Journal of Geophysical Research: Atmospheres

Evaluation and improvement of AMSU precipitation retrievals

Authors


Abstract

[1] The Advanced Microwave Sounding Unit-B (AMSU-B) high frequency channels (89 and 150-GHz) provide the ability to detect the scattering associated with precipitation sized ice particles and, indirectly with precipitation rate. Despite the fact that AMSU (a “sounder”) wasn't designed for rainfall retrieval, many studies show that it is possible to use this information for precipitation retrieval and therefore, many researchers use them in blending techniques to estimate rainfall at global scale. The main advantage of this approach is the availability of three NOAA POES satellites spaced approximately 4 h apart with a spatial resolution of 16 km at nadir and a wider swath than SSMI (2200 km), offers an excellent opportunity to reduce the errors associated with the inadequate sample of the diurnal cycle; while the weakness of this technique are related with its present inability to retrieve rain that has little or no ice; and with the cross-scan characteristics of the instrument (different footprints for different local zenithal angles). This situation tends to generate, in the first case, much less rain coverage over ocean than other algorithms based on emission techniques and a shift in the frequency distribution for low zenithal angles in the second case. An improvement to the current scattering based algorithm that uses cloud liquid water content (over ocean only) and the local zenithal angle will be developed and tested in this paper. In particular, the application of the proposed correction scheme produces a more realistic rainfall amount distribution function (especially over the tropics) and a light increase of rainfall occurrence in midlatitudes. The improvement is demonstrated by both global and regional studies over land and ocean.

1. Introduction

[2] Since the first launch of the Advanced Microwave Sounding Unit (AMSU) on board the NOAA 15 satellite in July 1998, many environmental passive microwave (PMW) products have been operationally generated by NOAA. New products such as Ice Water Path (IWP) and Ice Particle Size are also derived, owing to the unique AMSU-B millimeter wavelength channels [Weng et al., 2003]. The ability of AMSU-B to derive cloud and precipitation products using this information is remarkable for a variety of applications. These products, combined with those derived from the DMSP Special Sensor Microwave Imager (SSM/I), the Aqua Advanced Microwave Scanning Microwave Radiometer (AMSR-E) and the passive microwave (TMI) sensor on board the Tropical Rainfall Measuring Mission (TRMM) offers the scientific community an excellent source of global hydrological products. The high temporal frequency of rainfall retrievals will also help to better understand the diurnal cycle for different rainfall regimes around the world. It is also important to note that the high frequency channels on board AMSU-B offers the ability to detect lighter rain rates over land [Qiu et al., 2005] and even snowfall [Kongoli et al., 2004]. Additionally, there are several current and near term sensors with similar capability to AMSU-B, including the Microwave Humidity Sounder (MHS), carried on NOAA 18 and MetOp-A, while the National Polar-Orbiting Environmental Satellite System (NPOESS) Preparatory Program (NPP), will have the first Advanced Technology Microwave Sounder (ATMS). The ATMS has the same frequencies as the AMSU, with a broader scan swath to provide complete coverage of the globe twice a day. Another source of high frequency channel is the Special Sensor Microwave Imager Sounder (SSMIS) flying on DMSP F-16 and F-17. This is a conical scanner with 24 channels from 19 to 183 GHz and a swath width of 1700 km. Nevertheless, several researchers found that rainfall derived from the AMSU-B algorithm differs in many respects from SSM/I and other PMW estimation techniques. While SSM/I is equipped with channels that detect both emission and scattering signatures, the AMSU-B sensor only has high-frequency channels; thus, only precipitation that is detectable from a scattering signature can be estimated [Joyce et al., 2004]. Although the AMSU-A sensor has some limited ability to infer emission based signals over the ocean (i.e., through the use of 31 GHz channel), it has not been incorporated within the AMSU-based precipitation algorithm due to its much coarser spatial resolution (e.g., 48 km at nadir) and its saturation at low rain rates. However, there is potential to improve the current AMSU-B algorithm over ocean in light rain situations using such information.

[3] It is the purpose of this paper to describe an AMSU-B rain-rate correction scheme and to demonstrate its utility comparing the obtained results with other PMW rainfall retrievals. Section 2 will describe the database used in this study and a full explanation of the correction scheme; Section 3 presents several application examples and comparisons with other algorithms while Section 4 contains a summary of the paper.

2. Data and AMSU Rain-rate Correction Scheme

[4] The Advanced Microwave Sounding Unit-B (AMSU-B) is a 5 channel microwave radiometer. AMSU-B covers channels 16 through 20. The highest channels: 18, 19 and 20, span the strongly opaque water vapor absorption line at 183 GHz and provide data on the atmosphere's humidity level. Channels 16 and 17, at 89 GHz and 150 GHz, respectively, enable deeper penetration through the atmosphere to the Earth's surface. This is a cross-track, line scanned instrument designed to measure scene radiances in the 5 mentioned channels. At each channel frequency, the antenna beam width is a constant 1.1 degrees (at the half power point). Ninety contiguous scene resolution cells are sampled in a continuous fashion (beam positions from 1 to 90, where position 45 is in the nadir of a given line), each scan covering 50 degrees on each side of the subsatellite path. These scan patterns and geometric resolution translate to a 16.3 km diameter cell at nadir at a nominal altitude of 850 km, while this diameter is around 36 km in the limb of the image [Goodrum et al., 2000].

[5] The data used in this paper is the NOAA AMSU-B HMAP data set. This data set is generated operationally by NOAA/NESDIS in the Microwave Surface and Precipitation Products System (MSPPS) and it is available at the Cooperative Institute for Climate Studies (CICS) server (http://cics.umd.edu/∼lcao/datasets.html). This Level-3 daily mapped orbital product is available in HDF-EOS Grid format where the AMSU-A and AMSU-B swath products are remapped into a 0.25 degree global grid. Because this remapping procedure is based on the actual field-of-view (FOV), at the edge of the swath, an FOV can occupy more than one grid box and less than one grid box for nadir retrievals. For overlapping points from different passes, the closest grid box to the nadir FOV is retained. The main available products are: cloud liquid water, ice water path, rain rate and raw brightness temperature for different channels, among others. Other ancillary parameters like local zenithal angle, latitude and longitude of the closest remapped pixel, data and time and type of orbit are also stored in the HDF-EOS file. More details on MSPPS are given by Ferraro et al. [2005]. AMSU-A and AMSU-B swath data are also used for regional and case studies.

[6] Several data sets are used to compare AMSU-B retrievals with other rainfall retrieval sources. SSM/I EDRR (Special Sensor Microwave/Imager Environmental Data Records, Ferraro, 1997) global coverage data (0.5° spatial resolution), SSM/I GPROF V6.0 (Goddard Profiling Algorithm, Kummerow et al., 2001) global coverage data (0.25° spatial resolution), GPCP version-2 (Global Precipitation Climatology Project, Adler et al., 2003) monthly mean precipitation data (2.5° spatial resolution) and TMI 2A12 rain retrieval algorithm [Kummerow et al., 1996] are used to compare the improved algorithm results (CORR) with the previous version (UNCORR; Weng et al., 2003; Ferraro et al., 2005; Qiu et al., 2005).

[7] The main advantage of this scattering signature approach is that the availability of three NOAA POES satellites spaced approximately 4 h apart with a spatial resolution of 16 km at nadir and a wider swath than the SSM/I, AMSR-E and TMI sensors. This unique feature of NOAA POES satellites series offers an excellent opportunity to reduce the errors associated with the inadequate sample of the diurnal cycle. Nevertheless, the weakness of this technique is related with the inability to retrieve rain that has little or no ice and the cross-scan characteristics of the instrument (different footprints for different local zenithal angles). This situation tends to generate, in the first case, much less rain coverage over the oceans than other algorithms based on emission techniques [i.e., Kummerow et al., 2001] and a shift in the frequency distribution for different zenithal angles in the second case.

[8] Figure 1 shows the relative frequency histogram of AMSU-B derived rain rate bins (mmh−1) for different Local Zenithal Angles (LZA) with coincident SSM/I EDRR rain rate bins over the ocean around the globe. This analysis was performed using NOAA-15 AMSU-B and SSM/I F-13 satellites in the region limited by 60N–60S for year 2005. Only NOAA-15 and SSM/I F-13 collocated data where used in this case because they have similar Local Time Ascending Node (LTAN) orbit crossing times (approximately 18:00–19:00 LST).

Figure 1.

Relative frequency histogram for different Local Zenithal Angles (LZA) over the ocean. Crosses belong to NOAA-15 AMSU-B uncorrected algorithm while pointed-plus line represents the PDF for SSM/I F13 EDRR estimates.

[9] An unrealistic shift in the peak of the relative frequency histogram can be observed for low rain rates for AMSU-B retrievals. Beam positions around 43–48 (near nadir), have a larger shift; while beam positions 1–4 or 86–90 (pixels near the limb) has a smaller shift (and less dispersion also). This situation is also compared with a conical radiometer retrieval (SSM/I EDRR, in this case) which shows a more realistic behavior.

[10] On the other hand, it is also observed that, for high rain rates (Figure 2), the frequency of AMSU rain rate bins over ocean is lower than SSM/I EDRR rain rate bins. This situation is more evident for beam positions 1–4 and 86–90 (near limb). This situation is related with the cross-scan characteristics of the instrument (different footprint sizes for different beam positions).

Figure 2.

As in Figure 1 for rain rates higher than 10 mm/h.

[11] This behavior in the AMSU-B rain rate PDF is also observed over land (not shown) where the shift in the peak in the relative histogram for lower rain rates is less evident, but the absence of high rain rates (15 mmh−1) for near-limb positions is remarkable when compared with SSM/I. EDRR retrievals. A closer analysis on these collocated retrievals shows that the position of the peak in the histogram is not only related with the surface type but also with the latitude. The peak position in a LZA-latitude contoured 2-dimensional histogram (Figure 3) shows that the higher shift is observed in the tropical region for low zenithal angles (near nadir), while over the land (left panel) this effect is less evident than over ocean (right panel). High latitudes show a smaller shift in the peak position than low latitudes for both surface types (land and ocean) because the rain rate decrease naturally toward the pole, shifting the histogram toward lower values.

Figure 3.

AMSU-B derived rain rate histogram mode (histogram peak position in Figure 2) over land (left) and ocean (right) versus latitude and Local Zenithal Angles (LZA) in mm/h.

[12] To model this systematic bias, a second order polynomial interpolation was applied to the original data showing an excellent agreement. The absolute difference between both the original data and the model for both surface type (land and ocean) is less than 0.6 mm/h, while the square correlation coefficient (R2) is greater than 0.95. A similar analysis is performed for the standard deviation of the PDF over land and ocean respectively (not shown). The results are similar to those obtained for the previous analysis using the peak position in the 2-dimensional histogram.

[13] The normalization process of AMSU-B derived rain rates is carried out in two steps. The first one to correct the systematic bias is performed using a Gaussian PDF with μ (peak histogram position) and á (standard deviation of observed distribution) depending on LZA, latitude and surface type. For high rain rates, a linear correction scheme with a slope value depending on the square root of the ratio between SSM/I footprint, which it is constant, and AMSU-B footprint, which depends on LZA (Figure 4, solid line), is proposed to correct AMSU-B derived rain rates. The value of that slope is shown in Figure 4 (dashed line). The slope value ranges from 2, for beam position around 43–48 to 4 near the limb.

Figure 4.

AMSU-B footprint area (solid line – left axis) and calculated AMSU-B/SSM/I slope (dashed line – right axis) for different LZA.

[14] The derived correction scheme is presented in Figure 5 for three different local zenithal angles (LZA = 0, near nadir; LZA = 28, between nadir and limb; and LZA = 56, near limb) and two different latitudes (Lat = 10°, upper panel and Lat = 55°, lower panel) representing a tropical and a midlatitude environment for land (left panel) and ocean (right panel). While for low rain rates, this scheme generates a reduction of the uncorrected rain rates (depending on the LZA and latitude); for high rainfall rates, a general increase in the corrected values is observed after applying the proposed correction scheme due to a higher-than-one slope value.

Figure 5.

Correction scheme for AMSU-B derived rain rates (in mm/h) for three different local zenithal angles (LZA = 0, near nadir; LZA = 28, between nadir and limb; and LZA = 56, near limb) and two different latitudes (Lat = 10°, upper panel and Lat = 55°, lower panel) for land (right column) and ocean (left column).

[15] The second step in the correction scheme is based on the fact that the current technique presents an inability to retrieve rain that has little or no ice. Over the ocean, Cloud Liquid Water (CLW) (this parameter is generated within the current suite of AMSU-A operational products; Ferraro et al., 2005) is proposed as a proxy for retrieving rainfall. In addition to CLW, the convective index (CI) is also used to represent the environment where the cloud is developing. The CI is calculated using the AMSU-B moisture channels, which consist of a pair of narrow bandpass channels on either side of the H2O absorption line center at 183.3 GHz (a detailed definition of this parameter is given by Ferraro et al., 2005). Three different situations can be defined with this parameter: CI = 1 is an indicator of weak convection or stratiform rain; CI = 2 represents moderate convection while CI = 3 indicate the presence of strong convection in the environment. Figure 6 shows the mean value and standard deviation for collocated AMSU-B rain rates and AMSU-A CLW the stratiform case (CI = 1) and the strong convective environment (CI = 3) for year 2005.

Figure 6.

Mean AMSU-B derived rain rate and standard deviation for different environments where the cloud is developing. CI equal 1 represents a stratiform environment while CI equal 3 is a strong convective one. Satellite: NOAA 15 - 60N–60S: Year: 2005.

[16] Considering a conservative assumption, CLWcrit = 0.4 mm is considered as a threshold of rain/no-rain screening (CLWcrit = 0.2 in the work of Ferraro, 1997). Such a high value was chosen to avoid false signatures in moist, non-raining clouds and the uncertainty of the AMSU CLW calculations. A least squares linear method is used to fit the data obtained in the previous analysis. The correlation coefficient (R2) is greater than 0.90 in all cases, showing a good agreement between both parameters. The obtained relationship is used to fill those pixels in: (a) lack of ice structure (Tb89 < Tb150) because scattering of ice particles is larger for 150 GHz than for 89 GHz [see Zhao and Weng, 2002, Figure 4] and, (b) rr = 0 but CLW ≥ CLWcrit.

[17] Over land, where no CLW can be retrieved, only the Convective Index (CI) has been used to retrieve rainfall considering a mean rain rate value for each convection environment (CI) over those pixels with no ice structure (Tb89 < Tb150) only. It is important to point out that this issue is not so important over land as over ocean because the AMSU-B retrieval algorithm is based on a scattering approach (typically used over land for PMW estimations).

3. Evaluation

[18] This part of the paper will demonstrate the validation work done so far using the improved rain-rate algorithm. First, we briefly compare the correction scheme results (CORR) with the operational algorithm (UNCORR) on a monthly quasi-global scale (HMAP daily data - 60N to 60S). Matched data between AMSU-B NOAA-15 derived rain rates and SSM/I GPROF F-13 estimates for 2005 were used to perform this analysis. Both the amount of rainfall and frequency of rain events will be discussed in the next section. A second comparison analysis between Global Precipitation Climatology Project (GPCP) monthly data and a combined AMSU-B (using all available NOAA POES satellites) retrieval will be also performed.

[19] Then, we closely investigate the performance of the new algorithm in regional scale for different surface types and the inclusion of a couple of case studies will demonstrate how the correction technique works.

3.1. Global Scale

[20] Once the correction scheme had been applied to daily AMSU-B HMAP derived rain rate data set for 2005, the new corrected PDF (CORR) over the ocean was obtained (Figure 7). In the same figure, the operational algorithm (UNCORR) and SSM/I EDRR rain rates histograms bins are also included. An improvement for light rain rates (below 10 mm/h) is observed for all beam positions, but it is more remarkable near nadir. A similar result was obtained over land (not shown). The new corrected distribution function for both surface types is closer to SSM/I EDRR histogram than the operational algorithm. It is important to point out that no SSM/I data were used to build the correction scheme, so SSM/I retrievals may be considered as an independent data set to compare with CORR and UNCORR outputs.

Figure 7.

Relative frequency histogram for different Local Zenithal Angles (LZA) over the ocean. Crosses belong to NOAA-15 AMSU-B uncorrected algorithm, pointed-plus line represents the PDF for SSM/I F13 EDRR estimates, while corrected AMSU-B derived rain rates is represented by dashed-asterisks line.

[21] Figure 8 shows the global monthly mean precipitation rate (mm/day) for April 2005 derived from NOAA-15 data using the corrected (CORR) and operational algorithm (UNCORR). SSM/I GPROF retrievals are also included for comparison purposes. The correction scheme produces an overall reduction in the rainfall amount over ocean and over land, especially over the tropical region. This reduction can be most clearly seen in northern South America, the western Pacific Ocean and over the SPCZ. On the contrary, beyond 40°N and 40°S a small increase in the rainfall rate is observed. Similar results were found for other months (not shown). The zonal mean rain rate (Figure 9) shows a different behavior according to latitude and surface: while major improvements can be found in tropical region relative to the uncorrected AMSU-B product, a closer agreement between corrected AMSU-B retrievals and SSM/I GPROF (dashed line) is achieved for all latitudes and all surface types. In addition, the improved rain rate estimate results in lower bias error and RMSE when compared with SSM/I GPROF retrievals. Monthly mean absolute bias error and RMSE (Figure 10) for the tropical region (20° N, 20° S) reflect this agreement along the year. Similar behavior is observed for mid and high latitudes (not shown).

Figure 8.

Mean rain rate of AMSU-B NOAA-15 operational derived rain rates (mm/day) (upper panel), corrected values (middle panel) and mean derived rain rates (mm/day) for SSM/I F-13 GPROF v6.0 for April 2005.

Figure 9.

Zonal mean rain rate over ocean (left panel) and land (right panel) for AMSU-B operational product (dotted green line) corrected AMSU-B retrievals (solid black line) and SSM/I GPROF V6.0 for April 2005 (dotted red line).

Figure 10.

Monthly mean absolute bias (upper panel) and RMSE (bottom panel) for UNCORR and CORR algorithm compared with SSM/I GPROF V6.0 estimates for 2005. Left column belongs to comparisons over land for tropical region (20°N, 20°S), while the right column is over ocean for the same region.

[22] In 2003 the International Precipitation Working Group (IPWG) began a project to validate and intercompare satellite rainfall estimates [Ebert et al., 2006]. Some categorical statistics such as bias score (BIAS), probability of detection (POD), false alarm ratio (FAR), Critical Success Index (CSI) and Heidke Skill Score (HSS) can be computed for different rain rate thresholds as follows: 0.1, 1, 2, 5, 10 and 20 mm. All these parameters can be computed from a rain/no-rain contingency table and measure the performance of a given algorithm [see Wilks, 1995 for more details]. Figure 11 shows the bias score between the CORR and SSM/I GPROF retrievals and between UNCORR and SSM/I GPROF for April 2005 for different rain rate thresholds. This parameter shows that UNCORR algorithm tended to overestimate the rain rate (bias score >1) when compared with SSM/I GPROF. This is more evident for higher rain rate thresholds, so larger overestimates between UNCORR and CORR algorithms are expected for higher rain rates. This ‘close-to-one’ value of bias score observed for CORR for all thresholds means that all values in the CORR-SSM/I scatterplot are equally distributed along 1:1 line.

Figure 11.

Bias score for CORR and UNCORR algorithms for April 2005 over the ocean for tropical (left column) and mid latitudes region (right column).

[23] Table 1 shows a weighted average of all categorical statistics, using weighting factors equal to the corresponding rain rate threshold (higher rain rates are given more influence because of their greater hydrologic importance) for 2005. In this table, the results are generated for different regions (tropical and midlatitudes) and different surface types (land and ocean). An improvement in False Alarm Ratio (FAR) for CORR against UNCORR is observed because this parameter measures the fraction of AMSU-B occurrence of rainfall which is not observed in SSM/I estimates; while POD for UNCORR algorithm is higher than for CORR algorithm because UNCORR tends to overestimate the amount of rainfall, so missing higher rain rates which have more weight in the average (observed by SSM/I and not observed by AMSU-B uncorrected) are very little. The CSI and HSS show improvements with the new correction scheme, especially over the ocean for both regions (tropical and midlatitudes).

Table 1. Probability of Detection (POD), False Alarm Ratio (FAR), Bias Score (BIAS), Critical Success Index (CSI) and Heidke Skill Score (HSS) Categorical Statistics for Tropical Region and Mid-latitudes for Different Surface Types (Land and Ocean) When CORR and UNCORR are Compared Against SSM/I GPROF Retrievals
Tropical Region (0°–20°)
  FARPODBIASCSIHSS
LANDUNCORR0.460.932.020.530.59
CORR0.310.871.310.640.73
OCEANUNCORR0.330.901.400.630.71
CORR0.200.821.030.690.78
 
Mid Latitudes (20°–40°)
  FARPODBIASCSIHSS
LANDUNCORR0.290.670.970.530.65
CORR0.180.690.960.570.68
OCEANUNCORR0.410.711.300.480.60
CORR0.310.670.980.520.64

[24] Not only is the amount of rainfall an important variable to estimate, but also the frequency of rain events is also critical for several hydrological and environmental applications. According to the algorithm development described in the previous section, the first step of the correction scheme will not produce any change in the frequency of rain events (pixels with rain rates > 0) because this scheme only produce a ‘rescaling’ process in the rain rate estimate; but the second one, will fill up the gaps where the operational algorithm presents an inability to retrieve rain that has little or no ice structure, so the frequency of rain events is altered.

[25] To evaluate this frequency, the ratio between rain events (rr > 0 mm h−1) and valid observations (rr ≥ 0 mm h−1) was performed for the corrected (CORR) and operational algorithm (UNCORR) and SSM/I GPROF V6.0 retrievals. Figure 12 shows the zonal mean of that ratio for the different algorithms over the ocean and over land. Over the ocean, CORR shows a large improvement of relative frequency in all latitudes compared with UNCORR.

Figure 12.

Zonal Mean frequency ratio (rr > 0/rr ≥ 0) of AMSU-B NOAA-15 operational algorithm (UNCORR – green dotted line), for corrected values (CORR – black solid line) and for SSM/I F-13 GPROF V6.0 (red dashed line) for April 2005.

[26] Conversely, over land, the AMSU-B retrieval algorithm shows a larger frequency than SSM/I GPROF V6.0 (this fact is more evident over the equatorial region) because the higher frequency channels used in AMSU-B retrievals are more sensitive to lighter rain rates [Qiu et al., 2005]. In this case, there are almost no differences between CORR and UNCORR zonal profile. This result is in good agreement with the fact that AMSU-B algorithm is based on a scattering approach, so indeterminations due to lack of ice structure is an uncommon feature over land. Nevertheless, when this result is analyzed together with the zonal mean rain rate (Figure 9), it can be observed that although the frequency remains unaltered between CORR and UNCORR and greater than SSM/I GPROF, the CORR output is closer, in amount, to the mean SSM/I retrievals.

[27] Another source of reliable monthly global scale data is the Global Precipitation Climatology Project -version 2- product [Adler et al., 2003, GPCP herafter]. This globally complete, monthly analysis of surface precipitation is performed at 2.5° latitude × 2.5° longitude resolution, so the AMSU-B derived rain rates (HMAP data set) are remapped and monthly averaged along 2005 in order to compare both analyses. GPCP is a merged analysis that incorporates precipitation estimates from low-orbit satellite microwave data, geosynchronous-orbit satellite infrared data, and surface rain gauge observations. The merging approach utilizes the higher accuracy of the low-orbit microwave observations to calibrate, or adjust, the more frequent geosynchronous infrared observations.

[28] Figure 13 shows the zonal monthly mean for January, April and July 2005 for the operational algorithm (UNCORR), the corrected values (CORR) and GPCP. In this case, a good agreement between CORR and GPCP is observed over all surfaces, especially over the tropical and mid latitude region (40°N, 40°S). It is important to point out that both algorithms AMSU-B and GPCP use scattering approaches over land (Ferraro, 1997 for GPCP; Ferraro et al., 2005 for AMSU-B) as primary source of data, while a microwave emission based schemes are used to retrieve rainfall over the ocean by GPCP [Wilheit et al., 1991]. Nevertheless, the larger disagreement between GPCP and AMSU-B derived rain rates occurs for high latitudes (beyond 40°N). This fact is due to the inclusion of TOVS data in GPCP rainfall retrieval poleward 40° [Adler et al., 2003].

Figure 13.

Zonal mean rain rate over ocean (left panel) and land (right panel) for AMSU-B operational product (dotted green line) corrected AMSU-B retrievals (solid black line) and GPCP (dotted red line) for January (upper panel), April (middle panel) and July 2005 (bottom panel).

3.2. Regional Scale and Case Studies

[29] For the regional study, swath AMSU-A and AMSU-B/MHS data from NOAA 15, 16, 17, and 18 for July 2005 were processed to obtain the corrected rainfall estimations. In this case, high quality TMI 2A12 surface rain [Kummerow et al., 1996] is used for comparison.

[30] In pre-processing, both the AMSU-B swath data (CORR and UNCORR) and the TMI 2A12 estimations are remapped in a regular grid of 0.25 degree and, after that, those collocated pixels within 30 minutes difference between NOAA and TMI were retained to perform the statistics. Considering these constrains, and also keeping those non zero-zero pairs, less than 10 samples (in average) were obtained for each point. Three ocean regions (Atlantic, Western Pacific and Indian Ocean) and three land regions (America, Africa and Southeastern Asia) were selected to evaluate the performance of the correction.

[31] The Atlantic Ocean region (15 S–40 N, 105 W–15 W) shows the rainfall region has definitely increased relative to the UNCORR algorithm, but not enough to match TMI area extension (Figure 14). The overall rainfall amount for CORR is about 10% less than TMI. In this case, the zonal correlation coefficient (non zero-zero pairs at 0.25 degree half hourly matches – around 2000 cases per 10-degrees latitude band) shows some improvement for all latitude bands, especially in the southern hemisphere (between the Equator and 10°S) compared with UNCORR algorithm. Similar features are observed for the Indian Ocean (15°S–40°N, 45°E–135° E) (not shown), while over the western Pacific (15°S–40°N, 135°E–135°W) the biggest correlation improvement (relative to the UNCORR algorithm) seems to be on the edges of the ITCZ. The corrected version seems to increase the area of ITCZ, and relatively decrease the magnitude of convective cells within the core, generating a more reliable pattern compared with TMI 2A12 (Figure 15, left and central panel). In this case, due to an increase in the frequency of rain events, the number of cases per 10-degrees latitude band is around 4000 samples for July 2005. On the other hand, the new rain rate distribution curve compared to the UNCORR algorithm distribution shows a better agreement with high quality TMI estimates, especially for lower rain rates (Figure 15, right panel).

Figure 14.

Left panel: Mean rain rate for AMSU-B CORR estimates over Atlantic Ocean from 1–31 July 2005. Central panel: Idem for TMI 2A12. Only scenes within 30 minutes are considered in the average. Right Panel: correlation coefficient of different estimates compared against TMI 2A12.

Figure 15.

Left panel: Mean rain rate for AMSU-B CORR estimates over western Pacific Ocean from 1–31 July 2005. Central panel: Idem for TMI 2A12. Only scenes within 30 minutes are considered in the average. Right Panel: Rain rate distribution for AMSU CORR and TMI 2A12.

[32] Over the American continent (15°S–40° N, 125° W–35°W), CORR shows the same correlation as the UNCORR algorithm, but probably about 15% less rainfall than the TMI (Figure 16). The rain rate distribution for CORR algorithm has a good agreement with TMI in the rates heavier than 10 mm.h−1. Almost all of the CORR bins over land are under 10 mm.h−1. Over Africa (15°S–40°N, 25°W–65°E), there same slight total bias deficiency, but right on the TMI totals in the Southeastern Asia (15°S–40°N, 65°E–155°E) plot and also slight correlation improvement for all latitude bands in that region (not shown).

Figure 16.

Left panel: Mean rain rate for TMI-2A12 estimates over the Americas from 1–31 July 2005. Central panel: Idem for AMSU-B CORR. Right Panel: Idem for AMSU-B UNCORR. Only scenes within 30 minutes are considered in the average.

4. Summary and Conclusion

[33] This paper has presented an evaluation of a correction scheme for AMSU rain-rate algorithm through different comparisons with several PMW retrievals. The corrected AMSU rain-rate retrieval has been shown to have the ability to provide valuable rainfall estimates for different regions of the world and surface types.

[34] Taking into account the main advantage of having four NOAA satellite plus the recently launched MetOp European polar orbiter satellite, a new successful correction scheme has been developed based on the AMSU-B footprint geometry (different footprint size for different beam positions) and AMSU-A CLW rainfall proxy to fill in the gaps of AMSU-B retrievals over the ocean.

[35] After applying the proposed scheme, a useful improvement of rain rate distribution is observed for both surface types (land and ocean) compared with UNCORR distribution function. The new distribution is also in good agreement with other rain rate estimations such as the SSM/I EDDR on the global scale and with TMI 2A12 for different regions of the world and surface types. This improvement is more impressive for closer-to-nadir beam positions and lower rain (rr < 10 mm).

[36] On the global scale, this correction scheme produces different results according to the latitude band considered. Over the tropical region, the correction scheme produces an overall reduction in the rainfall amount over both surface types, while the area covered by rain shows an increase over the ocean but, over land, there are almost no differences between corrected and uncorrected frequencies (and higher than SSM/I). For midlatitudes, an increase in the amount and frequency of rain events is observed over the ocean, especially poleward of 40 degrees in both hemispheres.

[37] It is also notable that the new corrected outputs are in a good agreement when compared with other PMW techniques (i.e., SSM/I GPROF V6) and high quality collocated retrievals (TMI 2A12) at global and regional scales for both surface types.

[38] In the first case, the reduction of mean absolute bias, when SSM/I GPROF V6 is compared against CORR and UNCORR, is approximately 1.2 mm over land and 0.7 mm over ocean all around the year. Considering the area covered by rainfall, the bias score shows that CORR shows a better performance than the UNCORR estimates when compared with SSM/I GPROF V6 for high rain rates (rr > 10 mm).

[39] Looking at the regional analysis, the same features like global scale are observed for several regions around the world. It is notable that the correlation between TMI 2A12 and CORR collocated estimations is, in most of the cases, better than the UNCORR retrievals.

[40] A new coastal detection algorithm using AMSU measurements [Kongoli et al., 2007] has recently been combined with this correction scheme and the reprocessing of uncorrected AMSU retrievals is ongoing. Preliminary results can be found at the Microwave Surface and Precipitation Products System (MSPPS) Web site (http://www.orbit.nesdis.noaa.gov/corp/scsb/mspps/main.html).

Acknowledgments

[41] This research was supported under NOAA grant # 2007-2000820 to the Cooperative Institute of Climate Studies (CICS), Earth System Science Interdisciplinary Center (ESSIC), University of Maryland, College Park (UMCP). Also this work is sponsored by Christopher Miller of NOAA/Climate Program Office. The first author would also like to acknowledge to the two anonymous reviewers who helped to improve this manuscript.

Ancillary