hyperspectral microwave atmospheric sounding · sensing modality for atmospheric sounding, driven...

15
128 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 49, NO. 1, JANUARY 2011 Hyperspectral Microwave Atmospheric Sounding William J. Blackwell, Senior Member, IEEE, Laura J. Bickmeier, R. Vincent Leslie, Michael L. Pieper, Jenna E. Samra, Chinnawat Surussavadee, Member, IEEE, and Carolyn A. Upham Abstract—We introduce a new hyperspectral microwave remote sensing modality for atmospheric sounding, driven by recent ad- vances in microwave device technology that now permit receiver arrays that can multiplex multiple broad frequency bands into more than 100 spectral channels, thus improving both the vertical and horizontal resolutions of the retrieved atmospheric profile. Global simulation studies over ocean and land in clear and cloudy atmospheres using three different atmospheric profile databases are presented that assess the temperature, moisture, and pre- cipitation sounding capability of several notional hyperspectral systems with channels sampled near the 50–60-, 118.75-, and 183.31-GHz absorption lines. These analyses demonstrate that hyperspectral microwave operation using frequency multiplexing techniques substantially improves temperature and moisture pro- filing accuracy, particularly in atmospheres that challenge con- ventional nonhyperspectral microwave sounding systems because of high water vapor and cloud liquid water content. Retrieval performance studies are also included that compare hyperspectral microwave sounding performance to conventional microwave and hyperspectral infrared approaches, both in a geostationary and a low-Earth-orbit context, and a path forward to a new generation of high-performance all-weather sounding is discussed. Index Terms—Advanced Microwave Sounding Unit (AMSU), Advanced Technology Microwave Sounder (ATMS), data assimi- lation, geostationary, Geostationary Millimeter-wave Array Spec- trometer (GeoMAS), Geostationary Operational Environmental Satellite, humidity, hyperspectral, Hyperspectral Microwave Array Spectrometer (HyMAS), infrared, microwave, millimeter wave, moisture, National Polar-orbiting Operational Environmen- tal Satellite System (NPOESS), neural network (NN), precipita- tion, retrieval, sounding, temperature, water vapor. I. I NTRODUCTION R EMOTE measurements of the Earth’s atmospheric state using microwave and infrared wavelengths have been carried out for many years [1], [2]. Physical considerations involving the use of these spectral regions include the relatively Manuscript received August 4, 2009; revised December 7, 2009 and March 27, 2010; accepted May 31, 2010. Date of publication August 9, 2010; date of current version December 27, 2010. This work was supported in part by the Department of the Air Force under Air Force Contract FA8721-05-C-0002 and in part by the Lincoln Laboratory Advanced Concepts Committee of the Massachusetts Institute of Technology. Opinions, interpretations, conclusions, and recommendations are those of the authors and are not necessarily endorsed by the U.S. Government. W. J. Blackwell, L. J. Bickmeier, R. V. Leslie, M. L. Pieper, J. E. Samra, and C. A. Upham are with the Lincoln Laboratory, Massachusetts Institute of Technology, Lexington, MA 02420-9108 USA (e-mail: [email protected]). C. Surussavadee is with the Research Laboratory of Electronics, Massachusetts Institute of Technology, Cambridge, MA 02139 USA, and also with the Andaman Environment and Natural Disaster Research Center, Faculty of Technology and Environment, Prince of Songkla University, Phuket 83120, Thailand. Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TGRS.2010.2052260 high cloud-penetrating capability at microwave wavelengths and the relatively sharp weighting functions at infrared wave- lengths, particularly in the short-wave region near 4 μm where Planck nonlinearity further increases temperature sensitivity. Infrared spectrometer technology has advanced markedly over the last 15 years or so to allow the simultaneous spectral sam- pling of thousands of bands spaced along narrow atmospheric absorption features [3]. The Atmospheric InfraRed Sounder (AIRS), launched in May 2002, measures 2378 channels from 3.7 to 15.4 μm [4], and the Infrared Atmospheric Sounding Interferometer, launched in 2006, measures 8461 channels from 3.6 to 15.5 μm [5]. These sensors, and similar sensors to be launched as part of the National Polar-orbiting Operational Environmental Satellite System (NPOESS) and Meteosat Third Generation systems, substantially improve atmospheric sound- ing through the use of hyperspectral measurements, which yield greater vertical resolution throughout the atmosphere [6]. In this paper, we explore the potential use of hyperspectral microwave observations and make three primary contributions. First, we propose a frequency multiplexing technique that can be used to realize hyperspectral microwave measurements with conventional receiver hardware. The approach is easily scal- able and requires no new technology development. Second, we examine the relative merits of increased bandwidth versus increased channelization within an increased bandwidth and demonstrate that the optimal operating point depends on the available integration time. We also show that “multiplexed hyperspectral” operation always exceeds “sequential hyper- spectral” operation (where a fixed intermediate-frequency (IF) filter bank is swept across the measurement spectrum) for finite bandwidths and integration times. Third, we provide a set of comprehensive and global simulation analyses with state-of-the- art retrieval methods that illuminate many of the principal dimen- sions of the design and performance comparison trade space, including 60 GHz versus 118 GHz, hyperspectral microwave versus conventional microwave, and hyperspectral microwave versus hyperspectral infrared. We also consider both geostation- ary and low-Earth-orbit configurations. Note that these areas are certainly not mutually exclusive, and an optimized sounding system will likely contain several of these components. This paper is organized as follows. The hyperspectral mi- crowave concept is first introduced, and it is demonstrated that multiple receiver arrays can be used to multiplex a large set of channels onto a single spot on the ground. An overview of the data sets and physical models used in the simulation is provided. We next point out that opacity due to water vapor continuum absorption is a fundamental limitation of conven- tional millimeter-wave sounding and show how a hyperspec- tral millimeter-wave approach can be used to overcome this 0196-2892/$26.00 © 2010 IEEE

Upload: others

Post on 21-Oct-2019

4 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Hyperspectral Microwave Atmospheric Sounding · sensing modality for atmospheric sounding, driven by recent ad-vances in microwave device technology that now permit receiver arrays

128 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 49, NO. 1, JANUARY 2011

Hyperspectral Microwave Atmospheric SoundingWilliam J. Blackwell, Senior Member, IEEE, Laura J. Bickmeier, R. Vincent Leslie, Michael L. Pieper,

Jenna E. Samra, Chinnawat Surussavadee, Member, IEEE, and Carolyn A. Upham

Abstract—We introduce a new hyperspectral microwave remotesensing modality for atmospheric sounding, driven by recent ad-vances in microwave device technology that now permit receiverarrays that can multiplex multiple broad frequency bands intomore than 100 spectral channels, thus improving both the verticaland horizontal resolutions of the retrieved atmospheric profile.Global simulation studies over ocean and land in clear and cloudyatmospheres using three different atmospheric profile databasesare presented that assess the temperature, moisture, and pre-cipitation sounding capability of several notional hyperspectralsystems with channels sampled near the 50–60-, 118.75-, and183.31-GHz absorption lines. These analyses demonstrate thathyperspectral microwave operation using frequency multiplexingtechniques substantially improves temperature and moisture pro-filing accuracy, particularly in atmospheres that challenge con-ventional nonhyperspectral microwave sounding systems becauseof high water vapor and cloud liquid water content. Retrievalperformance studies are also included that compare hyperspectralmicrowave sounding performance to conventional microwave andhyperspectral infrared approaches, both in a geostationary and alow-Earth-orbit context, and a path forward to a new generationof high-performance all-weather sounding is discussed.

Index Terms—Advanced Microwave Sounding Unit (AMSU),Advanced Technology Microwave Sounder (ATMS), data assimi-lation, geostationary, Geostationary Millimeter-wave Array Spec-trometer (GeoMAS), Geostationary Operational EnvironmentalSatellite, humidity, hyperspectral, Hyperspectral MicrowaveArray Spectrometer (HyMAS), infrared, microwave, millimeterwave, moisture, National Polar-orbiting Operational Environmen-tal Satellite System (NPOESS), neural network (NN), precipita-tion, retrieval, sounding, temperature, water vapor.

I. INTRODUCTION

R EMOTE measurements of the Earth’s atmospheric stateusing microwave and infrared wavelengths have been

carried out for many years [1], [2]. Physical considerationsinvolving the use of these spectral regions include the relatively

Manuscript received August 4, 2009; revised December 7, 2009 andMarch 27, 2010; accepted May 31, 2010. Date of publication August 9, 2010;date of current version December 27, 2010. This work was supported in part bythe Department of the Air Force under Air Force Contract FA8721-05-C-0002and in part by the Lincoln Laboratory Advanced Concepts Committee of theMassachusetts Institute of Technology. Opinions, interpretations, conclusions,and recommendations are those of the authors and are not necessarily endorsedby the U.S. Government.

W. J. Blackwell, L. J. Bickmeier, R. V. Leslie, M. L. Pieper, J. E. Samra,and C. A. Upham are with the Lincoln Laboratory, Massachusetts Institute ofTechnology, Lexington, MA 02420-9108 USA (e-mail: [email protected]).

C. Surussavadee is with the Research Laboratory of Electronics,Massachusetts Institute of Technology, Cambridge, MA 02139 USA, and alsowith the Andaman Environment and Natural Disaster Research Center, Facultyof Technology and Environment, Prince of Songkla University, Phuket 83120,Thailand.

Color versions of one or more of the figures in this paper are available onlineat http://ieeexplore.ieee.org.

Digital Object Identifier 10.1109/TGRS.2010.2052260

high cloud-penetrating capability at microwave wavelengthsand the relatively sharp weighting functions at infrared wave-lengths, particularly in the short-wave region near 4 μm wherePlanck nonlinearity further increases temperature sensitivity.Infrared spectrometer technology has advanced markedly overthe last 15 years or so to allow the simultaneous spectral sam-pling of thousands of bands spaced along narrow atmosphericabsorption features [3]. The Atmospheric InfraRed Sounder(AIRS), launched in May 2002, measures 2378 channels from3.7 to 15.4 μm [4], and the Infrared Atmospheric SoundingInterferometer, launched in 2006, measures 8461 channels from3.6 to 15.5 μm [5]. These sensors, and similar sensors to belaunched as part of the National Polar-orbiting OperationalEnvironmental Satellite System (NPOESS) and Meteosat ThirdGeneration systems, substantially improve atmospheric sound-ing through the use of hyperspectral measurements, which yieldgreater vertical resolution throughout the atmosphere [6].

In this paper, we explore the potential use of hyperspectralmicrowave observations and make three primary contributions.First, we propose a frequency multiplexing technique that canbe used to realize hyperspectral microwave measurements withconventional receiver hardware. The approach is easily scal-able and requires no new technology development. Second,we examine the relative merits of increased bandwidth versusincreased channelization within an increased bandwidth anddemonstrate that the optimal operating point depends on theavailable integration time. We also show that “multiplexedhyperspectral” operation always exceeds “sequential hyper-spectral” operation (where a fixed intermediate-frequency (IF)filter bank is swept across the measurement spectrum) for finitebandwidths and integration times. Third, we provide a set ofcomprehensive and global simulation analyses with state-of-the-art retrieval methods that illuminate many of the principal dimen-sions of the design and performance comparison trade space,including 60 GHz versus 118 GHz, hyperspectral microwaveversus conventional microwave, and hyperspectral microwaveversus hyperspectral infrared. We also consider both geostation-ary and low-Earth-orbit configurations. Note that these areasare certainly not mutually exclusive, and an optimized soundingsystem will likely contain several of these components.

This paper is organized as follows. The hyperspectral mi-crowave concept is first introduced, and it is demonstrated thatmultiple receiver arrays can be used to multiplex a large setof channels onto a single spot on the ground. An overviewof the data sets and physical models used in the simulation isprovided. We next point out that opacity due to water vaporcontinuum absorption is a fundamental limitation of conven-tional millimeter-wave sounding and show how a hyperspec-tral millimeter-wave approach can be used to overcome this

0196-2892/$26.00 © 2010 IEEE

Page 2: Hyperspectral Microwave Atmospheric Sounding · sensing modality for atmospheric sounding, driven by recent ad-vances in microwave device technology that now permit receiver arrays

BLACKWELL et al.: HYPERSPECTRAL MICROWAVE ATMOSPHERIC SOUNDING 129

limitation. All retrievals presented in this paper are carried outusing neural networks (NNs) trained against cloud-resolvingphysical models, and we discuss the methodologies used toinitialize, train, and evaluate the NN retrievals. We then define avariety of notional systems and compare the hyperspectral mi-crowave approach with current and planned approaches that usemicrowave and hyperspectral infrared observations separatelyand in combination. Temperature, water vapor, and precipi-tation retrieval performance comparisons are then presented,and the impact of correlated error sources on performance isexamined. Finally, we summarize and provide suggestions forfurther research and development.

II. HYPERSPECTRAL MICROWAVE CONCEPT

A spate of recent technology advances driven in part bythe gigabit wireless communications industry [7], the semicon-ductor industry [8], and the National Aeronautics and SpaceAdministration (NASA) Earth Science Technology Office [9]has significantly and profoundly changed the landscape ofmodern radiometry by enabling miniaturized, low-power, andlow-noise radio-frequency receivers operating at frequenciesup to 200 GHz. These advances enable the practical use ofreceiver arrays to multiplex multiple broad frequency bandsinto many spectral channels, and we explore the atmosphericsounding benefit of such systems in this paper. We use the term“hyperspectral microwave” to refer generically to microwavesounding systems with approximately 100 spectral channelsor more. In the infrared wavelength range, the term “hyper-spectral” is used to denote the resolution of individual narrowabsorption features which are abundant throughout the infraredspectrum. In the microwave and millimeter wavelength range,however, there are substantially fewer spectral features, and thespectral widths are typically broad, and an alternate definitionis therefore appropriate.

We begin with an analysis of geostationary sounding sys-tems, and we consider low-Earth-orbiting systems later in thispaper. Detailed studies of the geophysical products that couldbe derived from a geostationary microwave sensor and theradiometric requirements of such a sensor date back many years[10], [11]. The persistent observations afforded by a geosta-tionary platform would allow temporal sampling over most ofthe viewable Earth hemisphere on timescales of approximately15 min. The capability to sound in and around storms wouldsignificantly improve both regional and global numericalweather prediction models. The tropospheric information con-tent of infrared observations, however, is compromised byclouds which attenuate the radiance to space from the at-mosphere below the cloud level. The ability to model andforecast hurricanes would greatly improve with microwavemeasurements from a geostationary orbit.

A simple example of a multiplexed hyperspectral microwavesystem is shown in Fig. 1 with eight instantaneous fields ofviews (IFOVs), i.e., four near 118 GHz and four near 183 GHz.Each IFOV is sampled by a single feedhorn measuring twoorthogonal polarizations (vertical and horizontal, for example)that are each fed to a ten-channel spectrometer. The cross-hatching in the figure indicates slightly different frequency

Fig. 1. Instantaneous beam pattern (3-dB contours) on the Earth for a notionalmultiplexed hyperspectral millimeter-wave array spectrometer.

bands. The 183-GHz IFOVs each measure the same spectralchannels. As the array is microscanned in 50-km (or finer) stepsin two dimensions, each 50-km spot on the ground is eventu-ally sampled by 80 channels near 118 GHz, and each 25-kmspot on the ground is eventually sampled by 20 channels near183 GHz for a total of 100 channels. Additional channelscould be added by increasing the number of feeds and receiverbanks, and hyperspectral microwave systems with hundredsof channels are therefore reasonable. Note that further chan-nelization within a given receiver bandwidth quickly reachesa point of diminishing returns due to the increase in thermalnoise as 1/

√B, where B is the bandwidth for a single channel.

Detailed analyses by other investigators (see [12], for example)demonstrate that a division into about eight to ten channels perreceiver yields near-optimal results.

We further illustrate the spectral multiplexing concept usingtemperature weighting functions. The temperature weightingfunction, given by the derivative of the transmittance functionwith respect to altitude, characterizes the degree to which eachatmospheric layer contributes to the radiances viewed fromspace at the indicated frequencies. The weighting functionsapproach zero at high altitudes, where the atmosphere becomestransparent, or at low altitudes, where the overlying atmosphereis so thick as to be fully opaque. The temperature weightingfunctions at nadir incidence for a nominal eight-channel re-ceiver operating in a 5-GHz bandwidth on the low-frequencyside of the 118.75-GHz oxygen line are shown in Fig. 2. The1976 U.S. Standard Atmosphere over a nonreflective surfacewas used in the calculations. The channels are equally spacedin frequency, and the bandwidth of each channel is 560 MHz.

A hyperspectral microwave system can be constructed byusing multiple receiver banks with a replicated, but frequency-shifted, version of the template channel set used to generateFig. 2. For example, eight receiver banks with eight channelseach could be used to create a 64-channel system by progres-sively shifting the IF band of each receiver by 70 MHz.

We conclude by comparing the temperature weightingfunctions of the 64-channel multiplexed system to that of a

Page 3: Hyperspectral Microwave Atmospheric Sounding · sensing modality for atmospheric sounding, driven by recent ad-vances in microwave device technology that now permit receiver arrays

130 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 49, NO. 1, JANUARY 2011

Fig. 2. Template temperature weighting functions for channels near118.75 GHz at nadir incidence. The 1976 U.S. Standard Atmosphere over anonreflective surface was used in the calculations.

Fig. 3. Weighting functions for 64 channels near 118.75 GHz and the eightATMS tropospheric temperature sounding channels near 50–57 GHz at nadirincidence. The 1976 U.S. Standard Atmosphere over a nonreflective surfacewas assumed in the calculations.

current-generation microwave sounder, the Advanced Technol-ogy Microwave Sounder (ATMS) [13] scheduled to fly onboardthe NPOESS Preparatory Project in 2011. In the upper panel ofFig. 3, the set of weighting functions for the 64-channel systemis indicated by thin lines, and the eight tropospheric ATMSchannels are indicated by heavy lines. The surface weights ofall the channels are shown in the lower panel of Fig. 3. The fre-quencies for the eight ATMS channels shown range from 50.3(most transparent) to 57.29 GHz. The effective atmosphericvertical sampling density of the 64-channel multiplexed systemis clearly superior to that of the conventional eight-channelsystem. Water-vapor-burden weighting functions are not shownhere, as similar arguments apply. The temperature and water va-por profiling advantage resulting from fine vertical atmosphericsampling is discussed in Section VII.

The temperature weighting function for the ATMS 50.3-GHzchannel is shown near the bottom of the top panel of Fig. 3(the temperature weight near the surface is ∼0.06 km−1), andthis channel is significantly more transparent than the most

Fig. 4. Geographical locations of the pixels contained in the MM5 andNOAA88b data sets. Global coverage, including a variety of land and oceanpixels, is achieved by each set.

transparent 118-GHz channel (the temperature weight nearthe surface is ∼0.09 km−1) due to water vapor continuumabsorption, but the weighting function has a comparable shape.Water vapor absorption appears to be a significant handicap formillimeter-wave sounding [10]. Hyperspectral millimeter-wavesystems, however, are not so adversely affected because of thevery dense vertical spacing of the weighting functions in theatmospheric boundary layer. We address this important point indetail in Section IV.

III. OVERVIEW OF DATA SETS AND PHYSICAL MODELS

The performance comparisons that we present in this paperare all based on simulated observations derived using physi-cal models and global ensembles of atmospheric states. Theselection of the ensemble of atmospheric states is a criticallyimportant part of any simulation study, and we have taken greatcare to ensure that the profiles included in the analysis aresufficiently representative of the cloudy moist atmospheres thatchallenge most atmospheric sounding systems.

A. NOAA88b Atmospheric Profile Data Set

The NOAA88b radiosonde/rocketsonde data set contains7547 profiles, globally distributed seasonally and geograph-ically (see Fig. 4). Atmospheric temperature, moisture, andozone are given at 100 discrete levels from the surface toaltitudes exceeding 50 km. Skin surface temperature is alsorecorded. NOAA88b water vapor measurements above approxi-mately 10 km are of questionable quality and are not consideredin this paper. Approximately 6500 profiles from the NOAA88bdatabase were selected for inclusion in this study by elimi-nating cases for which the surface pressure was not equal to1000 mbar. The mean and standard deviation of the temperatureand water vapor profiles in the NOAA88b data set are shown inFigs. 5 and 6, respectively.

B. MIT MM5 Precipitation and Atmospheric Profile Data Set

The Massachusetts Institute of Technology (MIT) MM5precipitation and atmospheric profile data set [14], [15] is com-posed of meteorological parameters for 122 storms simulatedusing the fifth-generation National Center for Atmospheric

Page 4: Hyperspectral Microwave Atmospheric Sounding · sensing modality for atmospheric sounding, driven by recent ad-vances in microwave device technology that now permit receiver arrays

BLACKWELL et al.: HYPERSPECTRAL MICROWAVE ATMOSPHERIC SOUNDING 131

Fig. 5. Mean and standard deviation of the temperature profiles in the MM5and NOAA88b data sets.

Fig. 6. Mean and standard deviation of the water vapor profiles in the MM5and NOAA88b data sets.

Research/The Pennsylvania State University mesoscale model(MM5). These storms are globally distributed, as shown inFig. 4, and span a year. Each storm has 190 × 190 pictureelements (pixels) spaced on a rectangular 5-km grid with 42pressure levels. About 46% of the 4.4 million pixels are precip-itating with nonzero rain water or snow at 1000 mbar [14]. Thispercentage of precipitating pixels is high because only verycloudy regions were sampled. The validity of this ensembleof storms has been shown by the statistical agreement betweenits simulated brightness temperatures and those coincidentallyobserved by the Advanced Microwave Sounding Unit (AMSU)instruments aboard the NOAA-15, NOAA-16, and NOAA-17satellites [14], [15]. All radiative transfer calculations wereperformed at the native horizontal spatial resolution of theMM5 data and were then convolved with the appropriateGeostationary Millimeter-wave Array Spectrometer (GeoMAS)spatial response functions.

Approximately 50 000 nonprecipitating pixels selected fromthe 122 storm cases were used in the temperature and mois-ture profile retrieval study; approximately half of these pixelswere cloudy (nonzero integrated cloud liquid water content).

Fig. 7. Histogram of the vertically integrated cloud liquid water content (i.e.,liquid water path) in the MM5 storm data set. The NOAA88b data set doesnot include measurements of cloud liquid water. Rain becomes likely when theintegrated cloud liquid water content exceeds approximately 0.2 mm [16], andprecipitating pixels have been removed from the data set.

Precipitation was screened by requiring that the sum of verti-cally averaged water mixing ratios in the form of rain, graupel,and snow was less than 0.01 kg/kg. The pixels were spatiallydownsampled to ensure that the distance between any twopixels always exceeds 25 km. The mean and standard deviationof the temperature and water vapor profiles in the MM5 dataset are shown in Figs. 5 and 6, respectively. The histogramof integrated cloud liquid water (liquid water path) is shownin Fig. 7.

The MM5 profiles are generally warmer and contain morewater vapor and cloud liquid water than the NOAA88b profilesand are highly representative of the atmospheric thermody-namic state that would be encountered near precipitation. TheNOAA88b profiles are characterized by a high level of variabil-ity due to the deliberate inclusion of more extreme cases.

C. Microwave/Millimeter-Wave Nonscattering RadiativeTransfer Model: TBARRAY

Simulated brightness temperature observations for at-mospheric profiles in the NOAA88b data set were calcu-lated using the TBARRAY software package of Rosenkranz[17]. TBARRAY is a line-by-line routine based on the Liebemillimeter-wave propagation model [18], [19]. Scattering wasnot modeled because cloud liquid water content was notrecorded in the NOAA88b data set. All radiative transfer calcu-lations for the temperature and water vapor retrieval simulationswere performed at a single angle at nadir incidence.

D. Microwave/Millimeter-Wave Scattering RadiativeTransfer Model: TBSCAT

Simulated brightness temperature observations for atmo-spheric profiles in the MIT MM5 data set were calculatedusing the TBSCAT software package of Rosenkranz [20].TBSCAT is a multistream initial-value radiative transfer routinethat includes both absorption and scattering. The collection of

Page 5: Hyperspectral Microwave Atmospheric Sounding · sensing modality for atmospheric sounding, driven by recent ad-vances in microwave device technology that now permit receiver arrays

132 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 49, NO. 1, JANUARY 2011

streams describes a system of coupled first-order differentialequations, and TBSCAT approaches the solution as an initial-value problem starting from the top of the atmosphere. The so-lution uses the backward Euler method of finite differences. Theabsorption coefficients are calculated identically in TBARRAYand TBSCAT. The scattering calculations comprise the Miecoefficients, the inverse-exponential drop-size distribution (forprecipitating cases only), the Liebe/Hufford permittivity model,and the Henyey–Greenstein phase function. The scattering co-efficients are calculated using the Deirmendjian implementa-tion in the Mie region and the Wiscombe implementation inthe Rayleigh region. The clouds in the nonprecipitating MM5data set were simulated using ten streams and with both thecloud liquid water and cloud ice as scattering hydrometeors.The cloud liquid water was given a radius of 0.02 mm, and thecloud ice was given a radius of 0.06 mm, as these values areconsistent with recommendations from other investigators [21].

E. Ocean Surface Emissivity Model: FASTEM and FASTEM2

English and Hewison [22] developed the FASTEM model,which parameterizes an “effective” ocean surface emissivityfor frequencies between 10 and 220 GHz for Earth incidenceangles less than 60◦ and for oceanic surface wind speeds lessthan 20 m/s. FASTEM2, an updated version of FASTEM, usesan approach that is similar to that of Petty and Katsaros [23]to compute the surface emissivity. FASTEM and FASTEM2both incorporate geometric optics, Bragg scattering, and foameffects. FASTEM was used for the simulations of the precipi-tating cases, and FASTEM2 (with the optical depth option setto zero) was used for the simulations of the nonprecipitatingcases. FASTEM and FASTEM2 calculations used the MM5oceanic wind speed at an altitude of 10 m, which ranged from0.14 to 11 m/s. The oceanic surface wind speed is not recordedin the NOAA88b data set, and the FASTEM2 wind speedinput for these cases was therefore randomized using a uniformdistribution between 0.5 and 10 m/s.

F. Land Surface Emissivity Model

Land surface emissivity values were assigned randomly us-ing a uniform distribution between 0.8 and 1.0. The sameemissivity value was used for all frequencies. Recent work hasshown that this simple model is fairly representative of mostnaturally occurring land emissivities [24], although improve-ments are planned in future work.

IV. ATMOSPHERIC TRANSMITTANCE AT MILLIMETER

WAVELENGTHS: IMPLICATIONS FOR

GEOSTATIONARY SOUNDING

Atmospheric extinction increases as ∼f4 due to Rayleighscattering and as ∼f2 due to absorption. Previous workshave demonstrated that the relatively high sensitivity of themillimeter-wave bands to hydrometeor scattering can be usedto improve precipitation sensing [25]–[27]. However, therelatively high levels of atmospheric absorption present inmillimeter-wave bands can hinder atmospheric sounding in the

Fig. 8. Microwave/millimeter-wave absorption spectrum. Two calculationsfor the percent transmission (nadir view) using the 1976 U.S. StandardAtmosphere are shown, i.e., one assuming no water vapor and one assuming1.5 g/cm2 (15 mm).

Fig. 9. Atmospheric transmittance at nadir incidence at 90 GHz as calculatedfrom the MIT MM5 data set. The transmittance values are averaged over binsof integrated water vapor and integrated cloud liquid water content.

boundary layer near the surface [10]. In this section, we explorethis apparent handicap in the context of geostationary sounding.

The microwave/millimeter-wave absorption spectrum isshown in Fig. 8 for a fixed amount of water vapor (15 mm)and no cloud liquid water. The characteristic decrease in trans-mittance with frequency is immediately apparent. The watervapor content in the 1976 U.S. Standard Atmosphere used tocompute the absorption spectrum shown in Fig. 8 is relativelylow compared to a typical tropical atmosphere (∼50-mm watervapor). Therefore, we now examine atmospheric transmittancewith a focus on profiles from the MIT MM5 data set, whichis characterized by high water content. The atmospheric trans-mittance at nadir incidence at 90 GHz is shown in Fig. 9as a function of the integrated water vapor content and theintegrated cloud liquid water content, and the transmittancepredictably decreases with increasing water content. To explorethe frequency and water content dependence in tandem, we

Page 6: Hyperspectral Microwave Atmospheric Sounding · sensing modality for atmospheric sounding, driven by recent ad-vances in microwave device technology that now permit receiver arrays

BLACKWELL et al.: HYPERSPECTRAL MICROWAVE ATMOSPHERIC SOUNDING 133

Fig. 10. Atmospheric transmittance at nadir incidence at 50 GHz minus thatat 90 GHz as calculated from the MIT MM5 data set. The transmittance valuesare averaged over bins of integrated water vapor and integrated cloud liquidwater content.

plot the difference between the transmittances at 50 GHz andat 90 GHz as a function of water content in Fig. 10. For lowwater contents, the transmittance at 90 GHz exceeds that at50 GHz, although the opposite is true for relatively high watercontents. The latter observation is a fundamental limitationof conventional (nonhyperspectral) millimeter-wave sounding.Hyperspectral millimeter-wave systems, however, are able toinfer and correct for the increasing atmospheric absorption,even in atmospheres with high water content, because of thevery dense vertical spacing of the weighting functions in the at-mospheric boundary layer. This important distinction betweenconventional and hyperspectral sounding will be highlighted inSection VII.

V. HYPERSPECTRAL MICROWAVE PHYSICAL

RETRIEVALS USING NNs

Recent works have demonstrated the utility of atmosphericprofile retrievals based on feedforward multilayer percep-tron NNs for both hyperspectral infrared and microwaveobservations [28], [29]. The execution time of NN retrievals,once trained using physical models, for example, is typicallyseveral orders of magnitude faster than iterated retrievals whileoffering improved retrieval performance. NNs are used in thiswork to retrieve temperature and moisture profiles and precipi-tation rates. The methodology is now briefly reviewed, and theinterested reader is referred to [30] for more details.

An NN is an interconnection of simple computational el-ements, or nodes, with activation functions that are usuallynonlinear, monotonically increasing, and differentiable. NNsare able to deduce input–output relationships directly from thetraining ensemble without requiring underlying assumptionsabout the distribution of the data. Furthermore, an NN withonly a single hidden layer of a sufficient number of nodes withnonlinear activation functions is capable of approximating anyreal-valued continuous scalar function to a given precision overa finite domain [31].

A multilayer feedforward NN consists of an input layer, anarbitrary number of hidden layers (usually one or two), andan output layer. The hidden layers typically contain sigmoidalactivation functions of the form zj = tanh(aj), where aj =∑d

i=1 wjixi + bj . The output layer is typically linear. Theweights (wji) and biases (bj) for the jth neuron are chosento minimize a cost function over a set of P training patterns. Acommon choice for the cost function is the sum-squared error,defined as

E(w) =12

∑p

∑k

(t(p)k − y

(p)k

)2

(1)

where y(p)k and t

(p)k denote the network outputs and target

responses, respectively, of each output node k given a patternp, and w is a vector containing all the weights and biases ofthe network. The “training” process involves iteratively findingthe weights and biases that minimize the cost function throughsome numerical optimization procedure. Second-order methodsare commonly used to carry out the optimization.

A. Preprocessing With the PPC Transform

Hyperspectral sounding systems typically measure atmo-spheric thermal emission in many spectral bands. The spectralinformation content is often correlated, and a linear preprocess-ing method such as the projected principal components (PPC)transform [28] can be effectively used to reduce the dimen-sionality and filter noise from the measured spectra, even inthe presence of clouds [32]. Furthermore, the PPC transformcan be used to optimally extract spectral radiance informationthat is correlated with a geophysical parameter, such as thetemperature or water vapor profile. The r-rank linear operatorthat captures the most radiance information that is correlated tothe profile is [28]

Lr = ErETr CTR(CRR + CΨΨ)−1 (2)

where Er = [E1|E2| · · · |Er] are the r most significant eigen-vectors of CTR(CRR + CΨΨ)−1CRT , CRR is the spectralradiance covariance, CΨΨ is the noise covariance, and CTR

is the cross-covariance of the atmospheric profile (temperatureor water vapor) and the radiance. The hyperspectral millimeter-wave measurements processed in this work were transformedfrom approximately 100 channels to 25 PPCs for the retrievalof both temperature and water vapor profiles. This factor-of-four reduction in the input dimensionality results in a significantimprovement in both the network training time (typically a fewhours on a 3-GHz Intel Xeon desktop workstation) and thegeneralization ability of the NNs.

B. Network Topologies

All the temperature and moisture retrievals in this workwere implemented using NNs with a single hidden layer of15 sigmoidal nodes and a linear output layer with approximately10 nodes. Approximately five NNs were aggregated (to achieve∼50 total outputs) to estimate the entire profile. Ad hoc attempts

Page 7: Hyperspectral Microwave Atmospheric Sounding · sensing modality for atmospheric sounding, driven by recent ad-vances in microwave device technology that now permit receiver arrays

134 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 49, NO. 1, JANUARY 2011

TABLE IGEOMAS CENTER FREQUENCY OFFSETS FROM 118.75 GHz,

CHANNEL BANDWIDTHS, AND ΔTrms FOR THE MOST

TRANSPARENT RECEIVER BANK (ONE OF EIGHT)

were made to optimize the network topology, and this configu-ration resulted in the best performance. The networks estimatethe atmospheric profile up to 20 km for temperature and up to10 km for water vapor. These estimates and the corresponding“truth” were averaged over 1-km layers prior to the computationof error statistics.

C. Network Initialization, Training, andPerformance Evaluation

The profile data were randomly divided into three nonover-lapping sets that we shall term the training set (80%), thevalidation set (10%), and the testing set (10%). The selectionof profiles for the data sets was identical for all GeoMAS andSynthetic Thinned Aperture Radiometer (STAR) performancecomparisons. The training set was used to derive the networkweights and biases. The network training was stopped if theerror on the validation set did not decrease after ten consecutivetraining epochs or if 300 epochs were reached. Each NN wastrained ten separate times with random initializations, and thevalidation set was used to select the best of the ten networks.All the NN retrieval results presented in this paper were derivedusing the testing set.

All networks were initialized using the Nguyen–Widrowmethod [33] and trained using the Levenberg–Marquardt op-timization procedure [34], [35]. The NETLAB NN softwarepackage [36] was used to train the networks. Random sensornoise (see the last column in Tables I and II) was added to eachsimulated measurement at the beginning of each training epoch.

VI. NOTIONAL SYSTEMS USED FOR COMPARISONS

We now assess the performance of the hyperspectralmillimeter-wave concept in the geostationary context using twonotional systems, i.e., an 88-channel system operating near 118and 183 GHz and a 10-channel system operating near 60 and183 GHz. The underlying physical and radiometric assumptionsfor both systems are identical, as discussed in the following, andidentical NN retrieval algorithms are used. We make no claimsof optimality for either of the two notional systems.

A. GeoMAS

The nominal GeoMAS sensor configuration comprises amodest 88-channel hyperspectral millimeter-wave spectrometer

TABLE IISTAR CENTER FREQUENCIES, CHANNEL BANDWIDTHS, AND ΔTrms.

THE INDIVIDUAL STAR CHANNEL SENSITIVITIES ARE POORER

THAN THOSE FOR COMPARABLE GEOMAS CHANNELS BECAUSE

THE STAR SYSTEM DESCRIBED IN [38] MUST SAMPLE

EACH CHANNEL SEPARATELY, THUS DECREASING

THE AVAILABLE INTEGRATION TIME

(GeoMAS) with 72 channels near the 118.75-GHz oxygen ab-sorption line and 16 channels near the 183.31-GHz water vaporabsorption line. This configuration could easily be realizedusing five dual-polarized antenna horns (sharing a commonreflector) with each feeding a simple nine-channel receiverbank. The spatial resolution of the water vapor channels couldbe doubled by arranging additional feeds in a 2 × 2 arrayconfiguration, thus raising the total number of feedhorns toeight (see Fig. 1). Additional feeds and channels could beadded to further improve the spatial resolution of both bands(additional SNR margin could be used to sharpen the effectiveantenna beam, for example), although we defer this analysis toa future paper.

We assume an integration time of 22.5 ms, which wouldallow the temperature profiles to be retrieved on a 200×200 gridspaced at 50 km (10 000 km × 10 000 km coverage area) in15 min. The water vapor profiles would be retrieved on a 400 ×400 grid spaced at 25 km (10 000 km × 10 000 km coveragearea) in 15 min.

The GeoMAS channel properties for a single receiver bank(of eight) of the 118-GHz temperature band are summarizedin Table I. System temperatures of 650 and 800 K wereassumed for the 118- and 183-GHz systems, respectively.The eight 118-GHz receiver banks (each with eight chan-nels split among a 5-GHz bandwidth near the 118.75-GHzO2 line and a ninth channel at 89 GHz) are offset from oneanother in frequency by 70 MHz, and the band edge of themost opaque channel is approximately 50 MHz from the O2

line center. A transparent channel at 89 GHz was added toeach 118-GHz receiver bank to improve profiling performanceat the surface and near the boundary layer, as a result of thediscussion in Section IV. The two 183-GHz receiver banks(each with eight channels split among an 8-GHz bandwidthnear the 183.31-GHz H2O line) are offset from one another infrequency by 500 MHz, and the band edge of the most opaquechannel is 10 MHz from the H2O line center. The ΔTrms

of each water vapor channel is 0.17 K. The frequency offsetand overlap values were obtained using simple trial-and-errorexperiments. Further channel optimizations should improve theGeoMAS retrieval performance but are beyond the scope ofthis paper.

Page 8: Hyperspectral Microwave Atmospheric Sounding · sensing modality for atmospheric sounding, driven by recent ad-vances in microwave device technology that now permit receiver arrays

BLACKWELL et al.: HYPERSPECTRAL MICROWAVE ATMOSPHERIC SOUNDING 135

B. STAR

A STAR has recently been suggested for geostationary im-plementation [37], [38] with approximately 900 receivers, i.e.,300 operating near 60 GHz and 600 operating near 183 GHz.We begin with identical fundamental assumptions used forthe GeoMAS system: 40 000 temperature profiles (50-km gridspacing) and 160 000 water vapor profiles (25-km grid spacing)derived in 15 min. Synthetic and unsharpened filled-aperturesystems yield comparable image noise levels for comparablereceivers and total integration times, provided that both systemssurvey the entire visible Earth and have the same bandwidthsand receiver noise temperatures [39].

The STAR configuration follows closely from [38], althoughwe have included broader bandwidths (therefore favorable toSTAR performance) to facilitate comparisons with AMSU-A/Bperformance. Ten STAR channels are used: six that are identicalto AMSU-A temperature bands near 60 GHz, three that areidentical to AMSU-B water vapor bands near 183 GHz, anda single band at 167 GHz. We assume the system temperaturesof 500 K for each 60-GHz receiver and 800 K for each 167-/183-GHz receiver. Because each receiver can sample only asingle channel at a time, the integration time available to agiven channel is equal to the total integration time per pixel(22.5 ms for 60 GHz and 5.625 ms for 183 GHz) divided by thenumber of channels in the receiver (six near 60 GHz and fournear 183 GHz). The STAR channel properties are summarizedin Table II.

VII. TEMPERATURE AND MOISTURE SOUNDING

PERFORMANCE COMPARISONS FOR NOTIONAL

GEOSTATIONARY SYSTEMS

We now examine temperature and water vapor profile re-trieval performance in nonprecipitating atmospheres for boththe GeoMAS and STAR notional systems. Both ocean andland cases are included in equal proportion, and the NOAA88band MM5 profile sets are analyzed separately. Performance isindicated by rms errors in 1-km atmospheric layers at nadirincidence (off-nadir performance differs negligibly).

A. Temperature

The temperature profile retrieval performance curves areshown in Fig. 11. Also shown is the surface temperature rmserror. The GeoMAS performance is excellent both for theNOAA88b profile set (high variability) and for the MM5 profileset (high water content). GeoMAS performance exceeds STARperformance for both profile sets at all atmospheric levels,including the surface. The performance difference in the rela-tively moist MM5 profile set is substantial, with approximately0.5-K difference in rms error throughout the troposphere,including the critical atmospheric boundary layer. GeoMASsurface skin temperature retrievals are substantially superior tothose of STAR, with a difference in rms error of approximately1 K. Note that the STAR performance closely resembles thatof AMSU-A/B, as the primary tropospheric sounding channelsare identical. The absence of relatively transparent channels at

Fig. 11. Comparison of GeoMAS (118-/183-GHz) and STAR (60-/183-GHz)temperature retrieval performances over land and ocean at nadir incidence. Theleft panel shows the rms error in 1-km layers using the NOAA88b data set(scattering was not modeled), and the right panel shows the rms error usingthe MIT MM5 data set (scattering was modeled). The surface temperature rmserror is indicated by a circle for GeoMAS and a triangle for STAR.

23.8 and 31.4 GHz slightly degrades the performance in thelower boundary layer, with the primary impact on water vaporretrieval over the ocean.

The GeoMAS performance advantage as shown in Fig. 11is due to two factors. First, the high density of the weightingfunctions in the vertical dimension, as shown in Fig. 3, enablesGeoMAS measurements to capture profile information withhigh vertical resolution. Second, the large total bandwidth af-forded by the GeoMAS system allows relatively high sensitivityto be achieved. We now explore separately the bandwidthbenefit and the channelization benefit.

Fig. 12 shows the rms temperature profile averaged over allatmospheric layers from 0 to 20 km as a function of the num-ber of receiver banks included in a hyperspectral microwavesystem. It is obvious that increased sensitivity due to increasedbandwidth will always improve the retrieval results, but we nowexamine two more interesting cases. First, a fixed amount ofbandwidth is shared among all the receiver banks (the solidcurves in Fig. 12). The improvement in this case as receiverbanks are added is due only to the advantage afforded byincreased channelization, resulting in a more dense verticalspacing of the weighting functions. Note that the performancefor this case is identical to that for a sequential hyperspectralsystem, whereby a range of channels (generated by sweepingthe receiver local oscillator frequency, for example) is sequen-tially measured. That is, receiver bank 1 is selected for one-eighth of the total integration time, then bank 2 for one-eighth,and so forth, until all eight receiver banks are utilized. Thecurves are shown for five different integration times, based on ascaling of the integration time used to generate Fig. 11. The topcurve corresponds to the same integration time used to generateFig. 11, and this integration time was increased by factors of10, 100, 1000, and infinity (no noise) to generate the othercurves. The second case (the dashed curves in Fig. 12) assumesthat a fixed amount of bandwidth is allocated for each receiverbank. We term this multiplexed hyperspectral operation, as

Page 9: Hyperspectral Microwave Atmospheric Sounding · sensing modality for atmospheric sounding, driven by recent ad-vances in microwave device technology that now permit receiver arrays

136 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 49, NO. 1, JANUARY 2011

Fig. 12. Incremental benefit of additional eight-channel receiver banks. Allcurves indicate the rms temperature profile error averaged over all atmosphericlayers from 0 to 20 km as a function of the number of receiver banks used. Thesolid curves were calculated using a fixed bandwidth that is shared among allreceiver banks (sequential hyperspectral), and the dashed curves are calculatedusing a fixed bandwidth for each receiver bank (multiplexed hyperspectral).Each pair of dashed and solid curves was calculated using a different noisefactor (Nfac) which is a simple scaling of the integration time. The dashedand solid curves labeled “no noise” are identical.

all channels eventually view the scene for the duration of thetotal integration time. Both increased bandwidth and increasedchannelization positively impact the performance in this caseas receiver banks are added. Note that sequential hyperspectraloperation is never superior to multiplexed hyperspectral oper-ation, and the performance difference is most pronounced forshort per-pixel integration times characteristic of geostationarysensing platforms.

These results are characterized by two interesting features.First, the marginal benefit of increased channelization is appar-ent in all cases. Second, the amount of channelization benefit ismore pronounced as the integration time is increased.

We further examine the temperature profile retrieval perfor-mance by assessing the accuracy in the lower boundary layer(the 1-km layer that is nearest the surface) as a function ofthe atmospheric transmittance, shown in Fig. 13. The errorsshown have been normalized by the a priori standard deviationof the temperature over all the profiles included in each ofthe 20 transmittance bins. That is, a value of 0.5 indicatesthat the rms error is one-half of the rms error that wouldhave resulted using the mean temperature value (over the bin)as the estimate. The bell-shaped curve can be explained asfollows. Relatively small errors toward the right of the figureresult because the high atmospheric transmittance allows thesurface temperature to be sensed with high accuracy, and thesurface temperature tends to be correlated with the temperaturein the lower boundary layer. As the transmittance decreases,the surface is increasingly obscured, and the error thereforeincreases. As the transmittance decreases beyond 0.6 or so,the contribution from the lower boundary layer to the overallmeasured radiance for most channels is maximized, and theerror therefore decreases. As the transmittance approaches zero,the radiative contribution from the lowest 1-km layer decreasesdue to obscuration by opaque layers above.

Fig. 13. Comparison of GeoMAS (118-/183-GHz) and STAR (60-/183-GHz)temperature retrieval performances in the 1-km atmospheric layer that is nearestthe surface as a function of transmittance at 90 GHz. The errors are shownrelative to the a priori standard deviation in the profile set.

Fig. 14. Comparison of GeoMAS (118-/183-GHz) and STAR (60-/183-GHz)water vapor retrieval performances over land and ocean at nadir incidence. Theleft panel shows the rms error (relative to the a priori variation in the validationset) in 1-km layers using the NOAA88b data set (scattering was not modeled),and the right panel shows the rms error using the MM5 storm data set (scatteringwas modeled).

B. Water Vapor

The water vapor profile retrieval performance curves areshown in Fig. 14. The errors shown are the rms profile retrievalerror divided by the a priori standard deviation in the profile set.The GeoMAS performance is excellent both for the NOAA88bprofile set (high variability) and for the MM5 profile set (highwater content). GeoMAS performance exceeds STAR perfor-mance for both profile sets at all atmospheric levels.

VIII. INVESTIGATION OF ADDITIONAL ERROR SOURCES

The simulation analyses that we have presented thus far havebeen performed under ideal circumstances to accentuate therelative merits of the various sensor systems. We now consider

Page 10: Hyperspectral Microwave Atmospheric Sounding · sensing modality for atmospheric sounding, driven by recent ad-vances in microwave device technology that now permit receiver arrays

BLACKWELL et al.: HYPERSPECTRAL MICROWAVE ATMOSPHERIC SOUNDING 137

additional error sources that would be expected under realisticconditions. For example, simulation analyses often assume per-fect sensor and atmospheric physics, as well as perfect groundtruth, and these assumptions are invalid in practice. Of par-ticular concern are correlated errors that could be particularlydetrimental to a hyperspectral microwave system comprisingmany channels with broad overlapping weighting functions.None of the error sources considered here was presented to theNN retrievals during training.

A. Two-Dimensional Simulation Methodology

We begin with an examination of the following two corre-lated error sources: 1) unknown array misalignment error and2) spatial inhomogeneity error. Array misalignment could resultdue to mechanical and/or electrical imperfections in the antennaarrays used to produce the footprints shown in Fig. 1. Subpixelspatial inhomogeneities introduced by atmospheric and surfacefeatures further degrade the performance. We assess the impactof both of these error sources using the high-resolution griddedMM5 profile set.

B. Array Misalignment Error

Static unknown misalignment errors of approximately 10%of each temperature sounding footprint and 20% of each watervapor sounding footprint were introduced to the GeoMASarrays during simulation of the brightness temperatures. Thislevel of unknown misalignment error is expected to be muchlarger than would be encountered on an actual system. TheSTAR alignment was assumed to be perfect.

C. Spatial Inhomogeneity Error

Clouds, surface features, and water vapor are highly variablein the horizontal dimension, and subpixel inhomogeneitiesacross 50-km temperature and 25-km water vapor footprints arelikely. Gaussian antenna beam patterns for the GeoMAS andSTAR systems were convolved with the high-resolution (5-km)brightness temperature fields calculated using the MM5 profileset to simulate the effects of spatial inhomogeneities.

D. Simulation Error

An uncorrelated Gaussian random error term with 0.2-Kstandard deviation and zero mean was added to all calculatedbrightness temperatures. This error accounts to first order forimperfections in the surface, transmittance, and radiative trans-fer models used to derive the retrievals, as well as any errors inthe ground truth.

E. Simulation Results

The temperature and water vapor retrieval results after theinclusion of all the error sources above are shown in Figs. 15and 16, respectively. The GeoMAS temperature profile retrievalrms performance exceeds 1.5 K in 1-km layers, the surface tem-perature rms performance exceeds 2 K, and the water vapor pro-file retrieval performance exceeds 30% in 1-km layers. While

Fig. 15. Comparison of GeoMAS (118-/183-GHz) and STAR (60-/183-GHz)2-D temperature retrieval performances over land and ocean at nadir incidenceusing the MIT MM5 data set with detailed modeling of correlated error sources.The surface temperature rms error is indicated by a circle for GeoMAS and atriangle for STAR.

Fig. 16. Comparison of GeoMAS (118-/183-GHz) and STAR (60-/183-GHz)2-D water vapor retrieval performances (relative to the a priori variation in thevalidation set) over land and ocean at nadir incidence using the MIT MM5 dataset with detailed modeling of correlated error sources.

performance does degrade relative to Figs. 11 and 14, it is en-couraging to note the high degree of robustness of the GeoMASsystem to these error sources. Other sources of error are cur-rently under study (for example, IF filter bank stability), and as-sociated analysis results will be reported in a future publication.

IX. PRECIPITATION PERFORMANCE COMPARISON FOR

NOTIONAL GEOSTATIONARY SYSTEMS

Airborne [40], [41] and spaceborne [27], [42]–[44] retrievalsof precipitation from passive opaque millimeter-wave mea-surements have demonstrated the potential of a geostationarymillimeter-wave sensor for precipitation mapping and tracking.We now compare the rain-rate retrieval performance of theGeoMAS and STAR systems.

Page 11: Hyperspectral Microwave Atmospheric Sounding · sensing modality for atmospheric sounding, driven by recent ad-vances in microwave device technology that now permit receiver arrays

138 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 49, NO. 1, JANUARY 2011

TABLE IIIRAIN-RATE RETRIEVAL PERFORMANCE (RMS ERROR

IN MILLIMETERS PER HOUR) FOR THE GEOMAS AND

STAR SYSTEMS AT 25-km SPATIAL RESOLUTION

Brightness temperature simulations were carried out usingTBSCAT and incorporated a fluffy-sphere ice scattering modelwith a wavelength-dependent density F (λ) that was tuned tomatch total scattering cross sections computed for spherical,hexagonal plate, and rosette hydrometeors using a discrete-dipole electromagnetic scattering model, i.e., DDSCAT [14].The precipitation retrieval accuracies for GeoMAS and STARwere computed using NNs trained using 122 MM5 stormsdescribed in Section III-B. The algorithms are the same asdescribed in [39], employing three NNs. If the rain-rate esti-mates from the first network were over 8 mm/h, then the secondNN was used to estimate the 15-min average precipitation rate;otherwise, the third network was used. The inputs used to trainall three networks include the MM5 precipitation rates blurredto 25 km, a land/sea flag, and the current channel brightnesstemperatures and those observed 15 min earlier. The three lay-ers of each network had ten, five, and one neurons, respectively,where the first two layers used a hyperbolic tangent sigmoidfunction. The best of the 100 tested networks was used foreach network and task, where “best” means the minimum rmsretrieval errors over the full dynamic range. The results, shownin Table III, indicate excellent performance of GeoMAS relativeto STAR in light to moderately heavy rain.

X. TEMPERATURE AND MOISTURE RETRIEVAL

PERFORMANCE SIMULATION COMPARISONS

FOR HYPERSPECTRAL MICROWAVE

AND AIRS/AMSU-A/HSB

We now examine the sounding performance for low-Earth-orbit systems. Retrieval simulations were performed usingthe channel sets of the AIRS/AMSU-A/Humidity Sensor forBrazil (HSB) sounding suite [4] currently flying on the NASAAqua satellite and for two notional hyperspectral microwavesystems. The notional Hyperspectral Microwave Array Spec-trometer (HyMAS) systems include channels near 60 GHz(HyMAS60) and near 118 GHz (HyMAS118). Both HyMASsystems include 16 water vapor channels near 183 GHz. TheHyMAS118 and GeoMAS systems contain identical channels(see Section VI-A and Table I). The HyMAS60 system contains64 channels near the oxygen lines in the 60-GHz region (seeTable IV). Each receiver bank has an offset in frequency by240 MHz. No attempts were made to optimize the channeliza-tion of either of the HyMAS systems. As with the GeoMASsimulations, system temperatures at 60, 118, and 183 GHzwere assumed to be 500, 650, and 800 K, respectively. Wehave assumed an integration time of 165 ms, which closelyapproximates that of the AMSU-A sensor. (The AMSU-A1

TABLE IVHyMAS60 CENTER FREQUENCIES, CHANNEL BANDWIDTHS,

AND ΔTrms FOR A SINGLE RECEIVER BANK

integration time is approximately 165 ms, and the AMSU-A2integration time is approximately 160 ms.) The channel prop-erties of the AIRS (2378 channels), AMSU-A (15 channels),and HSB (4 channels) sensors were obtained from the AquaLevel-1B Version-5 channel list files. The AMSU-A/HSBΔTrms values have been improved to reflect the system tem-peratures mentioned earlier. The actual AMSU/HSB ΔTrms

values are about twice those assumed here, where we haveincorporated the recent improvements reported in [9].

A. Simulation Methodology

The methods and assumptions used to carry out the retrievalperformance simulations presented in this section are verysimilar to those used earlier in this paper. Notable differencesinclude the profile database and the simulation and retrievalmethodology used for the combined microwave and infraredobservations.

1) AIRS/AMSU-A/HSB Level-2 Global Profile Database:Global atmospheric profile data derived from AIRS/AMSU-A/HSB observations dating back to August 30, 2002, areavailable from the NASA Data and Information ServicesCenter (http://disc.sci.gsfc.nasa.gov/AIRS/data-holdings). The“Version-5 Support Product” (AIRX2SUP) was used as theground truth in the performance simulation analyses presentedin this section. This product includes temperature, moisture,and cloud liquid water profiles reported on 100 atmosphericlevels with a 45-km horizontal spatial resolution at nadir. Thisproduct also includes the retrieved cloud-top pressures andcloud fractions for two cloud layers at each of the nine AIRSfields of view that comprise the AIRS/AMSU-A/HSB fieldof regard. All of the profile and cloud variables available inthe AIRX2SUP product were used in the radiative transfersimulations. Furthermore, the AIRS Level-2 Version-5 qualityflag “Pgood” was required to equal the surface pressure for theprofile to be included in the data set. The resulting distributionof cloud fractions in the data set was approximately uniformfrom 0% to 90%, with a spike near zero with a relative fre-quency of approximately 10%.

The AIRS Level-2 profiles and cloud products uniformlydistributed from December 2004 to January 2006 were used tosimulate a global training database of cloudy AIRS/AMSU-A/HSB observations over ocean at nadir incidence. A separatevalidation database was constructed using profiles from sevenAIRS focus days from September 6, 2002, to December 5, 2003.

Page 12: Hyperspectral Microwave Atmospheric Sounding · sensing modality for atmospheric sounding, driven by recent ad-vances in microwave device technology that now permit receiver arrays

BLACKWELL et al.: HYPERSPECTRAL MICROWAVE ATMOSPHERIC SOUNDING 139

Fig. 17. Comparison of HyMAS60 (60-/183-GHz), HyMAS118 (118-/183-GHz), AIRS/AMSU-A/HSB, and AMSU-A/HSB-only temperature re-trieval performances over ocean at nadir incidence for seven focus days from2002 to 2003. AMSU-A/HSB noise has been reduced from on-orbit values forconsistency with the Tsys values used throughout this paper.

A 3 × 3 “golfball” with nine AIRS/HSB footprints and oneAMSU-A footprint was simulated for each of the 80 000profiles in the simulation training and validation data set(40 000 profiles were included in each set). The Stand-AloneRapid Transmittance Algorithm [45] was used to simulate theAIRS observations, and the TBARRAY algorithm [17] wasused to simulate the AMSU-A, HSB, and HyMAS observa-tions. This data set was also used to simulate an ensembleof 45-km HyMAS60 and HyMAS118 footprints. The retrievalmethodology used for the HyMAS observations is identical tothat presented earlier in this paper for GeoMAS.

2) AIRS/AMSU-A/HSB Cloud Clearing and RetrievalMethodology: The cloudy AIRS observations were first “cloudcleared” using the microwave observations together with theAIRS 3 × 3 field of regard to estimate the radiances that wouldhave been observed by AIRS if the scene were cloud free.The stochastic cloud clearing (SCC) algorithm [46] was usedin this work. The SCC algorithm also produces a degree-of-cloudiness estimate that was used to identify “mostly clear”scenes (approximately the clearest 30%) and “mostly cloudy”scenes (approximately the cloudiest 20%). The microwaveand cloud-cleared infrared radiances were then used as inputsto the PPC/NN algorithm [28], and the temperature andmoisture profiles were estimated. This methodology has beenextensively validated with actual AIRS/AMSU observations[32], [47] using a variety of performance metrics.

B. Simulated Performance Comparisons of HyMAS SystemsWith AIRS/AMSU-A/HSB

The temperature and water vapor retrieval performancesof the HyMAS60, HyMAS118, AIRS/AMSU-A/HSB, andAMSU-A/HSB-only systems are shown in Figs. 17 and 18,respectively. We note that the AIRS/AMSU-A/HSB resultspresented here agree well with those presented by other in-vestigators [48], [49]. The same set of 40 000 profiles (fromthe seven-day 2002–2003 validation set) was used to derive

Fig. 18. Comparison of HyMAS60 (60-/183-GHz), HyMAS118 (118-/183-GHz), AIRS/AMSU-A/HSB, and AMSU-A/HSB-only water vapor re-trieval performances over ocean at nadir incidence for seven focus days from2002 to 2003. AMSU-A/HSB noise has been reduced from on-orbit values forconsistency with the Tsys values used throughout this paper.

the performance of each of the four systems. It is interest-ing to note that the performance of HyMAS60 in the mid-and lower troposphere is very similar to the AIRS/AMSU-A/HSB system. HyMAS118 water vapor sounding performanceis also very competitive with AIRS/AMSU-A/HSB. HyMASimprovements could be obtained by optimizing the channel sets(center frequencies and bandwidths) and/or by increasing thetotal number of temperature sounding receiver banks.

C. Simulated Performance Comparisons of HyMAS60 WithAIRS/AMSU-A/HSB in Mostly Clear and MostlyCloudy Conditions

We now examine the performance in mostly clear scenes(favorable to infrared sounding) and in mostly cloudy scenes(favorable to microwave sounding). The temperature and wa-ter vapor retrieval performances of the HyMAS60 and AIRS/AMSU-A/HSB systems are shown in Figs. 19 and 20, respec-tively. The same set of mostly clear or mostly cloudy profiles(from the seven-day 2002–2003 validation set) was used toderive the performance of each system. The midtropospherictemperature sounding performance of AIRS/AMSU-A/HSB isslightly superior to HyMAS60 in mostly clear scenes.HyMAS60 shows a slight advantage in boundary layer tem-perature retrieval in mostly clear and mostly cloudy conditionsand a pronounced advantage (0.4-K difference in rms) forsurface skin temperature retrieval in mostly cloudy conditions.HyMAS60 water vapor sounding performance exceeds AIRS/AMSU-A/HSB for both mostly clear and mostly cloudy scenes.

D. Discussion of Off-Nadir Performance

The results shown previously have been only for observationsat nadir incidence. Additional worst case analyses wereperformed at a sensor scan angle of 48◦, corresponding to theedge of scan for AMSU-A. The scan-angle dependence of theobservations is easily accommodated by the NN retrieval [28].

Page 13: Hyperspectral Microwave Atmospheric Sounding · sensing modality for atmospheric sounding, driven by recent ad-vances in microwave device technology that now permit receiver arrays

140 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 49, NO. 1, JANUARY 2011

Fig. 19. Comparison of HyMAS60 (60-/183-GHz) and AIRS/AMSU-A/HSBtemperature retrieval performances over ocean at nadir incidence for mostlyclear and mostly cloudy conditions for seven focus days from 2002 to 2003.AMSU-A/HSB noise has been reduced from on-orbit values for consistencywith the Tsys values used throughout this paper.

Fig. 20. Comparison of HyMAS60 (60-/183-GHz) and AIRS/AMSU-A/HSBwater vapor retrieval performances over ocean at nadir incidence for mostlyclear and mostly cloudy conditions for seven focus days from 2002 to 2003.AMSU-A/HSB noise has been reduced from on-orbit values for consistencywith the Tsys values used throughout this paper.

The temperature profile rms error degradation relative to thenadir-viewing configuration was no worse than 0.1 K through-out the troposphere for all cases shown in Fig. 17, and the watervapor profile rms error degradation relative to the nadir-viewingconfiguration was no worse than 1% (i.e., rmsnadir in percentminus rms48◦ in percent) throughout the troposphere for allcases shown in Fig. 18. The skin surface temperature retrievalrms error for HyMAS118 at 48◦ incidence degraded by 0.41 Krelative to the nadir-viewing configuration. The skin surfacetemperature rms error degradation for the AIRS/AMSU-A/HSB and AMSU-A/HSB cases was 0.16 K and 0.14 K,respectively. The skin surface temperature rms error degra-dation for HyMAS60 was less than the estimated uncertaintyof the retrieval simulation analyses, which is approx-imately 0.05 K.

XI. SUMMARY, CONCLUSION, AND FUTURE DIRECTIONS

We have presented a new modality of atmospheric soundingusing hyperspectral observations at microwave and millimeter-wave frequencies. The hyperspectral channel composition isachieved by multiplexing a number of independent spectrom-eters to build up a large number of densely spaced weightingfunctions. The performance of such a system, i.e., GeoMAS,was demonstrated using channels near 118 and 183 GHz, whichare well suited for geostationary implementation due to therelatively small antenna reflector sizes that would be requiredfor useful spatial resolution. A similar methodology couldbe adapted for low-Earth-orbit sensors using frequencies near60 GHz, and the analysis presented in this paper suggeststhat an 88-channel HyMAS system operating near 60, 89, and183 GHz could rival the performance of state-of-the-art micro-wave+infrared sounding suites at a small fraction of the cost.

Hyperspectral microwave sounding and precipitation map-ping performance could be further improved by spatial process-ing of Nyquist-sampled observations to sharpen the effectiveantenna beam. This sharpening amplifies sensor noise, but thisincreased noise could be offset by adding receiver banks. Apromising area of the current study is the investigation ofthe tradeoff between effective spatial resolution, receiver arraycomplexity, and retrieval performance. The antenna reflectordiameter required to meet a 50-/25-km temperature/water-vaporresolution goal from a geostationary orbit is approximately2.7 m for a filled-aperture system. However, the results pre-sented in this paper, together with other works on antenna beamsharpening [39], indicate that such requirements could be metwith a GeoMAS system with an antenna reflector diameterapproaching 2 m if image sharpening is used during groundprocessing. Furthermore, the resolution versus noise trade canbe dynamically optimized, which means that the level of sharp-ening can be selected “on the fly” based on the atmosphericscene being viewed. Scenes with high spatial frequency contentcould be detected, and sharpening could be used. The selectionof the degree of sharpening could be made during groundprocessing and could be tailored by each user.

While the results presented in this paper are encouragingand indicate great promise for future hyperspectral microwavesounding systems, additional analyses and concept demon-stration are clearly necessary to study the advantages andchallenges of this new sensing modality. Further studies ofthe error sensitivities of hyperspectral microwave systems arerecommended, for example, and deployment of airborne pro-totype hyperspectral microwave sensors would provide insightinto many of the noise and meteorological phenomena that aredifficult to accurately model and simulate.

Another interesting area of further research is the potentialsynergy resulting from the combined use of hyperspectral mi-crowave and hyperspectral infrared systems. For example, atemperature profile retrieval of high accuracy provided by ahyperspectral microwave system could be used to substantiallyimprove the estimates of atmospheric carbon dioxide usingan infrared sounder, as the relationship between temperatureand carbon dioxide could be decoupled using microwave andinfrared measurements.

Page 14: Hyperspectral Microwave Atmospheric Sounding · sensing modality for atmospheric sounding, driven by recent ad-vances in microwave device technology that now permit receiver arrays

BLACKWELL et al.: HYPERSPECTRAL MICROWAVE ATMOSPHERIC SOUNDING 141

ACKNOWLEDGMENT

The authors would like to thank the three anonymous review-ers for the helpful comments and suggestions which improvedthis paper, D. Staelin for the thoughtful guidance on all facetsof this work, and F. Chen, P. Rosenkranz, M. Shields, andJ. Solman for the helpful discussions.

REFERENCES

[1] D. H. Staelin, A. H. Barrett, J. W. Waters, F. T. Barath, E. J. Johnston,P. W. Rosenkranz, N. E. Gaut, and W. B. Lenoir, “Microwave spectrometeron the Nimbus 5 satellite: Meteorological and geophysical data,” Science,vol. 182, no. 4119, pp. 1339–1341, Dec. 1973.

[2] W. L. Smith, “Satellite techniques for observing the temperature structureof the atmosphere,” Bull. Amer. Meteorol. Soc., vol. 53, no. 11, pp. 1074–1082, Nov. 1972.

[3] W. L. Smith, “Atmospheric soundings from satellites—False expectationor the key to improved weather prediction?,” Q. J. R. Meteorol. Soc.,vol. 117, no. 498, pp. 267–297, Jan. 1991.

[4] H. H. Aumann, M. T. Chahine, C. Gautier, M. D. Goldberg, E. Kalnay,L. M. McMillin, H. Revercomb, P. W. Rosenkranz, W. L. Smith,D. H. Staelin, L. L. Strow, and J. Susskind, “AIRS/AMSU/HSB on theAqua mission: Design, science objectives, data products, and processingsystems,” IEEE Trans. Geosci. Remote Sens., vol. 41, no. 2, pp. 253–264,Feb. 2003.

[5] G. Chalon, F. Cayla, and D. Diebel, “IASI: An advanced sounder foroperational meteorology,” in Proc. 52nd Congr. IAF, Oct. 2001, pp. 1–5.

[6] W. L. Smith, H. Revercomb, G. Bingham, A. Larar, H. Huang, D. Zhou,J. Li, X. Liu, and S. Kireev, “Evolution, current capabilities, and futureadvance in satellite nadir viewing ultra-spectral IR sounding of the loweratmosphere,” Atmos. Chem. Phys., vol. 9, no. 15, pp. 5563–5574, 2009.

[7] B. Razavi, “A 60-GHz CMOS receiver front-end,” IEEE J. Solid-StateCircuits, vol. 41, no. 1, pp. 17–22, Jan. 2006.

[8] J. Powell, H. Kim, and C. Sodini, “SiGe receiver front ends for millimeter-wave passive imaging,” IEEE Trans. Microw. Theory Tech., vol. 56,no. 11, pp. 2416–2425, Nov. 2008.

[9] P. Kangaslahti, D. Pukala, T. Gaier, W. Deal, X. Mei, and R. Lai, “Lownoise amplifier for 180 GHz frequency band,” in Proc. IEEE MTT-S Int.Microw. Symp. Dig., Jun. 2008, pp. 451–454.

[10] W. J. Blackwell and D. H. Staelin, “Comparative performance analysesof passive microwave systems for tropospheric sounding of temperatureand water vapor profiles,” in Proc. SPIE GOES-8 Beyond, Aug. 1996,vol. 2812, pp. 472–478.

[11] D. Staelin, J. Kerekes, and F. J. Solman, “Final report of the geosynchro-nous microwave sounder working group,” MIT Lincoln Lab., Lexington,MA, Aug. 22, 1997, prepared for NOAA/NESDIS GOES Program Office.

[12] A. Lipton, “Satellite sounding channel optimization in the microwavespectrum,” IEEE Trans. Geosci. Remote Sens., vol. 41, no. 4, pp. 761–781, Apr. 2003.

[13] C. Muth, P. Lee, J. Shiue, and W. Allan Webb, “Advanced TechnologyMicrowave Sounder on NPOESS and NPP,” in Proc. IEEE IGARSS,Sep. 2004, vol. 4, pp. 2454–2458.

[14] C. Surussavadee and D. H. Staelin, “Comparison of AMSU millimeter-wave satellite observations, MM5/TBSCAT predicted radiances, and elec-tromagnetic models for hydrometeors,” IEEE Trans. Geosci. RemoteSens., vol. 44, no. 10, pp. 2667–2678, Oct. 2006.

[15] C. Surussavadee and D. H. Staelin, “Millimeter-wave precipitation re-trievals and observed-versus-simulated radiance distributions: Sensitivityto assumptions,” J. Atmos. Sci., vol. 64, no. 11, pp. 3808–3826, Nov. 2007.

[16] F. J. Wentz, “A well-calibrated ocean algorithm for SSM/I,” J. Geophys.Res., vol. 102, no. C4, pp. 8703–8718, 1997.

[17] P. W. Rosenkranz, “Absorption of microwaves by atmosphericgases,” in Atmospheric Remote Sensing by Microwave Radiometry,M. A. Janssen, Ed. New York: Wiley, Jul. 1993, ch. 2.

[18] H. J. Liebe, “MPM: An atmospheric millimeter-wave propagationmodel,” Int. J. Infrared Millim. Waves, vol. 10, no. 6, pp. 631–650,Jun. 1989.

[19] H. J. Liebe, P. W. Rosenkranz, and G. A. Hufford, “Atmospheric 60-GHzoxygen spectrum: New laboratory measurements and line parameters,”J. Quant. Spectrosc. Radiat. Transf., vol. 48, no. 5/6, pp. 629–643,Nov. 1992.

[20] P. W. Rosenkranz, “Radiative transfer solution using initial values in ascattering and absorbing atmosphere with surface reflection,” IEEE Trans.Geosci. Remote Sens., vol. 40, no. 8, pp. 1889–1892, Aug. 2002.

[21] F. T. Ulaby, R. K. Moore, and A. K. Fung, Microwave Remote SensingActive and Passive. Norwood, MA: Artech House, 1981.

[22] S. English and T. Hewison, “A fast generic millimeter-wave emissivitymodel,” Proc. SPIE, vol. 3503, pp. 288–300, Aug. 1998.

[23] G. W. Petty and K. B. Katsaros, “The response of the SSM/I to the marineenvironment. Part II: A parameterization of the effect of the sea surfaceslope distribution on emission and reflection,” J. Atmos. Ocean. Technol.,vol. 11, no. 3, pp. 617–628, 1994.

[24] F. Karbou, C. Prigent, L. Eymard, and J. R. Pardo, “Microwave land emis-sivity calculations using AMSU measurements,” IEEE Trans. Geosci.Remote Sens., vol. 43, no. 5, pp. 948–959, May 2005.

[25] A. J. Gasiewski, “Numerical sensitivity analysis of passive EHF andSMMW channels to tropospheric water vapor, clouds, and precipita-tion,” IEEE Trans. Geosci. Remote Sens., vol. 30, no. 5, pp. 859–869,Sep. 1992.

[26] C. Prigent, J. R. Pardo, and W. B. Rossow, “Comparisons of the millimeterand submillimeter bands for atmospheric temperature and water vaporsoundings for clear and cloudy skies,” J. Appl. Meteorol., vol. 45, no. 12,pp. 1622–1633, Dec. 2006.

[27] D. H. Staelin and F. W. Chen, “Precipitation observations near 54 and 183GHz using the NOAA-15 satellite,” IEEE Trans. Geosci. Remote Sens.,vol. 38, no. 5, pp. 2322–2332, May 2000.

[28] W. J. Blackwell, “A neural-network technique for the retrieval of at-mospheric temperature and moisture profiles from high spectral resolu-tion sounding data,” IEEE Trans. Geosci. Remote Sens., vol. 43, no. 11,pp. 2535–2546, Nov. 2005.

[29] W. J. Blackwell, “Neural network retrievals of atmospheric tempera-ture and moisture profiles from high-resolution infrared and microwavesounding data,” in Signal and Image Processing for Remote Sensing,C. C. Chen, Ed. Boca Raton, FL: Taylor & Francis, 2006, ch. 11.

[30] W. J. Blackwell and F. W. Chen, Neural Networks in Atmospheric RemoteSensing. Boston, MA: Artech House, 2009.

[31] K. M. Hornik, M. Stinchcombe, and H. White, “Multilayer feedfor-ward networks are universal approximators,” Neural Netw., vol. 4, no. 5,pp. 359–366, 1989.

[32] W. J. Blackwell, M. Pieper, and L. G. Jairam, “Neural network estima-tion of atmospheric profiles using AIRS/IASI/AMSU in the presenceof clouds,” in Proc. SPIE Asia-Pacific Remote Sens. Symp., Nov. 2008,pp. I-122–I-125.

[33] D. Nguyen and B. Widrow, “Improving the learning speed of two-layerneural networks by choosing initial values of the adaptive weights,” inProc. IJCNN, 1990, vol. 3, pp. 21–26.

[34] D. Marquardt, “An algorithm for least-squares estimation of nonlinearparameters,” J. Soc. Ind. Appl. Math., vol. 11, no. 2, pp. 431–441,Jun. 1963.

[35] M. T. Hagan and M. B. Menhaj, “Training feedforward networks with theMarquardt algorithm,” IEEE Trans. Neural Netw., vol. 5, no. 6, pp. 989–993, Nov. 1994.

[36] I. T. Nabney, NETLAB: Algorithms for Pattern Recognition. London,U.K.: Springer-Verlag, 2002.

[37] A. Tanner, W. Wilson, B. Lambrigsten, S. Dinardo, S. Brown,P. Kangaslahti, T. Gaier, C. Ruf, S. Gross, B. Lim, S. Musko, S. Rogacki,and J. Piepmeier, “Initial results of the Geostationary Synthetic ThinnedArray Radiometer (GeoSTAR) demonstrator instrument,” IEEE Trans.Geosci. Remote Sens., vol. 45, no. 7, pp. 1947–1957, Jul. 2007.

[38] B. Lambrigtsen, S. Brown, T. Gaier, P. Kangaslahti, and A. Tanner, “Abaseline for the decadal-survey path mission,” in Proc. IEEE IGARSS,Jul. 2008, vol. 3, pp. 338–341.

[39] D. H. Staelin and C. Surussavadee, “Precipitation retrieval accuracies forgeo-microwave sounders,” IEEE Trans. Geosci. Remote Sens., vol. 45,no. 10, pp. 3150–3159, Oct. 2007.

[40] W. J. Blackwell, J. W. Barrett, F. W. Chen, R. V. Leslie, P. W. Rosenkranz,M. J. Schwartz, and D. H. Staelin, “NPOESS aircraft sounder testbed-microwave (NAST-M): Instrument description and initial flight results,”IEEE Trans. Geosci. Remote Sens., vol. 39, no. 11, pp. 2444–2453,Nov. 2001.

[41] R. V. Leslie and D. H. Staelin, “NPOESS aircraft sounder testbed-microwave: Observations of clouds and precipitation at 54, 118, 183, and425 GHz,” IEEE Trans. Geosci. Remote Sens., vol. 42, no. 10, pp. 2240–2247, Oct. 2004.

[42] F. W. Chen and D. H. Staelin, “AIRS/AMSU/HSB precipitation esti-mates,” IEEE Trans. Geosci. Remote Sens., vol. 41, no. 2, pp. 410–417,Feb. 2003.

[43] C. Surussavadee and D. H. Staelin, “Global millimeter-wave precipitationretrievals trained with a cloud-resolving numerical weather predictionmodel, Part I: Retrieval design,” IEEE Trans. Geosci. Remote Sens.,vol. 46, no. 1, pp. 99–108, Jan. 2008.

Page 15: Hyperspectral Microwave Atmospheric Sounding · sensing modality for atmospheric sounding, driven by recent ad-vances in microwave device technology that now permit receiver arrays

142 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 49, NO. 1, JANUARY 2011

[44] C. Surussavadee and D. H. Staelin, “Global millimeter-wave precipitationretrievals trained with a cloud-resolving numerical weather predictionmodel, Part II: Performance evaluation,” IEEE Trans. Geosci. RemoteSens., vol. 46, no. 1, pp. 109–118, Jan. 2008.

[45] L. Strow, S. Hannon, and S. Desouza-Machado, “An overview of the AIRSradiative transfer model,” IEEE Trans. Geosci. Remote Sens., vol. 41,no. 2, pp. 303–313, Feb. 2003.

[46] C. Cho and D. H. Staelin, “AIRS observations versus numerical weatherpredictions of cloud-cleared radiances,” J. Geophys. Res., vol. 111, 2006.

[47] H. H. Aumann, E. Manning, C. Barnet, E. Maddy, and W. Blackwell,“An anomaly correlation skill score for the evaluation of the performanceof hyperspectral infrared sounders,” in Proc. SPIE Atmos. Environ. Re-mote Sens. Data Process. Utilization V—Readiness for GEOSS III, 2009,vol. 7456, pp. 745 60T-1–745 60T–7.

[48] J. Susskind, C. D. Barnet, and J. M. Blaisdell, “Retrieval of atmosphericand surface parameters from AIRS/AMSU/HSB data in the presence ofclouds,” IEEE Trans. Geosci. Remote Sens., vol. 41, no. 2, pp. 390–409,Feb. 2003.

[49] L. Shi, “Retrieval of atmospheric temperature profiles from AMSU—A measurement using a neural network approach,” J. Atmos. Ocean.Technol., vol. 18, no. 3, pp. 340–347, Mar. 2001.

William J. Blackwell (S’92–M’02–SM’07) receivedthe B.E.E. degree in electrical engineering from theGeorgia Institute of Technology, Atlanta, in 1994 andthe S.M. and Sc.D. degrees in electrical engineeringand computer science from the Massachusetts Insti-tute of Technology (MIT), Cambridge, in 1995 and2002, respectively.

Since 2002, he has been with the Lincoln Labo-ratory, MIT, Lexington, where he is currently a Se-nior Member of the Technical Staff with the SensorTechnology and System Applications Group. He is

the Integrated Program Office Sensor Scientist for the Advanced TechnologyMicrowave Sounder on the National Polar-orbiting Operational EnvironmentalSatellite System (NPOESS) Preparatory Project planned for launch in 2011and the Atmospheric Algorithm Development Team Leader for the NPOESSMicrowave Imager/Sounder planned for launch in 2016. He is a coauthor ofNeural Networks in Atmospheric Remote Sensing (Artech House, 2009). Hisprimary research interests are in the area of atmospheric remote sensing, in-cluding the development and calibration of airborne and spaceborne microwaveand hyperspectral infrared sensors, the retrieval of geophysical products fromremote radiance measurements, and the application of electromagnetic signalprocessing and estimation theory.

Dr. Blackwell held a National Science Foundation Graduate Research Fel-lowship from 1994 to 1997. He is a member of Tau Beta Pi, Eta Kappa Nu,Phi Kappa Phi, Sigma Xi, the American Meteorological Society, the AmericanGeophysical Union, and Commission F of the International Union of Radio Sci-ence. He is currently an Associate Editor of the IEEE Geoscience and RemoteSensing Society (GRSS) Newsletter and the Chair of the Boston Section of theIEEE GRSS and serves on the NPOESS Sounding Operational Algorithm Teamand the NASA Atmospheric InfraRed Sounder/NPOESS Preparatory ProjectScience Team. He was the recipient of the 2009 NOAA David Johnson Awardfor his work in neural network retrievals and microwave calibration.

Laura J. Bickmeier received the B.S. degree in elec-trical engineering from Boston University, Boston,MA, in 1998 and the M.S. degree in electrical en-gineering from the University of Alaska Fairbanks,Fairbanks, in 2000.

From 2000 to 2001, she was a Research Engineerwith Hovemere, Ltd., Tonbridge, U.K. From 2001to 2004, she was a Satellite Data Programmer withthe Alaska Volcano Observatory, Fairbanks. She iscurrently a Scientific Programmer with the LincolnLaboratory, Massachusetts Institute of Technology,

Lexington.

R. Vincent Leslie received the B.S. degree in elec-trical engineering from Boston University, Boston,MA, in 1998 and the S.M. and Sc.D. degrees in elec-trical engineering from the Massachusetts Institute ofTechnology (MIT), Cambridge, in 2000 and 2004,respectively.

Since 2004, he has been with the Lincoln Lab-oratory, MIT, Lexington, where he is currently aMember of the Technical Staff with the Sensor Tech-nology and System Applications Group.

Dr. Leslie is a member of Tau Beta Pi and theOrder of the Engineer.

Michael L. Pieper received the B.S. and M.S. de-grees in electrical engineering from NortheasternUniversity, Boston, MA, in 2007.

Since 2007, he has been an Associate Techni-cal Staff Member with the Sensor Technology andSystem Applications Group, Lincoln Laboratory,Massachusetts Institute of Technology, Lexington.He has been involved with the development of hy-perspectral infrared imaging detection and discrim-ination algorithms for the Advanced ResponsiveTactically Effective Military Imaging Spectrometer

sensor on the TacSat-3 satellite and the atmospheric profile retrieval algorithmdevelopment for the Atmospheric InfraRed Sounder/Advanced MicrowaveSounding Unit sensor suite on the NASA Aqua satellite.

Jenna E. Samra received the B.S. andM.S. degrees in electrical engineering fromThe Pennsylvania State University, University Park,in 2006 and 2008, respectively.

Since 2008, she has been an Associate Tech-nical Staff Member with the Sensor Technologyand System Applications Group, Lincoln Laboratory,Massachusetts Institute of Technology, Lexington.

Chinnawat Surussavadee (S’04–M’07) receivedthe B.Eng. degree in electrical engineering from theKing Mongkut’s Institute of Technology, Bangkok,Thailand, in 1999, the M.S. degree in electric powerengineering from the Rensselaer Polytechnic Insti-tute, Troy, NY, in 2001, and the Ph.D. degree in elec-trical engineering from the Massachusetts Institute ofTechnology (MIT), Cambridge, in 2007.

From 2004 to 2006, he was a Research Assistantwith the Remote Sensing and Estimation Group, Re-search Laboratory of Electronics (RLE), MIT. From

2006 to 2007, he was a Lecturer with Phuket Rajabhat University, Phuket,Thailand. He is currently a Research Affiliate with RLE and also a Lecturer andthe Head of the Andaman Environment and Natural Disaster Research Center,Faculty of Technology and Environment, Prince of Songkla University, Phuket.

Carolyn A. Upham received the B.S. degree inphysics from the University of Lowell, Lowell, MA,in 1986 and the B.S. degree in meteorology andthe M.S. degree in environmental studies, with aconcentration in atmospheric science, from the Uni-versity of Massachusetts, Lowell, in 1995 and 2004,respectively.

In 1995, she joined the Weather Sensing Group,Lincoln Laboratory, Massachusetts Institute of Tech-nology, Lexington, monitoring Terminal DopplerWeather Radar field sites, where she has been with

the Sensor Technology and System Applications Group analyzing data fromairborne and spaceborne sensors for the past 13 years.