HTML
--> --> -->The assimilation of MWHS data over land requires an estimate of emissivity. One option is to use microwave land emissivity models (Weng et al., 2001); another is using an emissivity atlas (Karbou et al., 2005a); and a third is to retrieve emissivity dynamically from microwave window channels (Karbou et al., 2006).
Physical land emissivity models require several inputs that must be estimated or derived from the NWP model or taken from the climatology. However, many input values are not usually available from land surface models. Therefore, most assimilation systems retrieve the land surface emissivity directly from the microwave observations at 50 GHz for temperature sounders (e.g., AMSU-A) and 89 GHz for humidity sounders (e.g., those onboard other satellites like NOAA or
EUMETSAT——referred to as MHS hereafter to make the distinction with MWHS/FY-3B) (Karbou et al., 2006).
In this study, we consider an approach in which emissivity is retrieved dynamically from MWHS/FY-3B channel 1 [150 GHz (vertical polarization, V)], as no 89 GHz channel is available (Zhang et al., 2008, Lu et al., 2011), and compared with the use of an emissivity atlas based on 89 GHz observations from other sensors.
The structure of the paper is as follows: section 2 describes the emissivity and experimental settings for adding data over the snow-covered surfaces; section 3 presents the results; and conclusions are provided in section 4.
Figure1. Difference between the retrieved dynamic emissivity from 150 GHz (V) of MWHS/FY-3B and the 89 GHz emissivity atlas at 0000 UTC 3 October 2013. Only regions where a successful 150 GHz retrieval can be performed are shown (see main text for details).2.1. Emissivity estimation
To assimilate MWHS/FY-3B over land, two approaches to specify surface emissivity are compared. One is to use an 89 GHz emissivity atlas derived from instruments on other satellites. This evolves over time using a Kalman filter and the emissivity is parameterized as a polynomial of the scan angle θ to account for the different effective polarization within the swath, assuming the emissivity changes occur slowly (Krzeminski et al., 2009). The other is to retrieve emissivity from the 150 GHz (V) channel; MWHS/FY-3B does not have an 89 GHz channel (Table 1). 150 GHz is more sensitive to water vapor than 89 GHz, so is useful when the total column water vapor is low, e.g., at high latitude, but more problematic at low latitudes. Figure 1 shows the difference between the retrieved dynamic emissivity from 150 GHz (V) of MWHS/FY-3B and the 89 GHz emissivity atlas at 0000 UTC 3 October 2013. For most locations, the differences are within 0.05. Larger differences are found over the snow-covered surfaces, as would be expected, since the emissivity of snow is more variable both in time and frequency than that of snow-free surfaces. This means we have low confidence in the emissivity estimate in these regions, as we do not know whether to trust the atlas or the 150 GHz dynamic emissivity estimate more. 150 GHz is more sensitive to the atmosphere, but the retrieved emissivity is also more representative of the emissivity at 183 GHz, due to being closer in frequency. We therefore compare using the atlas alone to using the retrieved 150 GHz emissivity. To ensure that 150 GHz has sufficient surface sensitivity to perform a reliable emissivity retrieval, we require that the surface-to-space transmittance is larger than 0.5. If this condition is not fulfilled (e.g., in tropical areas), the emissivity atlas is used instead. In addition, the dynamic emissivity is only used when the estimate differs from the emissivity atlas by less than 0.2. This threshold is designed to remove outliers for which the differences are much larger than what would be expected given typical uncertainties in the dynamic emissivity and the atlas emissivity.
Figure2. MWHS/FY-3B channel 3 averaged number of data used in Northern Hemisphere boreal winter without using the dynamic emissivity. Time period: 0000 UTC 1 January 2015 to 0012 UTC 31 March 2015.
Figure3. Time-averaged channel 3 first-guess departure standard deviations of (a) MHS/NOAA-18 and (b) MWHS/FY-3B clear data over land where the orography is lower than 1500 m and the skin temperature is no larger than 278 K. Time period: 0000 UTC 1 January 2015 to 0012 UTC 31 March 2015.2
2.2. Quality control over snow-covered surfaces
Noting the high level of uncertainty in the emissivity estimates over snow, MWHS/FY-3B observations are initially excluded where the skin temperature is lower than 278 K. 278 K is taken as a proxy for snow, but clearly also rejects many snow-free scenes. An example of the winter data coverage is shown in Fig. 2, taken from the first three months of 2015. Subsequent analysis shows that this quality-control criterion is too conservative for some channels. Figure 3 shows the first-guess departure standard deviations of the MWHS/FY-3B channel 3, which is sensitive to upper tropospheric humidity, with peak sensitivity around 400 hPa. The channel 3 clear data (Chen et al., 2015) are comparable to those of the equivalent channel in MHS/NOAA-18 for scenes where the skin temperature is less than or equal to 278 K (i.e., the observations rejected by this quality-control check). No abnormally large standard deviations are found for snow-covered surfaces, while there are some large standard deviations over lower latitudes related to some scan position problems, which can be removed by the quality-control process of the assimilation (Chen et al., 2015). Therefore, for the MWHS/FY-3B channel 3, the skin temperature check is unnecessary and data can be assimilated in snow-covered regions.For channel 4, the weighting function peaks at lower altitude, close to 600 hPa for a standard US atmosphere at nadir view. However, due to the broad weighting functions the impacts of the surface on this channel are significant (Figs. 4c and d). The investigations suggest that, unlike channel 3, there is value to retaining a skin temperature check. However, the value can be set to a low value. The final choice is 255 K, based on examining the first-guess departure standard deviations of the MWHS/FY-3B channel 4 clear data (Fig. 4b). These are comparable to MHS (Fig. 4a) for the skin temperatures between 255 K and 278 K, except that the MWHS observations have higher noise due to higher NEDT (noise equivalent differential temperature) and some points with gross error that the general quality-control procedure in four-dimensional variational (4D-Var) data assimilation can easily remove (Chen et al., 2015). As for the regions where the skin temperatures are lower than 255 K, the first-guess departure standard deviations of both MHS/NOAA-18 and MWHS/FY-3B are quite large (Figs. 4c and d). The reason why it is worse for the skin temperatures lower than 255 K is not yet understood, but may reflect changes in snow morphology that make the atlas less representative, or larger skin temperature errors in these extreme conditions may lead to poorer emissivity retrievals and larger departures.
So, to summarize, there are three scenarios being tested, testing sensitivity both to the choice of the emissivity estimate and quality-control checks, and these are described in the next section.
Figure4. (a, b) Time-averaged first-guess departure standard deviations of (a) MHS/NOAA-18 and (b) MWHS/FY-3B channel 4 clear data over land where the orography is lower than 1000 m and the skin temperatures are between 255 K and 278 K. (c, d) Time-averaged first-guess departure standard deviations of (c) MHS/NOAA-18 and (d) MWHS/FY-3B channel 4 clear data over land where the skin temperatures are below 255 K. The emissivity is based on the emissivity atlas. Time period: 0000 UTC 1 January 2015 to 0012 UTC 31 March 2015.2
2.3. Experimental settings
Assimilation experiments are conducted using the ECMWF 12-h 4D-Var data assimilation system, with a model spatial resolution of around 40 km, a final incremental analysis resolution of about 80 km, and 91 levels in the vertical direction. Background error covariance is based on an ensemble of data assimilations, providing situation-dependent estimates of the uncertainty in the short-range forecast. The control experiment is run from 1 January 2015 to 31 March 2015, and 10-day forecasts are run at 0000 UTC and 1200 UTC each day, which provide 180 forecast samples in total. The control run assimilates the same observations used operationally by ECMWF on these dates, i.e., including MWHS/FY-3B data over ocean. Scan positions 12-81 are assimilated for channel 5, whereas the full scan is used for other channels (see Chen et al., 2015). To remove observations strongly affected by ice cloud and precipitation, a 5 K check on the absolute value of the first-guess departure of channel 1 is made (Chen et al., 2015). This definition of "clear sky" is used in this paper: it does not mean no clouds; rather, that the radiative impact of clouds can be considered small, and is not analyzed. Note that this criterion will also reject data for which the emissivity or skin temperature estimate is significantly in error.The thresholds of the orography are 1500 m, 1000 m and 800 m for channels 3-5, respectively, for both MHS/NOAA-18 and MWHS/FY-3B, in order to avoid observations that are too sensitive to the surface, i.e., where errors in the surface emissivity or skin temperatures play a large role. These thresholds reflect the different surface sensitivities of the sounding channels——tighter for the lowest channel (channel 5) and less tight for the higher channels. Also, the assigned observation errors over land are the same as those over ocean (Chen et al., 2015).
Based on the control run settings, the BasicAtlas experiment assimilates MWHS/FY-3B over land using the emissivity atlas without adding data over the snow-covered surfaces; in the SnowAtlas experiment, MWHS/FY-3B is assimilated over land using the emissivity atlas to add data over the snow-covered surfaces (i.e., using the relaxed skin temperature check); the SnowDynamic experiment is the same as the SnowAtlas run but uses the dynamic emissivity retrieved from 150 GHz (V). For the SnowDynamic run, the MWHS channel 1 observations cannot be used for both the clouds and precipitation screening and retrieving the emissivity. We therefore use channel 5 to identify clear observations, as it is the next available lowest-peaking channel with the strongest cloud sensitivity. The criterion for "clear data" remains at absolute first-guess departures below 5 K, and the MWHS channel 5 observations are not assimilated. In addition, the same series of experiments are repeated for the Northern Hemisphere summer from 1 July 2014 to 30 September 2014, to test seasonal differences.
3.1. Change in the number of used data
Compared to the control run, many additional observations are assimilated over land in the BasicAtlas experiment (Fig. 2), and even more channel 3 observations are added in the SnowAtlas experiment (Fig. 5a). The time-averaged first-guess departure standard deviations of MWHS/FY-3B channel 4 data used in the SnowAtlas experiment are reasonable over the snow-covered surfaces (Fig. 5b), validating that the updated quality control is a good choice. The number of data used are further increased in snow-covered areas in the SnowDynamic experiment (Fig. 6), which indicates that adopting the retrieved 150 GHz emissivity does allow more observations over the snow-covered surfaces being used.
Figure5. (a) Average increase in data use coverage of MWHS/FY-3B channel 3 over snow-covered surfaces in the SnowAtlas experiment relative to the BasicAtlas experiment. (b) Time-averaged first-guess departure standard deviations of MWHS/FY-3B channel 4 data used in the SnowAtlas experiment. Time period: 0000 UTC 1 January 2015 to 0012 UTC 31 March 2015.
Figure6. Difference in the average number of MWHS/FY-3B channel 3 data used between the SnowDynamic experiment and SnowAtlas experiment. Time period: 0000 UTC 1 January 2015 to 0012 UTC 31 March 2015.2
3.2. Analysis and forecast impact
3.2.1. Analysis impactWe now discuss the impact of adding MWHS/FY-3B data over land in the context of the full observing system used operationally at ECMWF at the time. This provides a stringent test of using the data in the "clear sky" case considering that the full system has many humidity observations over both sea and land.
Figure7. Normalized difference in the root-mean-squared vector wind error at 500 hPa (top) and 850 hPa (bottom) as a function of forecast range (days) over the Southern Hemisphere (left), tropics (middle) and Northern Hemisphere (right) of the BasicAtlas experiment (green line), SnowAtlas experiment (red line), and SnowDynamic experiment (black line). Negative normalized differences indicate an improvement in forecast quality. Vertical bars show the 95% confidence range. The period merges 1 July 2014 to 30 September 2014 and 1 January 2015 to 31 March 2015 together. Statistics are based on a total of 364 forecasts and verified against the ECMWF operational analysis.Assimilating MWHS/FY-3B observations over land does not show any negative impacts on the first-guess fit to other instruments in the ECMWF observing system. Departure statistics for ATMS are shown in Fig. 7. Results are combined for the winter and summer trials, providing 364 samples in total. Adding data over land (BasicAtlas experiment) shows positive impacts on the first-guess departures, especially over the Northern Hemisphere (green line). Compared to the BasicAtlas experiment, the experiments with added data over the snow-covered surfaces indicate more positive impacts on the first-guess departures, but these differences are not statistically significant. The impacts are more neutral for the Southern Hemisphere and the tropics as there is less land (not shown). Improvements are also found in the first-guess departures of the humidity channels of other instruments, such as HIRS (High-resolution Infrared Radiation Sounder) and SSMIS (Special Sensor Microwave Imager/Sounder) (not shown).
3.2.2. Forecast impacts
With both seasons combined, the SnowDynamic experiment shows significantly positive forecast impacts for the vector wind in the Southern Hemisphere out to day 3 or 4 of the forecast, a positive impact at day 2 in the Northern Hemisphere, and neutral results in the tropics (Fig. 8). The stronger positive forecast impact over the Southern Hemisphere is somewhat unexpected, given the smaller amount of data added in these regions. A possible reason might be that the Southern Hemisphere is less well constrained by other observations, particularly conventional ones, so even smaller amounts of data can have a notable impact. There is some indication of a negative impact at longer forecast ranges in the Northern Hemisphere at days 5-9, but this is not statistically significant. Neutral results are found in the BasicAtlas and SnowAtlas experiment. The scores are verified against the ECMWF operational analyses. Similar conclusions can be obtained not only for the vector wind, but also for the geopotential height, temperature and humidity (not shown). Figure 9 shows the vertical structure of the forecast impact, demonstrating again the statistically significant positive impact in the Southern Hemisphere but inconclusive results in the Northern Hemisphere. The two seasons are also examined individually, and the results in both periods are broadly consistent.
Figure8. Standard deviations of the first-guess departures of ATMS data used in the Northern Hemisphere, normalized by values for the control experiment. Horizontal bars indicate 95% confidence intervals. Green line for the BasicAtlas experiment; red line for the SnowAtlas experiment; black line for the SnowDynamic experiment. The period merges 1 July 2014 to 30 September 2014 and 1 January 2015 to 31 March 2015 together.
Figure9. Zonal means of the difference in forecast errors of the vector wind with time between the SnowDynamic experiment and the control run, normalized by the control run. The period merges 1 July 2014 to 30 September 2014 and 1 January 2015 to 31 March 2015 together. Statistics are based on a sample of 364 forecasts over the study period and verified against the ECMWF operational analysis. Cross-hatching indicates 95% confidence.The quality control used initially rejects most data for the snow-covered surfaces. A revised quality-control procedure that allows more data over the snow-covered surfaces to be assimilated is tested. Analysis of the first-guess departure statistics indicates that assimilating more MWHS/FY-3B data over land, particularly in snow-covered areas, again significantly improves the fit of short-range forecasts to other observations in the Northern Hemisphere——most notably the ATMS humidity channels. However, the forecast impact of this configuration is again found to be neutral.
Finally, the configuration using 150 GHz (V) dynamic emissivity is found to give broadly similar results to the emissivity atlas with less stringent quality control in terms of the first-guess fit to ATMS, but this configuration does show some positive forecast impacts, especially in the Southern Hemisphere.
To summarize, the forecast impact from MWHS over land is strongest when 150 GHz emissivity is used with the less stringent skin temperature quality-control option, to increase data coverage in snow-covered regions. However, for the same instrument, adding an 89 GHz channel would be useful (as is done for MWHS-2), so that the same methods can be used as for MHS. Having 150 GHz and 89 GHz observations together would be beneficial for the forecasts in both higher and lower latitudes.
