MOST ACTIVE and passive microwave remote sensing

Transcription

1 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 49, NO. 12, DECEMBER Evaporation Correction Methods for Microwave Retrievals of Surface Precipitation Rate Chinnawat Surussavadee, Member, IEEE, and David H. Staelin, Life Fellow, IEEE Abstract Active and passive microwave remote sensing estimates of surface precipitation based on signals from hydrometeors aloft require correction for evaporated precipitation that would otherwise reach the ground. This paper develops and compares two near-surface evaporation correction methods using two years of data from 509 globally distributed rain gauges and three passive millimeter-wave Advanced Microwave Sounding Units (AMSUs) aboard National Oceanic and Atmospheric Administration (NOAA) satellites (NOAA-15, NOAA-16, and NOAA-18). The first type of correction is a scale factor that minimizes the bias between the means of annual AMSU and rain gauge precipitation accumulations (in millimeters per year) for each of 12 infrared-based surface classifications. The scale factor for the second correction method is computed using a neural network trained using both surface classification and annual average relative humidity profiles. AMSU surface precipitation retrievals using both methods were compared to the annual accumulations for 509 rain gauges uncorrected for wind effects, where different data were used for training and accuracy evaluation. The rms annual accumulation retrieval errors for AMSU using surface classification and relative humidity corrections were 236 and 222 mm/y, respectively, compared to 190 mm/y for corresponding Global Precipitation Climatology Project data, which utilizes more satellite sensors and over 6500 rain gauges. Index Terms Advanced Microwave Sounding Unit (AMSU), evaporation, microwave precipitation estimation, precipitation, rain, remote sensing, satellite, snow, virga. I. INTRODUCTION MOST ACTIVE and passive microwave remote sensing systems estimate surface precipitation rates based on signals from hydrometeors aloft that may partially evaporate before reaching the ground, leading to overestimates. This paper demonstrates how surface precipitation retrievals can be corrected for evaporation by comparing two correction methods developed for precipitation retrievals obtained using the passive millimeter-wave Advanced Microwave Sounding Units (AMSUs) aboard the operational U.S. National Oceanic and Manuscript received October 11, 2010; revised April 4, 2011; accepted July 3, Date of publication August 11, 2011; date of current version November 23, This work was supported by the National Aeronautics and Space Administration under Grant NNX07AE35G. C. Surussavadee is with the Research Laboratory of Electronics, Massachusetts Institute of Technology, Cambridge, MA USA, and also with the Andaman Environment and Natural Disaster Research Center, Faculty of Technology and Environment, Prince of Songkla University, Phuket 83120, Thailand ( pop@alum.mit.edu). D. H. Staelin is with the Research Laboratory of Electronics, Massachusetts Institute of Technology, Cambridge, MA USA ( staelin@ mit.edu). Digital Object Identifier /TGRS Atmospheric Administration satellites NOAA-15, NOAA-16, and NOAA-18 (Microwave Humidity Sounder replaces AMSU-B on NOAA-18.). The correction methods were tested for two years of surface precipitation rates retrieved using the AMSU Massachusetts Institute of Technology Precipitation retrieval algorithm (AMP) [1] [6], 509 globally distributed rain gauges [7], which are a subset of the 787 not-too-high nonhilly gauges used in [6], and the corresponding Global Precipitation Climatology Project (GPCP) estimates [8]. Section II reviews the AMP algorithm, and Section III demonstrates the need for evaporation bias correction. The evaporation correction methods are described in Section IV, the GPCP data are described in Section V, and the results are shown in Section VI. Section VII summarizes and concludes this paper. II. AMSU PRECIPITATION RETRIEVAL ALGORITHM Since May 1998, AMSU cross-track scanning spectrometers on the operational satellites NOAA-15, NOAA-16, NOAA-17, and NOAA-18 have been observing Earth at 20 frequencies ranging from 23 to 191 GHz. AMSU is composed of two units. AMSU-A has 50-km resolution near nadir and observes 15 channels in the 23-GHz water vapor and 53-GHz oxygen absorption bands. AMSU-B has 15-km resolution near nadir and observes five channels between 89 and 191 GHz, including the 183-GHz water vapor resonance [3], [6]. The AMP algorithm has evolved through several versions that retrieve global surface precipitation rates for both rain and snowfall (in millimeters per hour); water path estimates (in millimeters) for rain, snow, graupel, cloud liquid water, cloud ice, the sum of rain, snow, and graupel; and peak vertical wind (convective strength in meters per second) [1] [6]. It employs 13 AMSU channels and neural networks (NNs) trained using brightness temperatures simulated using a cloud-resolving numerical weather prediction (NWP) model [fifth-generation National Center for Atmospheric Research/Penn State Mesoscale Model (MM5)] and a radiative transfer model (RTM) for five hydrometeor species (rainwater, graupel, snow, cloud water, and cloud ice). The simulated brightness temperatures generally agree with those coincidentally observed by AMSUs aboard NOAA-15, NOAA-16, and NOAA-17 over 122 global storms, each 2850 km 2 [1]. This radiance-simulation model, i.e., the U.S. National Center for Environmental Prediction (NCEP)/MM5/TBSCAT/F(λ), utilizes the U.S. NCEP analyses [9], the MM5 [10], an RTM TBSCAT [11], and electromagnetic scattering models for icy hydrometeors, i.e., F(λ) [1] /$ IEEE

2 4764 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 49, NO. 12, DECEMBER 2011 Reference [2] shows that the simulated brightness temperatures are highly sensitive to assumptions in MM5 and the RTM. The version-3 AMP algorithm (AMP-3) successfully retrieves the global rain and snowfall rates, including those over snow-covered land and sea ice [5], but is excluded from the following: 1) high elevation land like the Himalayan, Greenland, and Antarctic plateaus and major mountain chains like the Andes and Alps, and 2) snow or ice-covered surfaces where the atmosphere is so cold and dry that even the opaque channels see that surface, potentially causing false detections [5]. AMP-3 does not retrieve surface precipitation when the surface elevation E is flagged as too high, which occurs when E>2 km for lat < 60, E>1.5 km for 60 < lat < 70, and E>0.5 km for lat > 70. The risk of falsely detecting precipitation is high if the air is extremely dry and transparent over snow or ice-covered surfaces, so they are flagged as too cold when AMSU-A channel 5 (53.6 GHz) is colder than 242 K for such surfaces where deep convection for such surfaces is less likely; this channel sounds temperatures centered near 4-km altitude. AMP-3 estimates are at 15-km resolution. Estimates below 0.5 mm/h are considered uncertain and are set to zero. Comparisons with CloudSat 94-GHz radar [12] and annual surface precipitation rate maps from the GPCP [5], [6] confirm the plausibility of these retrievals. III. NEED FOR EVAPORATION CORRECTION BASED ON RAIN GAUGE COMPARISONS The AMP-3 annual surface precipitation retrievals (in millimeters per year) for years 2006 and 2007 for NOAA-15 and NOAA-16 were compared with those observed by 787 rain gauges globally distributed in not-too-high (defined above) and nonhilly regions (defined as those for which the surface elevation varied by less than 500 m within a box of ±0.2 of longitude and latitude). The rain gauge data were from the NOAA Monthly Climatic Data for the World data set [7] and were obtained from 2000 worldwide quality-controlled surface data collection stations. Their quality was further enhanced by considering only the 1152 sites, reporting data for all 24 months in 2006 and These comparisons were made for 12 surface classifications defined using Advanced Very High Resolution Radiometer (AVHRR) infrared (IR) spectral images and one additional classification defined geographically (coastal within 55 km of coastline) [6], [13]. To compute the annual totals, AMP-3 estimates that were flagged as too cold were approximated as zero since such atmospheres are rare and sufficiently dry that they sometimes become transparent and permit surface snow and ice signatures to produce false precipitation detections. The average annual precipitation ratio (AMP-3)/gauge ranged from 0.88 for tundra to 1.25 for wooded grassland and therefore was reasonable for most surface classifications. It is interesting to note that AMP is tuned only to NCEP/MM5/TBSCAT/F(λ) physics and that the gauge reports are uncorrected for wind effects. Reference [6, Table V] suggests that the evaporation corrections for all surface classifications more arid than wooded grassland are significant beyond the six-sigma level, based on the observed rms variations in the AMP corrections deduced within each surface classification. AMP-3 significantly overestimated surface precipitation for undervegetated land, including grassland, shrubs over bare ground, and pure bare ground (desert) with precipitation ratios of 2.4, 3.1, and 9, respectively; a ratio of nine can be explained only by evaporation. Additional evidence of statistical significance is presented in Section VI. Reference [6] discusses three possible explanations for AMP-3 overestimation, including the following: 1) the AMP-3 training data use incorrect surface emissivity spectra for undervegetated land; 2) the undervegetated land affects hydrometeor habits and cloud drop size distributions such that they promote overestimation; and 3) the overestimation is due to terraincorrelated evaporation near surface (virga). The first hypothesis is less likely since AMP-3 employs only signals mostly insensitive to the surface. It would require very large systematic surface emissivity differences for undervegetated land to cause even factor-of-two retrieval errors, let alone factors of nine [6]. The second hypothesis is also less likely since larger numbers of small hydrometeors possibly promoted by terraindependent nucleating agents such as dust would reduce rather than increase AMP-3 retrieved precipitation values. In addition, 550-nm Moderate Resolution Imaging Spectroradiometer global maps of aerosol optical depth show low correlation with those land classifications where AMP-3 overestimates are high [5], [14]. The third hypothesis was studied in [6] using MM5 simulations. It was shown in [6, Fig. 2] that AMSU overestimates are highly correlated with a virga metric, arbitrarily defined for a given 15-km field of view (FOV) as the ratio between the average hydrometeor density at the highest MM5 layer and that at the lowest MM5 atmospheric layer. The systematic AMP-3 overestimation is thus most likely due to evaporation at altitudes beneath those of the hydrometeors sensed by AMSU [6]. If true, this implies that other precipitation remote sensing systems and retrieval algorithms may benefit from similar corrections. We know that errors in MM5 are not significant contributors to AMP overestimates since MM5 predicts unretrievable virga on the same order as observed, and residual errors in the RTM are evidently not responsible in view of the good agreement at all frequencies between observed AMSU brightness temperature histograms and those predicted using the coincident MM5/RTM model. The high AMP overestimates over desert and dry grassland are also definitive since evaporation there is expected and often observed visually. Finally, diurnal effects are not significant contributors since the AMP performance statistics derived in [6] are based on two satellites in different orbits (NOAA-15 and NOAA-16). In addition, our AMP evaporation correction methods developed in this paper employed data from three satellites (NOAA-15, NOAA-16, and NOAA- 18) for training and accuracy evaluation. IV. EVAPORATION CORRECTION METHODS A. Evaporation Correction Method Based on Surface Classifications The first evaporation correction method for the AMP-3 retrievals employs only surface classifications deduced from

3 SURUSSAVADEE AND STAELIN: EVAPORATION CORRECTION METHODS FOR MICROWAVE RETRIEVALS OF SURFACE PRECIPITATION RATE 4765 the AVHRR IR spectral images [13]. These 12 classifications include tundra, water, high-latitude deciduous forest, broadleaf evergreen forest, mixed coniferous forest, coniferous evergreen forest, broadleaf deciduous forest, wooded grassland, cultivated crops, grassland, shrubs and bare ground, and bare ground, in order of increasing aridity. Due to the possibility of significant coastal irregularities and tidal variation and, therefore, uncertain pixel surface emissivity within slightly uncertain fields of view, areas within 55 km of a coastline were classified separately as coastline instead of using the AVHRR surface classification. Another separate class defined geographically was the latitudes above 75 N where AMP-3 generally cannot operate during very cold weather. Eliminating coastline gauges from the original global set of 787 left 509 that were noncoastal, nonhilly, and not too high. The highest latitude for these 509 gauges is 75 N. The first evaporation correction method minimizes for each surface classification the bias between the AMP-3 and rain gauge annual accumulations (in millimeters per year). The accumulation correction factor for each of the 12 surface classifications is therefore the ratio between the means of its annual rain gauge and AMP-3 accumulations, independent of season and gauge locations within each region. The AMP-3 retrievals corresponding to each gauge were bilinear interpolations of the 1 gridded AMP-3 retrievals to the gauge location. The reciprocal of the average correction factor for each classification then multiplied the AMP-3 retrievals to yield the AMP-4 retrievals [6]. B. Corrections Based on Both Surface Classification and Annual Relative Humidity Profiles Although [6] shows by means of AMSU/gauge data that extreme AMSU overestimation (evaporation) is strongly correlated with surface classification and aridity, land surface classification does not fully reveal global atmospheric variations or predict variations over undifferentiated ocean or large lakes. Therefore, the second evaporation correction method employs both surface classification, which indicates aridity over land by means of a single integer (1 12) corresponding to the surface classifications listed in Section IV, and the annual relative humidity profile, which provides similar aridity information in a global multivariate form. We find that the additional use of climatological relative humidity profile information not only permits its use over ocean but also improves rms retrieval performance in areas receiving below 900 mm/y, as discussed later, thereby further justifying this approach. NCEP relative humidity analyses [9] with 1 resolution are available on 21 discrete pressure levels ranging from 1000 to 100 mbar. Since altitudes above 300 mbar lie above most precipitation, the relative humidity was bilinearly interpolated to gauge locations at only the bottom 17 pressure levels, which ranged from 1000 to 300 mbar. Annual accumulations from 509 gauges contributed 1018 samples for the two-year period studied. Fig. 1 suggests the utility of annual relative humidity profiles for evaporation correction. The pluses in Fig. 1 show correlation coefficients between 19 meteorological variables and the base-ten logarithm of ob- Fig. 1. (Pluses) Correlation coefficients between the base-ten logarithm of observed AMP-3/gauge precipitation accumulation ratios and the 19 variables. (Squares) Correlation coefficients between the base-ten logarithm of AMP-3 annual precipitation accumulations and the 19 variables. Black and gray symbols indicate gauges 2000 and gauges > 2000 mm/y, respectively. The correlations were computed using the 991 training samples. served AMP-3/gauge annual accumulation ratios for the 991 annual training samples, which are from 509 gauges for years 2006 and 2007 and have AMP-3/gauge annual accumulation ratios between 0.3 and 20. Twenty-seven samples with AMP-3/ gauge annual accumulation ratios less than 0.3 or greater than 20 were not included in the training samples to avoid outliers, but they were included in the evaluating samples. In numerical order, the 19 variables include the following: an integer ranging from 1 to 12 in order of increasing presumed surface aridity (excluding coastline and locations above 75 N), the base-ten logarithm of AMP-3 annual precipitation accumulation, and the base-ten logarithm of annual average relative humidity profiles for 17 pressure levels (1000, 975, 950, 925, 900, 850, 800, 750, 700, 650, 600, 550, 500, 450, 400, 350, and 300 mbar, respectively). The squares in Fig. 1 show the correlations between the base-ten logarithm of AMP-3 annual precipitation accumulations and the 19 variables, where self-correlations (variable two) were not plotted. Black and gray symbols represent gauge annual accumulations 2000 mm/y (971 samples) and > 2000 mm/y (20 samples), respectively, where the breakpoint at 2000 mm/yr was arbitrarily chosen to illustrate clearly the influence of accumulation upon the correlation. The first observation is that surface classification (variable one) is correlated ( 0.7) with AMP-3/gauge annual accumulation ratios (black and gray pluses) for both accumulation regimes; thus, surface classification matters even for rainy sites. Second, for gauges 2000 mm/y, although the AMP-3/gauge annual accumulation ratio is correlated (black plus = 0.7) with surface classification (variable one), it is nearly uncorrelated ( 0.1) with AMP-3 annual precipitation (variable two). This demonstrates that surface classification predicts the needed correction significantly better than does the annual accumulation. Third, the AMP-3/gauge annual accumulation ratios for gauges < 2000 mm/y are strongly negatively correlated ( 0.8) with the average annual relative humidity at pressures

4 4766 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 49, NO. 12, DECEMBER 2011 below 750 mbar (variables three to ten). This suggests that, when the air is dry at altitudes below the sensed hydrometeors, hydrometeor evaporation and, hence, AMP-3 overestimates increase. Thus, climatologically low relative humidity below 750 mbar is correlated with overestimation of precipitation by satellites. It therefore provides an additional means for correcting surface precipitation-rate retrievals for evaporation with the added benefit that similar corrections for relative humidity should apply over ocean whereas land surface classification cannot. In contrast, AMP-3 annual accumulations for gauges < 2000 mm/y are nearly uncorrelated (black squares = 0.1) with the average annual relative humidity at pressures below 750 mbar (variables three to ten). Fourth, for gauges > 2000 mm/y, the AMP-3/gauge annual accumulation ratios are poorly correlated (gray pluses < 0.4) with average annual relative humidity below 750 mbar (variables three to ten). Thus, average relative humidity contributes little additional information about near-surface evaporation in regions that rain heavily, e.g., in the highly humid tropics where fractional evaporation is reduced. The second evaporation correction method uses an NN to estimate the correction based on the 19 input variables listed earlier, thus taking advantage of relative humidity climatology at all altitudes below 300 mbar. The NN has three layers composed of three, two, and one neurons, where the transfer functions for the first two layers are hyperbolic tangent sigmoids and the output layer is linear. The NN output is the baseten logarithm of observed AMP-3/gauge annual accumulation ratios; the desired correction factor is the reciprocal of this ratio. For each of the two independent subsets of the gauge annual accumulations (in millimeters per year), the NN training used a fixed 50% of the accumulation reports for training, 25% for validation (i.e., to decide when to stop training), and 25% for testing (i.e., to decide which NN performed best). Each of the two gauge data sets A and B randomly included approximately half of all 1018 annual accumulations, 509 of which were obtained in each of the years 2006 and 2007 but no data in common. Since both sets spanned the globe, erroneous correlations between the pair should be negligible relative to the large dynamic range among gauges. The best of 200 trained NNs with respect to rms error was selected for each subset of approximately 509 samples. For each subset of A and B, the best NN yielded a 1 resolution correction map for each of the 509 sites; these two subset results were then averaged to yield the correction map used to convert AMP-3 retrievals into AMP-5 retrievals. This dual-nn approach reduced the sensitivity to NN training errors as discussed in Section VI. Since annual near-surface relative humidity is not very useful for estimating evaporation when annual accumulations are greater than 2000 mm/y, AMP-4 correction ratios were used instead when the AMP-4 retrievals exceeded 2000 mm/y. V. GPCP PRECIPITATION DATA The international GPCP was established by the World Climate Research Program. Its initial product was the GPCP satellite-gauge precipitation product (SG), which offers monthly precipitation estimates on a 2.5 latitude/longitude grid. Version 2.1 is the latest and is available from January 1979 to the present [15], [16]. In order to provide precipitation estimates with finer spatial and temporal resolutions, the GPCP 1 daily product (1DD) was developed [8]. The 1DD product provides daily precipitation estimates on a 1 latitude/longitude grid and is available from October 1996 to the present. Its latest version is version 1.1. The 1DD daily precipitation totals are scaled to sum to the SG monthly precipitation totals. The SG product incorporates the precipitation information from the following: 1) the Global Precipitation Climatology Centre (GPCC) gauge analyses based on rain gauges worldwide [17]; 2) microwave emission precipitation estimates over ocean [18] and microwave scattering precipitation estimates over land [19] from the Special Sensor Microwave Imager (SSM/I) aboard the Defense Meteorological Satellite Program; 3) IR precipitation estimates for latitudes 40 S 40 N from geostationary satellites (the Geostationary Operational Environmental Satellites, the Geosynchronous Meteorological Satellite, and the Meteorological Satellite) and from the AVHRR aboard NOAA polar-orbiting satellites when geostationary data are unavailable [20]; 4) precipitation estimates from the Television and Infrared Observation Satellite Operational Vertical Sounder (TOVS) [21], which also utilizes the AMSU data; and 5) precipitation estimates based on the use of low-earth orbit satellite outgoing long-wave radiation observations [22]. TOVS precipitation estimates were replaced by Atmospheric Infrared Sounder estimates [21] in May The finer spatial and temporal resolutions of the 1DD data were achieved by the use of fine time and space sampling of the geostationary IR data and the fairly accurate instantaneous estimates from microwave radiometers. Within GPCP, wind-loss corrections are applied initially only to the GPCC gauge analysis, but a smoothed large-scale gaugebias adjustment based on these wind-loss-corrected analyses is applied to the multisatellite estimate before it is combined directly with the gauge analysis. The smoothing operation involves convolution with square boxcars of sizes varying from 12.5 to 15 in longitude and latitude. Antarctica, Iceland, and parts of Greenland are not adjusted. Wind losses arise in part when wind blowing past a rain gauge is partially forced upward by the gauge so as to prevent some precipitation from entering it; the needed correction depends in part on local turbulence, gauge geometry and location, etc. The same wind can also promote limited evaporation within the gauge itself before the precipitation is next measured while exiting; the time between measurements thus affects the wind bias, as does the relative humidity and gauge temperature. These wind and evaporation errors are proportionally larger for lower rain rates and snowfall because snowflakes and small droplets are blown and evaporated more easily, and time intervals between such measurements are typically longer. The correction factors employed by GPCP range between 0.87 and 3. The quality of rain gauge reports also varies from site to site, perhaps due to management practices. For example, of the 2000 NOAA rain gauge sites that were surveyed, only 1152 sites reported values for all 24 months (years 2006 and 2007), and only these more regular sites were utilized.

5 SURUSSAVADEE AND STAELIN: EVAPORATION CORRECTION METHODS FOR MICROWAVE RETRIEVALS OF SURFACE PRECIPITATION RATE 4767 The resulting 1DD version-1.1 GPCP product used in this paper was interpolated to the 509 gauges locations, and the 2.5 interpolated monthly global climatological rain gauge wind-loss ratio adjustment available only over land [23] was incorporated, where the maximum adjustment used by GPCP was three in order to avoid potentially large unrealistic adjustments (D. T. Bolvin, personal communication, 2010). This adjustment was undone in this paper to make GPCP compatible with the 509 gauges that were uncorrected for wind loss and that were used here for training the AMP algorithms and as a standard for comparison. TABLE I RMS ERRORS, BIASES (Estimate Gauge), AND CORRELATION COEFFICIENTS FOR AMP-3, AMP-4, AMP-5, AND GPCP FOR 509 GAUGE ANNUAL OBSERVATIONS (1018 ANNUAL SAMPLES) VI. RESULTS Both evaporation correction methods were evaluated by comparing the AMP-3 surface precipitation retrievals for the years 2006 and 2007 with 509 rain gauges located in noncoastal and nonpolar ( lat < 75 ) sites, which are a subset of the 787 not-too-high and nonhilly gauges used in [6]. Only gauges for which 24 consecutive monthly reports were available were used because they are believed to be more reliable. AMP-3 surface precipitation retrievals for NOAA-15, NOAA-16, and NOAA- 18 were averaged into a 1 grid separately for years 2006 and The annual accumulations (in millimeters per year) were then computed and interpolated to rain gauge locations. To avoid using the same rain gauge records for both training and evaluation of both the surface classification and NN methods, the gauge subset A was used for training 200 NNs and choosing the best, and subset B was used for accuracy evaluation, and then, the roles of A and B were reversed so as to provide two independent sets of performance results that could be combined. These results led to the entries for AMP-4 and AMP-5 shown in Table I. Some fraction of these errors is due to the finite size of the training and evaluation data sets. This rms fractional error was estimated to be 11% for the 1018-sample annual accumulation data set, based on the scatter between the independently derived A and B correction factors for the 509 sites. This accuracy could be improved by using more annual samples to derive the corrections. Thus, correction factors between 0.9 and 1.1 are not sensibly different from unity, whereas those above approximately 1.33 become increasingly statistically significant beyond the three-sigma level, a conclusion that is generally consistent with the rms dispersions found within the various surface classes reported in [6, Table V]. The rain gauges and gauge-based adjustments used for the AMP-4 and AMP-5 retrievals did not incorporate the windloss adjustment because those adjustments are believed to be outdated due to changes in many gauges over the intervening years. They could readily be applied to AMP-5 retrievals, however, so as to yield better agreement with GPCP averages. Fig. 2 shows the scatter plots of AMP-3 [5], AMP-4 [6], AMP-5, and GPCP versus gauge annual precipitation accumulations (in millimeters per year). AMP-5 performs better than AMP-4, and both AMP-4 and AMP-5 perform much better than AMP-3, reducing both its biases and rms scatter. For example, the apparent cutoff below 200 mm/yr for AMP-3 retrievals due to uncorrected evaporation at the driest sites essentially disap- Fig. 2. Scatter plots of AMP-3, AMP-4, AMP-5, and GPCP versus gauge annual precipitation accumulations for 1018 samples. pears for AMP-4 and AMP-5. The good agreement between the GPCP analyses and rain gauges having low annual accumulations arises partly because GPCP incorporates rain gauge measurements from the GPCC gauge analyses when satellite data are less reliable, e.g., in dry climates. The effects of this GPCC gauge-truth overlap are most evident in the reduced scatter for annual accumulations below 500 mm/y shown in the GPCP plot in Fig. 2. Table I presents the rms errors, biases (estimate gauge), and correlation coefficients for AMP-3, AMP-4, AMP-5, and GPCP when compared with the set of 509 rain gauges (1018 samples) and then compared separately for gauge annual accumulations 900 mm/y (795 samples) and > 900 mm/y (223 samples). These gauges are all located below 75 N latitude because those above 75 N are either coastal, in hilly sites, or at excessive altitude. The break at 900 mm/y was chosen to make the differences more evident; other breakpoints in annual accumulations begin to blur the distinction between wet and dry meteorological regimes and their relative correction performance. Consistent with the scatter plots in Fig. 2, AMP-4 and AMP-5 greatly improve AMP-3, and AMP-5 performs better than AMP-4 in terms of rms errors and correlation coefficients. AMP-4 has less bias than AMP-5 because the AMP-4 land-class correction ratios were tuned to make the expected biases zero. AMP-5 reduces the rms errors of AMP-3 by 60.5% and 18.5% for gauge annual accumulations 900 and > 900 mm/y, respectively. Biases and correlations were also substantially improved. The inclusion of relative humidity profiles at low altitudes in AMP-5 corrects AMP-3 overestimates quite well. There

6 4768 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 49, NO. 12, DECEMBER 2011 appears to be only limited room for further improvement in the 222-mm/y rms AMP-5 accumulation performance because the GPCP 190-mm/y rms discrepancies were artificially reduced by the degree to which GPCP incorporates GPCC gauge data that include some of the same gauges being used here for comparison; the reliance of GPCP upon GPCC data increases when the GPCP satellite-only data are sparse or uncertain, as they are in arid areas. The effect of this data inbreeding can arguably be seen in the GPCP scatter plot in Fig. 2 for gauges reporting < 500 mm/y and in the artificially high correlation coefficient (0.89) between GPCP and those gauges reporting < 900 mm/y relative to those reporting more precipitation (0.82). For those gauges reporting more than 900 mm/y, the correlation coefficients for AMP-5 and GPCP are about the same (0.85 versus 0.82, respectively); for such gauges, evaporation is less important since high accumulation is correlated with high humidity and reduced evaporation. Moreover, part of this AMP-5/GPCP performance gap can be remedied by using more AMSU and Special Sensor Microwave Imager/Sounder (SSMIS) [24], [25] satellites to capture an even larger fraction of any brief intense precipitation events. Fig. 3 shows the global maps of two-year mean annual precipitation discrepancies (estimate gauge, in millimeters per year) between 509 gauges and AMP-3, AMP-4, AMP-5, and GPCP. AMP-4 and AMP-5 provide much improvement over AMP-3, and AMP-5 particularly improves the results over the central U.S. and southeast Australia. Some sites pose special problems for AMP such as the Australian dry lake; these can be individually addressed in the future. The somewhat regional character of these biases suggests that they may have geophysical origins that future gauge and algorithm adjustments can reduce. Fig. 4 shows the same data as those in Fig. 3 but with the fractional errors of the two-year mean annual precipitation, (estimate gauge)/(gauge ), for AMP-3, AMP-4, AMP-5, and GPCP for 509 gauges. The small number, 0.001, was added to the denominator when computing the fractional errors in order to avoid singularity. Fig. 4 makes evident that the ratio overestimates by AMP-5 and GPCP are mostly associated with outlier gauges in desert areas where the gauge values are very low and the ratios are very high, making accurate satellite retrievals very difficult. Also, the overestimates evident in Fig. 3 for AMP-5 over the U.S. Great Plains and for GPCP over northern Europe become relatively small when re-expressed as ratios in Fig. 4. The numbers of gauges out of 509 for which the magnitude of the fractional errors exceeds 0.5 are 171, 80, 59, and 56 for AMP-3, AMP-4, AMP-5, and GPCP, respectively. Fig. 5 shows a global map of the AMP 5 fractional error GPCP fractional error for 509 gauges. Figs. 3 5 show that, over most of the globe, the AMP-5 annual accumulation retrievals are comparable to the gauges and to GPCP, the main exceptions being North Africa and the Middle East. Fig. 5 also shows that, relative to GPCP, AMP-5 tends to underestimate slightly more at the northernmost gauges where some AMSU soundings are flagged as too cold and not accumulated and to scatter more in the desert areas, with more Fig. 3. Global maps of two-year mean annual precipitation error (estimate gauge) for AMP-3, AMP-4, AMP-5, and GPCP for 509 gauges. overestimates than underestimates. The relative stability of GPCP in these dry border areas is perhaps due to its increased use of gauge data there, embedded in GPCC. VII. SUMMARY AND CONCLUSION Two methods for near-surface evaporation correction have been developed for microwave precipitation sensors and have been incorporated in the AMP-4 and AMP-5 retrieval algorithms. AMP-4 employed evaporation information obtained from surface classification alone, whereas average annual relative humidity profiles were also used in AMP-5 and merged by means of an NN. Both methods greatly reduce retrieval errors due to evaporation below the remotely sensed hydrometeors. AMP-5 yields annual accumulation estimates superior to AMP-4

7 SURUSSAVADEE AND STAELIN: EVAPORATION CORRECTION METHODS FOR MICROWAVE RETRIEVALS OF SURFACE PRECIPITATION RATE 4769 and approaches the performance of GPCP, which incorporates data from many more satellite sensors and some overlapping rain gauge data via GPCC. An important next step in improving AMP-5 would involve the use of relative humidity profile information that is more nearly concurrent with the sounding of interest rather than relying on annual averages within a 1 box. An obvious modification to AMP-5 would simply be to replace the annual average relative humidity profile with the corresponding profile taken from an operational NWP forecast for that time and location. One difficulty with this approach is that there is little reliable global training or validation rain gauge data on the spatial and 5 15-min time scales of convective precipitation. Global validation of AMP-5 or other retrievals over ocean is also problematic. Ground-based radar data suffer similar evaporation challenges except in well-instrumented but geographically limited cases, so it unfortunately cannot replace rain gauges for globally training or validating evaporation corrections. Highperformance satellite radar should provide better insights globally, as should well-tuned NWP models. Rain evaporation correction factors can be derived for other microwave instruments observing precipitation by using the methods demonstrated here. Such corrections would differ among instruments such as radar or passive microwave conically scanning spectrometers, depending on their sensitivities to various types of hydrometeors at various altitudes. ACKNOWLEDGMENT Fig. 4. Global maps of the fractional errors of the two-year mean annual precipitation, (estimate gauge)/(gauge ), for AMP-3, AMP-4, AMP-5, and GPCP for 509 gauges. The authors would like to thank the Pennsylvania State University and the University Corporation for Atmospheric Research for providing them with the fifth-generation National Center for Atmospheric Research/Penn State Mesoscale Model and technical support, the Alliance for Computational Earth Science at Massachusetts Institute of Technology for assistance with computer resources, P. W. Rosenkranz for his forward radiance program, TBSCAT, and helpful discussions, G. J. Huffman and D. T. Bolvin for the help in understanding the Global Precipitation Climatology Project data, L. von Bosau for the help with the rain gauge data, and A. Graumann, C. Nichols, and L. Zhao of National Oceanic and Atmospheric Administration/ National Environmental Satellite, Data, and Information Service for the help with Advanced Microwave Sounding Unit data. Fig. 5. Global maps of the AMP 5 fractional error GPCP fractional error for 509 gauges, where the fractional error is (estimate gauge)/(gauge ). REFERENCES [1] C. Surussavadee and D. H. Staelin, Comparison of AMSU millimeterwave satellite observations, MM5/TBSCAT predicted radiances, and electromagnetic models for hydrometeors, IEEE Trans. Geosci. Remote Sens., vol. 44, no. 10, pp , Oct [2] C. Surussavadee and D. H. Staelin, Millimeter-wave precipitation retrievals and observed-versus-simulated radiance distributions: Sensitivity to assumptions, J. Atmos. Sci., vol.64,no.11,pp ,Nov [3] C. Surussavadee and D. H. Staelin, Global millimeter-wave precipitation retrievals trained with a cloud-resolving numerical weather prediction model, part I: Retrieval design, IEEE Trans. Geosci. Remote Sens., vol. 46, no. 1, pp , Jan [4] C. Surussavadee and D. H. Staelin, Global millimeter-wave precipitation retrievals trained with a cloud-resolving numerical weather prediction model, part II: Performance evaluation, IEEE Trans. Geosci. Remote Sens., vol. 46, no. 1, pp , Jan

8 4770 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 49, NO. 12, DECEMBER 2011 [5] C. Surussavadee and D. H. Staelin, Satellite retrievals of arctic and equatorial rain and snowfall rates using millimeter wavelengths, IEEE Trans. Geosci. Remote Sens., vol. 47, no. 11, pp , Nov [6] C. Surussavadee and D. H. Staelin, Global precipitation retrievals using the NOAA/AMSU millimeter-wave channels: Comparison with rain gauges, J. Appl. Meteorol. Climatol., vol. 49, no. 1, pp , Jan [7] NOAA, Cited 2007: Monthly Climatic Data for the World. [Online]. Available: [8] G. J. Huffman, R. F. Adler, M. M. Morrissey, D. T. Bolvin, S. Curtis, R. Joyce, B. McGavock, and J. Susskind, Global precipitation at one degree daily resolution from multi-satellite observations, J. Hydrometeor., vol. 2, no. 1, pp , Feb [9] CISL Data Support Section Nat. Center Atmos. Res., Dataset ds083.2 Updated Daily: NCEP FNL Operational Model Global Tropospheric Analyses, Continuing From July 1999, Boulder, CO. [Online]. Available: Dataset ds083.2 [10] J. Dudhia, D. Gil, K. Manning, W. Wang, and C. Bruyere, PSU/NCAR Mesoscale Modeling System Tutorial Class Notes and Users Guide (MM5 Modeling System Version 3)2005, Jan.. [Online]. Available: [11] P. W. Rosenkranz, Radiative transfer solution using initial values in a scattering and absorbing atmosphere with surface reflection, IEEE Trans. Geosci. Remote Sens., vol. 40, no. 8, pp , Aug [12] 2B-GEOPROF2007, Dec. 5. [Online]. Available: cira.colostate.edu/ [13] M. C. Hansen, R. S. DeFries, J. R. G. Townshend, and R. Sohlberg, Global land cover classification at 1 km spatial resolution using a classification tree approach, Int. J. Remote Sens., vol. 21, no. 6/7, pp , [14] NASA, Cited 2009: 2006 Annual Mean Aerosol Optical Depth at 550 nm From MODIS. [Online]. Available: viewimage.php?id5199 [15] R. F. Adler, G. J. Huffman, A. Chang, R. Ferraro, P. Xie, J. Janowiak, B. Rudolf, U. Schneider, S. Curtis, D. Bolvin, A. Gruber, J. Susskind, P. Arkin, and E. Nelkin, The version 2 Global Precipitation Climatology Project (GPCP) monthly precipitation analysis (1979 present), J. Hydrometeor., vol. 4, no. 6, pp , Dec [16] G. J. Huffman, R. F. Adler, D. T. Bolvin, and G. Gu, Improving the global precipitation record: GPCP version 2.1, Geophys. Res. Lett., vol. 36, p. L17 808, doi: /2009gl [17] B. Rudolf, Management and analysis of precipitation data on a routine basis, in Proc. Int. WMO/IAHS/ETH Symp. Precipitation Evaporation, 1993, pp [18] T. Wilheit, A. Chang, and L. Chiu, Retrieval of monthly rainfall indices from microwave radiometric measurements using probability distribution functions, J. Atmos. Ocean. Technol., vol. 8, no. 1, pp , Feb [19] N. C. Grody, Classification of snow cover and precipitation using the Special Sensor Microwave/Imager (SSM/I), J. Geophys. Res., vol. 96, pp , [20] G. J. Huffman, R. F. Adler, M. M. Morrissey, D. T. Bolvin, S. Curtis, R. Joyce, B. McGavock, and J. Susskind, Global precipitation at onedegree daily resolution from multisatellite observations, J. Hydrometeor., vol. 2, no. 1, pp , Oct [21] J. Susskind, P. Piraino, L. Rokke, L. Iredell, and A. Mehta, Characteristics of the TOVS pathfinder path A dataset, Bull. Amer. Meteorol. Soc., vol. 78, no. 7, pp , Jul [22] P. Xie and P. A. Arkin, Global monthly precipitation estimates from satellite-observed outgoing longwave radiation, J. Clim., vol. 11, no. 2, pp , Feb [23] 85 pp D. R. Legates, A Climatology of Global Precipitation, vol. 40, Newark, DE1987, 85 pp. [24] N. Sun and F. Weng, Evaluation of Special Sensor Microwave Imager/Sounder (SSMIS) environmental data records, IEEE Trans. Geosci. Remote Sens., vol. 46, no. 4, pp , Apr [25] D. B. Kunkee, G. A. Poe, D. J. Boucher, S. D. Swadley, Y. Hong, J. E. Wessel, and E. A. Uliana, Design and evaluation of the first special sensor microwave imager/sounder, IEEE Trans. Geosci. Remote Sens., vol. 46, no. 4, pp , Apr Chinnawat Surussavadee (S 04 M 07) received the B.Eng. degree in electrical engineering from the King Mongkut s Institute of Technology at Ladkrabang, Bangkok, Thailand, in 1999, the M.S. degree in electric power engineering from the Rensselaer Polytechnic Institute, Troy, NY, in 2001, and the Ph.D. degree in electrical engineering from the Massachusetts Institute of Technology (MIT), Cambridge, in From 2004 to 2006, he was a Research Assistant with the Remote Sensing and Estimation Group, Research Laboratory of Electronics (RLE), MIT. From 2006 to 2007, he was a Lecturer with the Phuket Rajabhat University, Phuket, Thailand. Since 2006, he has been a Research Affiliate with RLE. He joined the Faculty of Technology and Environment, Prince of Songkla University, Phuket Campus, Phuket, Thailand, in 2007, where he is currently an Assistant Professor of electrical engineering and is also the Head of the Andaman Environment and Natural Disaster Research Center. David H. Staelin (S 59 M 65 SM 75 F 79 LF 04) received the B.S., M.S., and D.Sc. degrees in electrical engineering from the Massachusetts Institute of Technology (MIT), Cambridge, in 1960, 1961, and 1965, respectively. He was an Assistant Director with the Lincoln Laboratory, MIT, from 1990 to He joined the MIT faculty in 1965, where he is currently a Professor of electrical engineering and teaches electromagnetics and signal processing. He was the Principal Investigator for the Nimbus-E Microwave Spectrometer and Scanning Microwave Spectrometer experiments on the National Aeronautics and Space Administration (NASA) Nimbus-5 and Nimbus-6 satellites and a Coinvestigator for the NASA experiments Atmospheric Infrared Sounder/Advanced Microwave Sounding Unit/Humidity Sounder for Brazil on Aqua, Scanning Multichannel Microwave Radiometer on Nimbus 7, and Planetary Radio Astronomy on Voyager 1 and 2. He was a member of the U.S. National Polar-orbiting Operational Environmental Satellite System Sounder Operational Algorithm Team and the NASA science teams for Precipitation Measurement Missions and the U.S. NPOESS Preparatory Program.