In a companion article (Paper I), a water-vapour retrieval algorithm has been developed using microwave sounding observations. An operational chain derived from this algorithm is being used for the Megha-Tropiques mission, launched in autumn 2011. The water vapour is retrieved for clear and cloudy scenes, excluding precipitation cases, over ocean and land surfaces. By-products are also calculated by the algorithm, including surface temperature and microwave emissivities over land.
The French–Indian mission Megha-Tropiques, launched on 12 October 2011, carries a passive microwave humidity sounder (Sondeur Atmosphérique du Profil d'Humidité Intertropicale par Radiométrie (SAPHIR)) and a microwave imager (Microwave Analysis and Detection of Rain and Atmospheric Systems (MADRAS)); see the companion article by Aires et al. (2012, hereafter Paper I) for more detailed instrument characteristics. SAPHIR is a passive microwave sounder with six channels around the 183.31 GHz absorbing line at ±0.2, ±1.1, ±2.8, ±4.2, ±6.8 and ±11 GHz. SAPHIR has a cross-track viewing geometry, with 130 pixels per scan line from nadir to ±42.96°. This means that the pixel size of 10 km at nadir increases with the scanning angle. MADRAS is a passive microwave imager that measures the radiation at nine frequency bands: 18.7, 23.8, 36.5, 89 and 157 GHz at vertical and horizontal polarizations, except for 23.8 Ghz which is vertical-only. It has a conical viewing geometry: the incidence angle of the measure has a constant angle of 53.5° (i.e. about 45° on-board angle). Its scan coverage is ±65° and the swath is about 1700 km wide. In the companion article (Paper I), it has been shown that the sounding capacities of SAPHIR benefit greatly from the surface and integrated information provided by MADRAS, so a combined SAPHIR/MADRAS chain has been developed, mainly for the retrieval of atmospheric humidity.
Various retrieval algorithms have been developed for the retrieval of atmospheric water vapour from microwave instruments. Special Sensor Microwave Instrument (SSM/I) observations have been exploited for example in Alishouse et al. (1990), Tjemkes et al. (1991) and Jackson and Stephens (1995), mostly over ocean. In Deeter (2007), an algorithm is built for ocean and land surfaces. Grody et al. (2001) and Jethva and Srinivasan (2004) use Advanced Microwave Sounding Unit (AMSU) observations over ocean to retrieve water vapour. Tropical Rainfall Measuring Mission (TRMM) Microwave Imager (TMI) observations are used in the work of Wang et al. (2009) and Gentemann et al. (2004) over the tropical belt, again mostly over ocean surfaces.
The processing chain developed in the companion article (Paper I) for Megha-Tropiques operates over the ocean and land, for clear or cloudy/non-precipitating situations. It is based on a neural network (NN) inversion, trained on a large collection of radiative transfer simulations using the Radiative Transfer for TIROS Operational Vertical Sounder (RTTOV) code (Saunder et al., 1999) and European Centre for Medium-Range Weather Forecasts (ECMWF) analyses. It uses a priori information on the land-surface emissivities to help solve the inversion problem over land (Aires et al., 2011a). It also relies on an innovative methodology to calibrate the observations prior to the inversion (Aires et al., 2010). A separate NN is used for four different configurations: land/ocean and clear/cloudy situations. A precipitation and cloud flag is required. The ECMWF temperature profile is also used as a priori information.
The algorithm and its theoretical assessment are presented in the companion article (Paper I), but since no real observations are available yet from SAPHIR/MADRAS, the chain is tested on two existing platforms: AQUA and MetOp. In this article, a near real-time retrieval chain is tested on both of these platforms. The passive microwave measurements from the AQUA platform are composed of collocated observations from two radiometers: the Humidity Sounder for Brazil (HSB, only operated from May 2002–February 2003) and the Advanced Microwave Scanning Radiometer–Earth observing system (AMSR-E). The HSB cross-track sounder is a nearly identical copy of AMSU-B with four moisture-sounding channels. Three out of four are located around the strong water-vapour absorption line at 183.31 GHz (±1.0, ±3.0 and ±7.0) and the fourth channel is a window channel at 150 GHz. During one scan, HSB samples 90 scenes of 1.1° between ±49.5°, with a footprint size of 13.5 km at nadir. AMSR-E is a dual-polarized radiometer operating at frequencies of 6.9, 10.7, 18.7, 23.8, 36.5 and 89 GHz. The instrument has a conically scanning antenna that provides multichannel observations at a constant incidence angle of 55° across a 1445 km swath. The spatial resolution of AMSR-E varies from approximately 60 km at 6.9 GHz to 5 km at 89 GHz.
MetOp carries two passive microwave sounders, the Microwave Humidity Sounder (MHS) and AMSU-A (1 and 2) for temperature sounding. MHS provides measurements in the 183.31 GHz water-vapour absorption line, at ±1, ±3 and ±7 GHz, plus at two window channels at 89 and 150 GHz that enable deeper penetration through the atmosphere down to the Earth's surface. Each swath is made up of 90 contiguous individual pixels that have a diameter of approximately 16 km at nadir. AMSU-A is dedicated to the retrieval of atmospheric temperature profiles, with 12 sounding channels between the 50 and 60 GHz O2 band and three other channels at 23.8, 31.4 and 89 GHz. It is a cross-track scanning radiometer, with ±48.3° from nadir and a total of 30 Earth fields of view of 3.3° per scan line, providing a nominal spatial resolution of 48 km at nadir, and its swath is approximately 2000 km.
The water-vapour (WV) estimates are evaluated in section 2. The retrievals of surface temperature and microwave emissivities over continents are examined in section 3. A validation in the space of the satellite observations is developed in section 4. Finally, conclusions are drawn in section 5.
2. Evaluation of the water-vapour retrievals
In this section, the evaluation of the WV retrieved by the inversion chain (see the companion article) is presented, in comparison with the ECMWF analyses and radiosonde measurements.
It should be clear, first, that the ECMWF analyses are already a good estimation of the state of the atmosphere, taking into account a lot of information from satellite data (including AQUA and MetOp observations) and from radiosoundings. The analyses can be used as a reference, but the goal of the satellite retrieval is to improve the analysis, in the sense that the retrieved profile, when given as input to the radiative transfer (RT) model, should provide a simulation closer to the satellite observations than the simulation using the analysis. Any departure from the analysis can be treated as either error or improvement in the retrieval. These considerations bring the necessity to validate the retrievals in the space of the satellite observations, and this will be developed further in section 4.
The evaluation is performed in the framework of the Megha-Tropiques mission. Although the algorithms are globally applicable, they are specifically tested here over the Tropics (±30° in latitude), the region that will be observed by Megha-Tropiques. Furthermore, since it has been shown in the companion article that the AQUA retrieval uncertainties are closer theoretically to Megha-Tropiques than MetOp, the results will be presented mostly using the AQUA retrievals. The statistics commented on below are processed on a systematic sampling selection from the two months of satellite observations described in section 3.4 of the companion article.
2.1. Total column water vapour
The retrieval of total column water vapour (TCWV) can be performed in two different ways. First, a dedicated NN can be trained to retrieve the TCWV directly. Secondly, it is possible to integrate the TCWV vertically from the NN retrieved WV profile. The second scheme appears to be more satisfactory in our tests (not shown). This might be surprising, as the direct retrieval might be thought to be simpler. A possible explanation could be that the profile retrieval uses more neurons in the inversion process and therefore that the inversion scheme is able to obtain more information from the observations.
Figure 1 represents a map of the TCWV retrieved from the AQUA observations, together with the corresponding ECMWF analysis, for the morning of 7 September 2002 from 0000–1140 UTC over the Tropics. The AMSR-E tracks are collocated with the HSB observations. This map (and the following maps) shows the retrieval results for the descending and ascending orbits: the first ones are localized on the west of the map and are observed during local night-time, the eastern part of the map corresponding to ascending orbits during local daytime. The missing pixels correspond to unprocessed data (precipitating scenes or high elevation, as specified in the companion article).
The spatial structures of the retrieved TCWV are very similar to the analysis (Figure 1). This is very encouraging considering that no a priori information on the WV is used in the retrieval scheme. Some differences can be noted however, for instance in the Indian Ocean south of India or west of Brazil. There is a good continuity in the field between ocean and land surfaces (e.g. west of Africa or Central America), confirming that the land and ocean inversion schemes are robust and coherent with one another: this is a key point, as so far the inversion of satellite WV information has been limited over land. The retrievals over continents are possible thanks to the surface constraints on the microwave emissivities. The retrieved fields show slightly more spatial noise, but this can be expected as the analyses are smooth by nature.
The TCWVs from the analyses are compared directly to the retrievals in the scatterplots of Figure 2 (left), over both ocean and land. The bias is limited (0.39 kg m−2), for a standard deviation of the errors equal to 3.59 kg m−2. The retrieval slightly underestimates the very high TCWV, especially over land (with a slope of the regression slightly lower than unity). This is a general behaviour of statistical retrieval schemes: they tend to dampen the dynamics of the retrieved geophysical variables. Note that this reduction of the dynamics is limited, partly because three neural network inversion models have been trained on three TCWV ranges, as given by the a priori ECMWF analysis (see the companion article).
In order to check the quality of the retrieval further, September 2002 and January 2003 radiosondes are extracted from the ECMWF operational sounding archive used during the assimilation process in the ERA40 / ERA-Interim re-analyses (Uppala et al., 2005). The temperature and humidity measurements have been quality-controlled in order to discard incomplete profiles (thresholds of 30 hPa for the temperature and 350 hPa for the humidity), and a vertical extrapolation is applied up to the top of the atmosphere using a climatology (R. Armante, 2008, personal communication, Laboratoire de Meteorologie Dynamique (LMD)). The radiosoundings must be ±0.3° apart at maximum from the closest AMSR-E observation, with a maximum time difference of ±90 min. AMSR-E scenes with scattering signatures have been discarded. 72 match-ups fitting the coincidence criteria above have been found; they are almost equally distributed in cloud-free and cloudy situations, for night and day observations.
The radiosonde measurements only come from continental areas so the two right panels in Figure 2 are for land situations only. For each panel, the standard deviation of the errors is provided together with the bias and the R2 statistics.
These results show that the AQUA retrieval is as good as the analysis compared with the radiosoundings: both have standard deviation (StD) differences as well as a R2 coefficient in the same range (4.1 and 4.6 kg m−2 for the StD and 85–89% for the R2). The retrieval is less biased than the analysis, which is surprising considering that the radiosonde observations are assimilated in the analysis. The StD of the difference from the radiosondes is comparable with the departure of the retrieval from the analysis, reinforcing the significance of the comparisons with the analysis. A negative bias of 2.1 kg m−2 is observed on the retrieval of night and cloud-free situations, while other situations have a bias lower than 0.5 kg m−2. Root-mean-square (RMS) errors show no strong link with the time of observation or the cloud cover. Currently, TCWV is estimated accurately only over ocean regions, for non-precipitating conditions. For comparison purposes, Deblonde and Wagneur (1997) obtained a bias of −0.2 kg m−2 and a StD of 3.7 kg m−2 between SSM/I TCWV estimates and radiosonde observations from small islands. Retrieval results agree better with ECMWF analysis over ocean than over land. Over ocean, our results give a bias of 0.3 kg m−2 and a RMS of 2.8 kg m−2.
A previous comparison under similar conditions between SSM/I retrieval and a numerical weather prediction (NWP) analysis showed larger differences, with a bias of 1.08 kg m−2 and a RMS of 3.25 kg m−2 (Deblonde et al., 1997). Aires et al. (2001) have developed a similar retrieval scheme using SSM/I observations and obtained 3.8 kg m−2 for clear cases and 4.9 kg m−2 for cloudy cases, both over land. In Wang et al. (2009) a semi-statistical retrieval scheme is developed for TMI observations. The RMS error compared with that for radiosondes is equal to 3.5 kg m−2 (2.5 when the 5% of worst cases are suppressed from the statistics), but this scheme is performed for oceanic cases and only for the tropical belt.
2.2. Atmospheric water-vapour profiles
The retrieval of WV profiles with instruments such as AMSU-B, MHS, HSB or SAPHIR is a true challenge. Although their weighting functions (see figure 1 in the companion article, Paper I) sample the vertical column relatively well, they overlap significantly, meaning that the degrees of freedom or the number of independent pieces of information are rather limited (about 3–4). When these satellite observations are assimilated in a NWP scheme, the a priori information from the atmospheric circulation model helps to constrain the vertical structure of the retrieval. It is much more difficult to estimate the vertical structure in a direct retrieval, without a priori vertical information.
Similarly to Figure 1, Figure 3 presents the relative humidity between 920 and 1013 hPa from the analysis and from the retrieval. Note that this layer is particularly difficult to retrieve, especially over land, due to the surface contribution. Comparison of the maps reveals that the main structures of the WV are captured by the retrieval and the transition from ocean to land is smooth, even for this lower layer. Even mesoscale structures in the WV layer over land are correctly estimated, for instance in Sudan, Africa, or the east of Brazil. This is particularly encouraging because the use of microwave observations over continents is particularly difficult, especially for the surface-sensitive channels in the lower troposphere (Aires et al., 2011a). This is to be expected since the surface emissivities have been well constrained in the algorithm. However, some noticeable differences can be observed, even over ocean, for instance in the South Pacific and Indian Oceans, where WV structures present in the analysis are not reproduced by the retrieval.
Figure 4 represents the RMS profile of the difference of the retrieval and the ECMWF analysis. These ‘departures' are not errors because it cannot be said a priori which one is closest to reality. This statement is part of the challenge in such an evaluation. Retrieval seems closer to the analysis over ocean than over land. This can be related to the fact that the analysis assimilates more observations from microwave sensors over ocean than over land (in both clear and cloudy situations), providing better agreement. The departures seem overall to be close to the theoretical uncertainty estimates (i.e. between 5% and 10% of the RMS error in percentage of relative humidity, see figure 4 of the companion article). For oceanic cases, the theoretical estimates are higher than the departures. Furthermore, the departures have a flat vertical error structure where theoretical errors are stronger in the lower troposphere and smaller in higher atmospheric layers.
For both estimates, the errors are higher for the lower layers over land but the theoretical errors are larger (about 12%) than the departures (less than 10%). Over land, microwave observations cannot be easily assimilated by the analysis for the lower layers of the atmosphere, due to the important surface contribution. It leads to a less accurate description of the relative humidity (RH) profile. The situation is worst in cloudy cases, where there is no contribution of information from real observations to the forecasts. In this configuration, the analysis is not expected to fit reality exactly and thus should not be used as a target for our retrieval.
One should also understand that these statistics depend upon the space–time thresholds used in the coincidences. Plus or minus 1.5 h are tolerated in these statistics in order to keep enough data samples. If 10 min were used instead, the RMS profile could be one or two percentage points lower. The RMS profile is represented only up to 250 hPa because the information provided by the satellites beyond this value is low (see weighting functions in figure 1 of the companion article). As expected, the retrieval of the lower layer is better over ocean than over land. Overall, MetOp and AQUA retrievals have similar statistics. Generally, the AQUA platform performs better close to the surface, due to the larger number of window channels provided by the AMSR-E imager. The opposite case usually prevails for the higher layers, likely due to the lower instrument noise of MHS on board MetOp compared with HSB on board AQUA. The overall statistics are satisfactory, with a departure from analysis that is lower than 10%. The retrieval statistics are stable for angles lower than 49° (the statistics are integrated over all angles up to 49° and higher angles are excluded).
Another diagnostic is to compare the variability of the resulting WV profiles. Figure 5 represents the statistical variability of the WV profile from the ECMWF analysis and from our AQUA and MetOp retrievals. Over 90% of the variability of the analysis is retrieved by the NN inversion for the two sets of microwave instruments. The variability of the vertical structure is well reproduced, with only a slightly lower variability of the retrieval compared with the analysis. For the lower atmospheric layers, an increase of the variability is even observed in the retrieval, possibly due to the good treatment of the surface contribution in our algorithm which can improve the analysis, especially over the continents. Another reason is that the profiles from the retrieval have a thinner horizontal resolution than the ECMWF analysis, which has smoother 1.125° data.
In Figure 6, the retrieved WV profiles are also compared with the radiosoundings, for the same in situ measurements as in Figure 2. The corresponding ECMWF analyses are added to the comparison. For the lower layers below 560 hPa, the analyses are systematically closer to the radiosoundings than the retrieval, although the difference between the retrieval and the in situ measurements is limited, especially close to the surface. With the radiosoundings already assimilated in the analyses, it is expected that they are very close. Nevertheless, this illustrates the difficulty of comparing these different kinds of measurements, radiosoundings providing a very local view compared with the 20×20 km2 microwave sensor pixels. WV profiles over land are subject to strong spatial gradients in the low troposphere. Note, however, that the comparison between the retrievals and the radiosoundings shows differences within 15% in percentage of relative humidity, except for the lower layer (Jethva and Srinivasan, 2004). A previous study using radiosoundings from the Radiation, Cloud and Climate Interactions in the Amazonia/Large Scale Biosphere–Atmosphere Experiment (RaCCI/LBA) field campaign and HSB retrieved water vapour showed results approaching this for clear-sky situations and more important differences in cloudy cases (Lima et al., 2006).
The difference statistics between the AQUA/MetOp retrieval and the analysis and radiosoundings can be compared with the previous theoretical uncertainty estimations of the Megha-Tropiques and AQUA/MetOp platforms of the companion article (Paper I). They can be compared with the theoretical learning errors (figure 5 in the companion article) or the information content analysis estimate (figure 6 in the companion article). The information content analysis (around 10% for the high troposphere and 20% for the close-to-surface layers) is closest to the statistics using radiosondes. However, from this theoretical and practical analysis, it can be expected that the uncertainties in the WV profile retrieval from Megha-Tropiques over both ocean and land will be below 20% for the lower atmospheric layers and below 15% in higher layers in percentage of the relative humidity, with an uncertainty of ∼5 kg m−2 in the TCWV.
3. Evaluation of the surface temperature and emissivity retrievals over the continents
The main objective of this study is to validate the WV products; however, a preliminary evaluation of the auxiliary products is initiated in this section. The validation results are only presented here by comparison with the first-guess, namely the ECMWF analysis for the surface temperature and the atlases for the land-surface emissivities.
3.1. Land-surface temperature
The land skin surface temperature (TS) is not only a very important geophysical variable driving the weather and climate (in land/atmosphere interaction for instance) but also plays a major role in the inversion of the surface-sensitive sounding channels (English et al., 2008) along with the land-surface emissivities. TS is routinely estimated from satellites using infrared measurements. However, to overcome the limitations of infrared observations under cloudy conditions, methods have been developed using microwave measurements (e.g. Aires et al., 2001; Holmes et al., 2009). The present study contributes to this effort.
Results are presented for the AQUA platform, which carries instruments very similar to Megha-Tropiques. MetOp does not have a microwave imager with a large number of window channels and, as a consequence, the surface parameter estimates from this platform are not as accurate (see the companion article). Figure 7 shows a scatterplot of retrieved TS versus the TS from the ECMWF analysis. The comparison of surface skin temperature is very challenging and the time agreement is of course an important factor. The ECMWF analysis is available only every six hours. The coincidence with satellite observations is accepted in our collocation process only if the time difference is less than 90 min. Furthermore, a temporal interpolation is used to obtain a better match. In agreement with the theoretical results (see the companion article), the standard deviation of the difference between retrieval and analysis is significantly lower than the uncertainty in the first-guess error and the bias is null. Note nevertheless that the retrieval presents a negative bias of 1.3 K during the day under clear-sky conditions, more than twice the bias during the night or under cloudy conditions. With an Equatorial Crossing Time of AQUA around 1:30PM (ascending mode), the large temporal variability of TS at that time of the day inevitably induces large errors under clear conditions, with a temporal difference between the analysis and observations of up to 1.5 h and with the diurnal cycle of the TS ECMWF analysis still questionable.
Figure 8 represents the skin surface temperature for 7 September 2002 (from 0000–1140 UTC) from AQUA (top) and its departure from the first guess derived from the ECMWF analysis. The spatial structures are accurately represented, showing the expected gradient around topography features or in the tropical forests. The cloud cover is also shown (bottom panel). Transitions from clear to cloudy conditions do not introduce discontinuities in the TS retrieval compared with the analysis. This is very encouraging, with the objective of using the microwave retrieval to complement the infrared estimates in cloudy regions. The ascending orbits around noon local time, on the eastern part of the map, show more differences than the descending orbits during local night. As already discussed, differences are higher during the day than during the night, especially in areas where the expected diurnal cycle is large (e.g. in arid regions).
3.2. Land-surface microwave emissivities
The analysis of the retrieved surface emissivity is limited to 23.8 GHz for both orthogonal polarizations. Similar conclusions are derived for the other channels of the microwave sensors used in this study. The WV channel around 23 GHz has been selected, as it is common to all the satellites in this study. Figure 7 presents the surface emissivity retrieved from AQUA as a function of its first guess calculated from the Tool to Estimate Land Surface Emissivities at Microwaves (TELSEM) and the pre-calculated emissivity atlases (Aires et al., 2011a). The results are not biased and the standard deviation is even lower than the expected theoretical departure. We checked that the results do not depend upon the time of the day, which illustrates the robustness of the retrieval with respect to TS. Neither do they depend upon the cloud cover.
Figure 9 shows maps of the surface emissivity at 23.8 GHz vertical polarization retrieved from AQUA for 7 September 2002 (top), along with its departure from the emissivity first guess (bottom). The emissivities have the expected spatial structures (e.g. low emissivities for the hydrological structures, low emissivities over carbonated outcrops in Oman or high emissivities over the deserts). The differences from the first guess do not exhibit any obvious contamination due to the presence of clouds (see Figure 8, bottom) or any errors that would be related to specific surface changes.
4. TB-space validation
In a physically based retrieval scheme, the quality criterion of the inversion process is based on the difference between (1) the actual observations and (2) the RT simulations when using the retrieved atmospheric and surface parameters. The objective of the inversion is to minimize the difference between the observations and these simulations. As a consequence, the improvement of the inversion compared with a first guess or an a priori can be measured by the comparison of the differences between the observations and the RT simulations when using (1) the first guess or a priori information on the state of the atmosphere and (2) the retrieved parameters. If the retrieved variables degrade these statistics, the inversion has not improved the a priori information or has mathematically diverged and failed to perform the retrieval. The accuracy in the final retrieval product is largely dependent on the accuracy of RT calculations. The objective of the inversion is to perform retrieval accurately based on the RT simulations, which are supposed to represent the real world.
Note that the differences between the observations and the simulations are not only related to the quality of the retrieval. The satellite observations include instrument noise (depending on the sensors) and the RT also induces errors. The objective of the calibration step prior to the inversion process is to limit these two sources of error, but the calibration is never perfect. In addition, differences in space and time between observations and analysis can also be critical.
4.1. Brightness-temperature comparisons
The geophysical variables retrieved by the inversion scheme described in section 4.3 of the companion article (Paper I) are used as inputs for the RT model (i.e. RTTOV). These include the WV profile, the surface temperature and emissivities over land and the wind speed at the surface over ocean. The simulations are also performed using the initial ECMWF analysis in order to compare the potential relative improvements provided by the retrievals.
Figure 10 represents the statistics for the MetOp instruments (AMSU-A and MHS, left) and for the AQUA instruments (AMSR-E and HSB, right). Both land (upper part) and ocean (bottom part) are considered. RMS statistics are provided for clear (black) and cloudy (grey) situations. The statistics are presented for RT simulations performed on the retrievals (continuous line) and on the analysis (dashed line). The differences between the simulated TB result uniquely from the parameters that the inversion scheme retrieves. The simulations with the retrieved products (continuous lines) always have a lower RMS than the simulations with the a priori information (dashed ones): this clearly indicates that the retrieved variables represent an improvement in the TB space, compared with the analysis. When the retrievals are used as inputs to the RT code instead of the ECMWF analysis, the simulated TB becomes closer to the calibrated observations. This is exactly the objective of the retrieval scheme. In particular, the simulations of the 183 GHz channels are improved by the retrieved WV profile for the four configurations, for AQUA and MetOp instruments. The statistics for the WV sounding channels are similar for all cases (land/ocean, clear/cloudy, AQUA/MetOp), meaning that the retrieval is very robust regardless of the conditions. The window channels are improved by better surface temperatures and emissivities over land, and surface wind over ocean. Over continental surfaces, better statistics are obtained under cloudy conditions for the surface-sensitive channels: this is due to the smaller dynamics of the surface temperature in cloudy cases. Note that part of the improvement can be related to a better space/time coincidence between the retrieval and the satellite data (by construction, the retrieval is performed at the same time and location as the satellite observation).
Figure 11 presents the calibrated observations at 23.8 GHz for vertical polarization from HSB on AQUA (middle), along with the simulations from ECMWF analysis (top) and the retrieval (bottom). This channel is very sensitive to the TCWV, and using the retrieval improves the comparison with observations. For instance, the observed structures in the Pacific Ocean around 150°E and 20°N or around 140°W and 20°S are better reproduced using the retrieval in the simulation. Over ocean, simulation from analysis seems to present a strong positive bias in very humid atmospheres (up to 5 K from the observations) that is not so well reproduced by the retrieval (at 10° of latitude). This figure can be related to Figure 1, where retrieval values were way below analysis forecasts over these areas. The possibility of a positive bias from our radiative transfer model is not enough to explain this phenomenon, since the RTTOV model is largely used by NWP centres for forecasts. Furthermore, a bias between observed and simulated TB would have been reduced by our calibration step.
Figure 12 presents the calibrated observations at 183.81±3 GHz from HSB on AQUA and the difference between these calibrated observations and the simulation using the retrievals. This channel is strongly correlated to the water vapour in the low troposphere (see figure 2 in the companion article). Simulations show significant differences from observation when using the analysis, with often ≈3 K differences, over both ocean and land. A time mismatch alone cannot explain these large discrepancies, since the time difference is less than 30 min over some areas such as the South Atlantic or Africa. When using the retrieved products in the simulations, the agreement with observations is significantly improved over both ocean and land, without noticeable discontinuity between ocean and land.
4.2. Contribution of each retrieved variable
In order to assess the impact of each retrieved variable in the improvement in TB space, a similar experiment is conducted in which only some of the retrieved variables are used as inputs to the RT code. Similarly to Figure 10, Figure 13 shows RMS statistics but for AQUA observations (AMSR-E and HSB), for different RT inputs: (1) the ECMWF initial analysis (in blue), (2) all the retrieved variables (black), (3) only the retrieved atmospheric humidity profile (red) and (4) only the retrieved surface parameters (green). For the 183 GHz channels and for both platforms, as expected, the retrieval of WV is the primary contributor to the reduction in the difference from the observations, the impact of the surface to the signal being very small or null. Even for the window channels, for oceanic situations the surface retrieval has a limited impact. Firstly, the quality of the oceanic surface parameters in the analysis is good and the departure from observations when using the analysis is already small. Secondly, the surface contribution to the radiation is reduced compared with the land case, due to the lower emissivities of the ocean. Over land, the retrieval of TS and emissivity yields to a significant improvement for all channels up to 100 GHz (with the residual RMS higher for clear cases, likely due to the larger variability of TS under clear conditions, as already discussed).
4.3. Evaluation of the improvement in TB space
Various diagnostics can be designed to monitor the quality of the retrieval automatically: these criteria can help to filter out retrievals that have not satisfactorily converged, as well as to assess the degree of confidence in the retrieval.
A measure of the improvement provided by the retrieval with respect to the analysis is given by
where RT(ana) is the RT simulation for the analysis, RT(ret) for the retrieval and CAL(obs) the calibrated observations. When this quantity is positive, the retrieval has reduced the difference from observations compared with the analysis.
The histograms of this quantity at 150 and 183.31±3 GHz from HSB (Figure 14) are biased towards the left, indicating that the retrieval improves the RT simulation compared with the calibrated observations in more than 50% of cases. This is true for the four configurations (over land or ocean and for clear or cloudy scenes), with an improvement of up to 10 K. For 183.31±3 GHz, the distributions are identical for the four configurations, which means that the type of improvement our retrieval brings is similar for cloudy/clear cases and for land/ocean surfaces. In contrast, the distributions vary for the 150 GHz channel: they are more symmetrical and slightly biased to the left for the land cases and more skewed to the left with longer tails for the ocean cases. This means that improvements can be higher for the oceanic case (they can reach 10 K) and less important for the land case (i.e. up to 5 K). It is important to note that in our convention the negative cases are improvements. However, positive cases do not necessarily have to be a degradation of the analysis: traditionally in operational centres this kind of diagnostics is used to test whether the retrieval improves the analysis and in the case in which it degrades it the retrieval is simply discarded (only a small percentage of satellite data are actually used in the assimilation system).
The previous figure provides information on the amplitude of the potential improvements of our retrievals. Figure 15 presents the percentage of situations that are improved when the retrieval is used instead of the analysis, for all AMSR-E/HSB channels on board AQUA. Regardless of the situation (land/ocean and clear/cloudy) or the channel, more than 50% of cases are improved. The percentage of improvements is the highest for window channels over land under cloudy conditions. Note that the percentages of improved situations are of the same order for land and ocean cases, confirming that our retrieval, in most cases, provides similar performance over land and ocean. The higher improvement rates in the 183 ± 3 and ±7 GHz channels illustrate the ability of our algorithm to retrieve the lower part of the troposphere humidity, regardless of the cloud cover. This is key information that cannot be obtained from infrared sensors.
The statistics that have been presented showed that the retrievals improve the analysis in the majority of cases, with an impact that is rather significant in TB space (a few K). This proves that an increased amount of microwave observations could potentially be assimilated or used in the retrieval process in order to improve weather forecasts, in particular over the continents.
A retrieval algorithm has been developed to retrieve water-vapour profiles over both ocean and land, from microwave instruments such as SAPHIR/MADRAS on board Megha-Tropiques, the French-Indian mission to be launched in autumn 2011. Its performance is tested on existing observations with similar characteristics from the AQUA and MetOp platforms. The WV profile and total column content are compared with the ECMWF analysis, as well as radiosounding measurements. It has been shown that it is important for Megha-Tropiques that SAPHIR and MADRAS instruments are used together for the retrieval of the water vapour. The estimated errors in the TCWV content are comparable over land and ocean and under clear and cloudy cases, with a RMS error of ∼5 kg m−2 with respect to the radiosondes. The water-vapour profiles are retrieved for six atmospheric layers with a maximum RMS error of 20% in relative humidity in the lower layer over land, without any a priori use of water-vapour information. The retrieved water-vapour fields do not show any obvious discontinuity related to ocean/land or to cloudy/clear transitions. The retrieval of the surface temperature and emissivities over land has also been assessed. The microwave retrievals are a very promising complement to the infrared estimates, especially when sounding and window channels are used together. Microwave observations are particularly interesting for sounding in the presence of clouds.
The retrieved products have also been evaluated in terms of simulated brightness temperatures. After the inversion process, the retrieved parameters are used as inputs to the RT code and the simulations are compared with observations. These differences are a direct test of the performance of the retrieval and can be used to filter out unsuccessful retrievals. The statistics show that the retrievals are an improvement compared with the analysis in the majority of cases and that the improvement in TB space is significant (a few K). This improvement in TB space is the result of an improvement of the geophysical variables (e.g. humidity profile). However, validation using insitu measurements is also necessary for a direct validation of the humidity retrievals.
In the framework of preparation for Megha-Tropiques, the algorithm has been tested over the Tropics, although the retrieval scheme is global. Its implementation on a global scale should not be a problem, except over snow and ice conditions where the high spatial and temporal variability of the surface emissivities makes characterization of the lower atmospheric layer more difficult. Another perspective for this work is the combined use of microwave and infrared observations for the water-vapour retrieval scheme. A good synergy exists between these two types of observations (Aires etal., 2012) and the neural network approach is a good candidate to exploit it.
We thank Rémy Roca for discussions linked to the Megha-Tropiques mission. We are grateful to the ARA group at LMD for providing their radiosonde database. Didier Renault, Anne Liefermann and more generally the CNES (Centre National des Études Spatiales) are acknowledged for their support for part of this study, in the framework of the development of the Megha-Tropiques algorithms.