profile is reasonable and if it is stable over the long integration period Comparison against Brewer- Dobson Fig. 5 shows the MSR- 2 level 4 data set compared to ground- based data from the WOUDC database. Note, first of all, that these ground observations are already used in generating the MSR data, so this is not a fully independent data set. The validation described here is 10

11 Fig 5. Fitted offset (MSR- 2 minus ground) between the MSR- 2 level 4 data and all selected Dobson and Brewer ground measurements in the period partly a consistency check and shows the difference between the level 4 (assimilated) satellite data set compared to the ground truth. However, the groundbased data has only been used to determine a limited set of global correction factors to improve seasonality and trends of the satellite data compared to the Brewer/Dobson network. This is based on comparisons with all stations simultaneously (van der A et al., 2010). Comparisons with individual measurements of the stations separately still provide useful validation information. Figure 5 shows that, on average the offset is small. A large amount of stations have a yellow color, meaning that the overall bias with the final MSR- 2 product is small. No systematic structures are obvious in the geographical distribution, and the yellow points occur in all regions. However, also outlier stations are visible, often close to a station with a very small offset, which suggests that the offset may be station related. Examples are Antarctica, where yellow, orange and light green points occur, suggesting that the bias of the MSR- 2 is small in this region. In the Arctic there are a few purple, orange, but also green points, again suggestion that the mean bias may be small. A dense net of observations exists in Europe. Here we also observe a few red, purple, green and blue points close to yellow points. This strongly suggests that the individual stations may have offsets up to +- 5 %. When fitting the MSR data to the full set of groundbased observations, the mean fitted offset was found to be smaller than 0.2 DU, and both the trend (0.02 DU/year) and seasonal variation (effective ozone temperature dependence of DU/K) were negligible. 11

12 Fig. 6. The global distribution, gridded on 1 x 1 degrees, of the observation- minus- forecast in DU of the MSR- 2 dataset averaged for the month January The MSR- 2 data for this month is based on satellite observations from SBUV, GOME, SCIAMACHY, GOME2 and OMI. The inset shows the same OmF distribution for the MSR OmF, OmA Figure 6 shows an example of the OmF gridded for January 2008 as function of geographical location. In general the mean OmF is between 3 and +3 DU. In the northern latitudes some higher variations are found caused by the vicinity of the polar night where observations are lacking. No obvious patterns as function of ground elevation or surface type are visible. Compared to the OmF of the MSR- 1 data set (see inset) the deviations have become somewhat smaller. For the assimilation of observations of a single sensor with full daily coverage, the observations at a certain location are typically available once a day and therefore the OmF statistics result from an unconstrained model forecast time step of roughly 1 day. Smaller time steps occur near the poles. For the assimilation of SBUV observations the revisit time is typically only 1 week, given that the assimilation has a spatial correlation length of 500 km. For the assimilation of data from multiple sensors this is different, and time steps between observations vary from half an hour to one day. In this case the OmF, OmA and ozone assimilation results are more restrained by the observations. In Fig. 7 the OmF is shown as function of the geo- parameters solar zenith angle, latitude, cloud fraction and viewing angle for January In addition the root- mean- square (RMS) values of the OmF and OmA are plotted in the figure. The systematic effects found for these parameters are all much smaller than the typical RMS. On average the RMS difference between new satellite observations and the short- range model forecast (1 day) is small: about 6 DU, or roughly 2% for this month. The RMS error of the OmF is smaller than the observational error compared to ground observations, because representative errors of the 12

13 Fig. 7. The observation- minus- forecast in DU (blue line) and the observation- minus- analysis (red line) as a function of solar zenith angle (a), latitude (b), cloud fraction (c), and viewing zenith angle (d). The dashed lines represent the RMS value of the observation- minus- forecast (blue) and the observation- minus- analysis (red) distribution. All data are averaged over January satellite observations compared to level 4 data are likely to be smaller than for satellite observations compared to ground observations. For high solar zenith angles the RMS value increases, because these measurements are usually associated with the highly variable ozone concentrations in and around the polar vortex. Plots similar to Fig. 7, shown in Fig. 8, show that the bias between the forecast and the satellite columns is generally smaller than 1% after 1979, apart from the highest solar zenith angles. The bias between analyses and observations is in general even smaller (less than 1 DU), which shows the effect of data assimilation. Compared to the MSR- 1 results (van der A et al., 2010), the bias has slightly decreased probably as a result of the higher spatial resolution for the MSR- 2 and the improved data assimilation. The bias in the analysis is negligible. The RMS is typically of the order of 2-3%. Figure 9 is similar to Fig. 7 but here the statistics are for the complete year 1971, when only BUV observations were available. In this period the time between observations of the same air mass is generally much longer than 24 hours. There is no plot as function of viewing angle since BUV is observing under a fixed angle in nadir direction. The mean OmF and OmA values are in general still small (less than 5 %), but the RMS values are higher than the period after 1979, up to 10 %. The region on the Southern hemisphere shows a high forecast error, especially visible in the plot as function of latitude. This is because all BUV data in and 13

14 Fig. 8. The observation- minus- forecast in DU (blue line) and the observation- minus- analysis (red line) as a function of latitude for Jan 1986 (top- left), Jan 1995 (top- right) and June 2000 (bottom), to be compared with Fig. 7- b. The dashed lines represents the RMS value of the observation- minus- forecast (blue) and the observation- minus- analysis (red) distribution. around the South Atlantic Anomaly has not been made available and therefore the forecasts are for longer time steps. The structures seen in the latitude dependence in Fig. 9 reflect the ozone model errors (drifts), which may be related to issues with the ERA- 40 meteorology mainly in the tropics (Uppala et al., 2005) and the removal of BUV data in the South Atlantic Anomaly region Evaluation of the error fields An example of an error field supplied in the ozone product is given in Fig. 3. In TMDAM the forecast error covariance matrix is written as a product of a time independent correlation matrix and a time- dependent diagonal variance. The various parameters in this approach are fixed and are based on the observation minus forecast (OmF) statistics accumulated over the period of one year (2000) using GOME observations. This method produces detailed and realistic time- and space- dependent forecast error distributions. The error field is an important component of the assimilation process and determines the relative contribution of the observations and model to the analysed ozone amount, as prescribed by the analysis equation of the Kalman filter, which leads to a reduction of the error of the model distribution close to the observations. In between measurements the error variance is growing due to the model error, and the variance field is advected in the same way as ozone itself (van der A et al., 2010). Those three aspects of the time dependence of the error field can be recognized in Figure 3. Since the number of observations is much higher in 14

15 Fig. 9. The observation- minus- forecast in DU (blue line) and the observation- minus- analysis (red line) as a function of solar zenith angle (a), latitude (b), and cloud fraction (c). There is no change in viewing angle in this period, because BUV has only nadir observations. The dashed lines represent the RMS value of the observation- minus- forecast and the observation- minus- forecast distribution. All data are averaged over (and available for more locations) the error is in general much lower than for instance in Since no observations over the Antarctic exist in summertime (bottom figure), the error is much higher there, although it is partly reduced by advection of ozone information into this region. To check if these parameters are valid for the MSR- 2 in other years, we have compared the observed OmF (by comparing the forecast with individual observations) with the estimate of the OmF (computed from the sum of the observation error covariance and the model forecast error covariance). The latter is calculated from the combination of the model forecast error as computed in TMDAM and the given individual measurement error bars on the observations. This approach can be seen as an extension of the much used χ 2 test, which checks basically if the mean of both quantities are consistent. In Fig. 10 we show this comparison for two extreme cases in the MSR- 2 time series, the first for the complete year 1971 (see Fig. 10, top left) when only the sparse BUV observations were available, and the second for 30 April 2010 (see Fig. 10, bottom right) when there is a very high density of observations from a wide range of satellite instruments which strongly reduces the errors. The grey area shows the number of observations with that specific forecasted OmF (bin size is 0.2 DU). The black line indicates 15

16 Fig 10. OmF (Observation minus Forecast) from the data assimilation as function of the theoretical OmF as calculated from model error and the individual measurement errors. The grey area and y- axis on the right indicate the number of observations per OmF value (bin size is 0.2). the perfect situation where the observed OmF in the bin would be equal to the forecasted OmF. As one can see from the Figures, in both cases the OmF values are remarkably comparable, especially in the grey area corresponding to the bulk of the observations, and the strong error reduction due to the extra observations is well described. In the figure we also showed two intermediate situations, namely January 1986 and In 1986 we have also a satisfactory result, although the error is a bit overestimated (conservative). In 1995 we find the largest differences, and the too conservative error estimates in the file do not represent the high quality of the analysis, which is in fact comparable to 2010 with it's high density of observations. Based on these results we decided not to change the model error parameters in the Kalman Filter as compared to MSR- 1. An improvement could be achieved by fitting the error growth coefficients for each period separately, but this is a lot of work Evaluation of the ozone profiles The MSR- 2 data product consists of ozone column fields only, because this aspect of the model is tightly constrained by the model. Although the data assimilation is analysing total columns, the model describes the vertical distribution of ozone in the stratosphere and troposphere, and an update of the total ozone is distributed over the vertical profile in ratio to the modelled profile. If the profile shape in the model deteriorates, this will also affect the quality of the analysed ozone columns and will result in larger forecast errors. 16

17 Fig. 11. Validation of the ozone distribution in the model using a selection of about ozone sonde stations. The mean difference between the sonde and MSR- 2 ozone profile is shown for 2001 (blue) and 2011 (red). For reference the average ozone profile of the sondes is shown (dotted line). We have checked the ozone distribution in different years by comparing with ozone sondes (see Fig. 11). The validation is a bit limited because the 3D ozone field is only stored at the beginning of each month. No significant drift in the ozone profile shape was visible over a time period of 10 years of assimilating ozone columns. As can be seen in Fig. 11 the bias between ozone sondes and the model ozone profiles is less than 10-20%, which is satisfactory, given that the ozone profile is not at all constrained by observations and the chemistry in the model is very simplified. These relatively small profile biases will not significantly impact the ozone column analysis, as demonstrated by the good performance of the assimilation system. 17

18 References van der A, R. J., Allaart, M. A. F., and Eskes, H. J.: Extended and refined multi sensor reanalysis of total ozone for the period , Atmos. Meas. Tech. Discuss., 8, , doi: /amtd , van der A, R. J., M. A. F. Allaart, and H. J. Eskes, Multi sensor reanalysis of total ozone, Atmos. Chem. Phys., 10, , doi: /acp , chem- phys.net/10/11277/2010/, Eskes, H. J., van Velthoven, P. F. J., Valks, P. J. M. and Kelder, H. M.: Assimilation of GOME total ozone satellite observations in a three- dimensional tracer transport model, Q. J. R. Meteorol. Soc., 129, ,