For 20 years the Naval Research Laboratory has been making continuous water vapor profile measurements at 22.235 GHz with the Water Vapor Millimeter-Wave Spectrometer (WVMS) instruments, with the program expanding from one to three instruments in the first 6 years. Since the initial deployments there have been gradual improvements in the instrument design which have improved data quality and reduced maintenance requirements. Recent technological developments have made it possible to entirely redesign the instrument and improve not only the quality of the measurements but also the capability of the instrument. We present the fourth-generation instrument now operating at Table Mountain, California, which incorporates the most recent advances in microwave radiometry. This instrument represents the most significant extension of our measurement capability to date, enabling us to measure middle atmospheric water vapor from ∼26–80 km.
 Water vapor is a vital trace gas for climate monitoring. It is the reservoir of odd hydrogen in the middle atmosphere, and this is important to ozone chemistry. It is also unique among long-lived tracers in that the amount of water vapor entering the middle atmosphere is sensitive to the temperature at the tropical tropopause where mass exchange occurs between the stratosphere and the troposphere. The Naval Research Laboratory Water Vapor Millimeter-Wave Spectrometer (WVMS) instruments have provided water vapor measurements for the Network for the Detection of Atmospheric Composition Change (NDACC) since 1992. These instruments have retrieved water vapor profiles in the upper stratosphere and mesosphere (40–80 km) by measuring the pressure broadened 22 GHz water vapor line. They are preferably deployed at dry, high-altitude sites, where the water vapor in the dry stratosphere and mesosphere can be observed through a minimal amount of tropospheric water vapor.
 The first Water Vapor Millimeter-Wave Spectrometer (WVMS1) was developed in the late 1980s at JPL's Table Mountain Facility in California (34.4°N, 242.3°E) and was deployed to the National Institute of Water and Atmospheric Research (NIWA) at Lauder, New Zealand (45.0°S, 169.7°E) in 1992 where it operated until 2011 [Nedoluha et al., 1995, 1996]. In the second-generation instrument the number of discrete filters was increased from 50 to 90 and the spectral resolution of the center ten channels was improved from 200 KHz to 50 KHz. This generation also added a new phase lock loop to stabilize the center frequency. The first instrument of the second generation (WVMS2) was developed alongside WVMS1. It made continuous measurements from 1991 to 1997 and at various times until August 2008 when it was decommissioned [Nedoluha et al., 1998]. A third instrument (WVMS3) was also developed at Table Mountain, California, and deployed to Mauna Loa, Hawaii (19.5°N, 204.4°E) in 1996 where it operated until 2011 [Nedoluha et al., 2003]. This instrument was similar to WVMS2 and of the same generation. The instruments are numbered in the order in which they are built and not by generation. The improved spectral resolution and frequency stability of the second-generation instruments were particularly important for increasing the capability of the measurement in the upper mesosphere.
 From 1997 to 2005 many improvements were made to all three instruments in pointing, cabling, calibration, cryogenics, RF interconnections, and weatherproofing. These improvements were made using the WVMS2 instrument at Table Mountain as an experimental platform. In 2005 WVMS2 was modernized with a new computer and USB relay control and temperature monitoring. This third-generation instrument utilized a single board computer and a 5 V to 3.3 V TTL converter to operate the discrete filters. This modification allowed the instruments to operate independently of the host computers, whereas they were previously limited by an I/O card operating on the ISA bus of a DOS PC. The software that operates the system could then be developed in C to run in a Windows environment. In 2006 these changes were implemented in WVMS3 at Mauna Loa [Nedoluha et al., 2009].
 The original instruments employed cryogenic amplifiers which required constant monitoring and yearly maintenance. Nevertheless, they sometimes failed unexpectedly. In 2007 a new room temperature low-noise amplifier (LNA) was installed in WVMS2. The noise temperature increased by ∼80 K, but eliminating the cryogenics removed a key failure point and made it possible to eliminate yearly maintenance trips. A new room temperature LNA was also installed on WVMS1 in 2008.
 In 2008 the fourth-generation instrument was developed and deployed at JPL's Table Mountain facility in California, replacing the now decommissioned WVMS2 instrument. WVMS4 is based on the WVMS2 antenna and mechanical parts but includes an improved simplified RF front end, a room temperature LNA and an Acqiris AC240 digitizer with Fast Fourier Transform firmware that creates a 16,384 channel spectrometer. This large increase in the number of channels available for the retrieval allows for better instrumental baseline characterization and thus an increase in the altitude range over which retrievals provide useful information.
 Long-term measurements require a stable, maintainable instrument. To that end WVMS4 has been designed to be robust and to require minimal maintenance. It is also constructed primarily with commercially available components.
2.1. Measurement Technique
 WVMS4 detects the 616-523 H2O transition line at 22.235 GHz. The principle of operation of millimeter wave receivers is well documented by Parrish et al. , among others. The instrument uses a flat mirror positioned at 45° to the axis of rotation, which is defined by the center of the antenna beam, to direct the incoming signal from the desired direction.
 The measurement cycle begins with a “tipping measurement,” during which the sky is scanned at 29 discrete angles. The 11 angles from 45° to 75° above the horizon are used in the calculation of the tropospheric optical depth near 22 GHz. The additional 18 angles are used to diagnose pointing issues, antenna issues, absorber bar problems, and other parameters that are helpful to determine failure points remotely.
 Following the tipping measurement, the instrument enters a noise-balancing routine. The mirror alternates between two positions so that the beam is first pointed at a predetermined reference angle near zenith containing both sky and a reference absorber bar at ambient temperature. The ambient temperature reference absorber bar, developed in 2002, consists of a 12 × 65 cm piece of AN74 mounted on a 9 × 61 cm plank of 1.2 cm thick poplar wood. It is positioned 1 m from the antenna at 45° to the incident radiation in two axes to reduce reflections, as shown in Figure 1. The instrument then determines a low-elevation (signal) angle at which the total power of the measurement is similar to the total power of the measurement at the reference angle. The elevation angle of the reference position is chosen to give a signal angle of at least ∼20° above the horizon when the tropospheric opacity is low, ensuring that the antenna beam does not intersect the ground in the signal position. We note that, while it is possible to balance the system by fixing the signal angle and varying the reference angle instead, this has proven to be unsatisfactory because variations in the position of the reference absorber in the antenna beam can result in changes to the instrumental baseline.
 The measurement scan then begins using this combination of signal and reference angles. Ten signal-reference pairs are collected (a subscan) before the balancing routine is run again. Five subscans constitute a full scan, after which the data are recorded. The total integration time of a full scan is about 18 min. The measurement cycle consists of four full scans, each taken at slightly different combinations of two antenna positions and two reference angle positions. Between full scans the antenna is moved along its axis by 1/4 wavelength, i.e., 3.378 mm at 22.2 GHz, in order to both monitor and allow for the cancelation of any standing waves which may occur between the antenna and any external structure in the beam (most importantly the reference absorber) as shown in Figure 2. In addition, the reference angle is generally alternated between two angles after each pair of 1/4λ antenna movements. This allows us to monitor the sensitivity of the measurement to the position of the reference absorber bar in the beam.
2.2. Front End
 The instrument stand is constructed of 6 mm aluminum square tubing and is rugged enough to withstand mild earthquakes and the very high winds which are common at Table Mountain. Since the primary goal of these measurements is to detect trends in water vapor, it is extremely important that the positions and pointing of the antenna and mirror remain stable over a period of years. If the true pointing is ∼1° closer to the horizon than is assumed in the retrieval, then retrieved mixing ratios at all altitudes will be ∼7% too large at a typical angle of measurement due to the resultant error in the assumed air mass. Therefore, the instrument has been designed to be free standing, bolted only to the floor, and not in any way connected to the wall. Any variation in this stability can be easily detected with a precision inclinometer placed anywhere on the stand. The stand is 1.2 m tall and cantilevers the center of the mirror 80 cm from the building to keep the edge of the roof out of the antenna beam.
 The mirror is a 56 × 38 cm ellipse constructed of a lightweight honeycomb core aluminum panel and mounted at a 45° angle to the axis of rotation. The mirror is rotated by an Empire Magnetics U42 stainless steel waterproof stepping motor, which is required to hold the mirror stationary to within 0.1° in 110 km/h wind. This motor has been designed to survive in the harshest of weather conditions including being buried in snow and ice at Table Mountain. It is operated by a Parker Gemini 8-A motor driver which is set to 20,000 counts per revolution, giving an angular resolution of 0.018°. A Parker 6K2 two-axis controller interfaces the motor to the computer and provides feedback from the encoder and home sensor. The mirror is brought to the home position marked by the motor's internal home sensor every four scans to ensure precise movement. The antenna and rotating mirror are aligned using a helium neon laser mounted on the rear of the horn with a custom fixture, and alignment is verified through 360° of rotation. During some seasons it is possible to use the sun as a target to check the alignment of the system, but the utility of these measurements is limited since they can only be used to infer errors along the axis of rotation of the mirror.
 The antenna is a corrugated feed horn with a 15 cm aperture and 43 cm in length producing a gain of ∼25 dB and a FWHM, or 3 dB, beam width of ∼8°. It was manufactured by hand in the early 1990s, and due to its length and conical nature consists of five segments. These segments pose a problem for long-term measurements, since good electrical contact between them is essential to maintain the beam shape [DeWachter et al., 2009]. Water can work its way into the seams between the segments and leave behind dirt and mineral deposits. These deposits can build up over time and separate the segments, causing artifacts in the data. Care is taken to cover each segment with silicon tape, and they are periodically taken apart and cleaned. Heat tape is used to heat the entire horn above the ambient temperature. The horn is also insulated with fiberglass and covered with heavy aluminum foil.
 In previous generations of WVMS instruments the interior of the horn has been protected by using a resonant window made of two sheets of Dow Chemical's polyvinylidene chloride microwave plastic placed at an angle to the face of the horn and separated by 1/4λ. The interior of the horn was then placed under positive pressure with warm dry air passing between the plastic sheets. However, this plastic is flexible and moves slightly with atmospheric pressure, causing traveling baseline waves. While retrievals in the upper stratosphere and mesosphere were possible with such a covered feed horn, retrievals in the midstratosphere and lower stratosphere, which use bandwidths as large as 400 MHz, were found to be significantly affected by the resonant window. At this time, both the instrument at Table Mountain and the instruments at Mauna Loa are operated without the cover; however because of insects it may not be possible to operate an instrument without a cover at Lauder.
 The antenna, RF chain, programmable attenuator, and spectrometer all ride on a carriage mounted on cross roller bearing Micro Slides. This carriage is coupled to the frame via a Parker 404 motion table with home and limits proximity sensors. The table is driven by a Parker OS series U23 stepper motor giving the carriage motion control to better than 25 μm. The carriage is moved 3.378 mm (1/4λ) between each full scan. This motor is driven by a Parker Gemini 5-A driver and controlled by the same Parker 6K2 controller as the outdoor mirror motor.
 The instrument is controlled using custom software and a custom built industrial PC. In addition there is an auxiliary box containing the following three USB devices. The noise diodes are operated via relay by a Measurement Computing (MC) Switch and Sense USB device. The indoor, outdoor, bar, and hot load temperatures are measured with YSI 46040 glass thermistors. These are attached to an MC USB Temperature device which converts resistance to temperature. The third USB device, an MC 1608, records voltages produced by the Macro Sensor's Linear Variable Differential Transformer (LVDT) that precisely measures the position of the translation table. The WVMS4 instrument can be accessed, powered on/off, and operated remotely. An external Internet-accessible camera equipped with motion detection monitors the viewing angles and mirror positions.
 The analog signal received by the horn travels through an RF/IF front end with two down conversions before being digitized by the spectrometer, as shown in Figure 3. The instrument is calibrated after each subscan (five times per scan) using a Noise Com 5142 noise diode as a reference. This noise diode injects 15 dB (ENR) noise at 22.235 GHz. The noise output variation with temperature is less than 0.01 dB/°C and the noise output variation with voltage is less than 0.1 dB/1% ΔV. The noise diode is coupled into the signal using a Flann Microwave 20 dB broadwall coupler. There is a second noise diode for redundancy, and the two diodes are compared every four scans. The signal travels though the couplers into a Miteq AMFWW-SF-21982248-16-10P room temperature amplifier with a noise figure of 1.3 dB at 22.235 GHz, a passband of 500 MHz, and gain of 35 dB, as shown in Figure 3. This is followed by a 500 MHz band-pass filter which removes the lower sideband. The RF signal then passes to the first IF mixer and preamp, Spacek Labs P22-1.25 with another 25 dB of gain. The first LO frequency is 21.26 GHz which is provided by a Miteq PLDRO with a stability of ±0.01 ppm. This mixes the line center frequency down to 975 MHz. The cabling between the amplifier, oscillator, and the first IF mixer is critical for stable measurements.
 The 975 MHz signal then passes through another 500 MHz band-pass filter, through a Mini-Circuits ZFL-2000 20dB amplifier, and into a Mini-Circuits ZFM-2 mixer. The second LO frequency is 723 MHz, provided by a Nova Engineering NovaSource G6 which also utilizes a 10 MHz reference stable to ±1.0 ppm.
 Although the back end spectrometer measures from 0 to 500 MHz there is an artifact, further discussed in section 2.3, that occurs precisely in the middle of this range (at 250 MHz). Hence the second IF is centered at 252 MHz in order to avoid placing this artifact precisely at the center frequency of the molecular emission. The signal then passes through an antialiasing low-pass filter and into the final gain stage provided by a Mini-Circuits ZFL-2000 amplifier with 20 dB of gain. The RF front end is contained in an aluminum enclosure with power connections, providing EM shielding. The components heat the interior of the box above ambient temperature, but the box is not temperature controlled. The SMA connector on the output of the box passes the 500 MHz bandwidth signal from the antenna centered at 252 MHz with a total power of ∼6 dBm.
2.3. Back End
 The Acqiris AC240, now sold by Agilent as the U1080A-001, is an 8-bit high-speed 6U compact PCI Digitizer with onboard signal processing utilizing the Xilinx Vertex II PRO FPGA. This device is installed in an Adlink 6130R 1U compact PCI two slot chassis. An Adlink PCI to PCIX bridge occupies the second slot which allows communication with the PC. For use as a 16,384 channel spectrometer, the AC240 requires 32 Kpoint Fast Fourier Transform analyzer firmware also sold by Agilent. For this application we use the entire dynamic range of the spectrometer at a 1 Gs/s sampling rate. In this setup the spectrometer has a bandwidth of 500 MHz and a resolution of 30.52 KHz. 20,000 samples are accumulated by the unit for each mirror position. With communication, transfer time, and screen refresh, this takes less than three seconds. At the 2285 m altitude of the Table Mountain site, the FPGA temperature is maintained at about 40°C. To account for variations in room temperature, the firmware provides an internal calibration of the unit, which is performed with each measurement.
 The spectrometer is very sensitive to input power. The manufacturer suggests that care should be taken to select an input voltage range that will allow the signal to be recorded using as much of the dynamic range of the digitizer as possible. A nominal total input power of 1 dBm is preferable. To accommodate this requirement the RF signal is passed through an Aeroflex-Weinschel programmable attenuator between the IF box and the AC240. The attenuation is set to provide a total power of 1 dBm to the AC240 from the front end analog components when the instrument is installed. The attenuation is increased by 2 dB during tipping measurements, noise diode comparisons, and manual calibrations. Depending on pointing, sky conditions, and calibration sources, the total input power to the spectrometer varies slightly.
 The output of the AC240 contains several periodic artifacts. Some of these artifacts become more or less prominent depending on input power. The unit operates on a 62.5 MHz clock that is doubled to 125 MHz which produces harmonic spikes. The 16th subharmonic 7.8125 MHz is prominent, and most of the larger spikes are multiples of this value with the largest spikes at 125 MHz and 250 MHz. With the input of the unit terminated, the zero measurement produces a spectrum in several pieces with six major segments. Each of these segments has a slightly different zero level, as shown in Figure 4. Most of these patterns can be accounted for by recording the noise floor and subtracting it from the measured spectrum. A zero level is recorded during observations by increasing the attenuation on the variable attenuator to 64 dB. However, there is a step 885 kHz wide centered at 250 MHz. The size of this step changes with input power, as shown in Figure 5. Another prominent feature at the zero input level is a 122.0703125 kHz pattern which is the 1024th subharmonic of 125 MHz. This pattern persists throughout the raw data measurements. Since these patterns are nearly identical in the signal and reference positions, they should not affect the spectra used in the retrievals. Nevertheless, we have centered the spectrometer at 252 MHz so that none of these features falls near the center of the emission, since measurements from these channels very near the line center provide most of the information for the upper stratospheric and mesospheric retrievals.
3. Calibrated Spectrum
 The data product of the Acqiris AC240 is 16,384 channels of raw counts with an integration time of 18 min for both the signal and reference positions. The result of a 24 h averaged uncalibrated signal and reference spectrum is shown in Figure 6a. Each scan is calibrated using the Noise Com 5142 noise diode. This noise source is used to calibrate each channel five times per scan as well as once for every angle in the tipping measurement. The brightness temperature of a noise-balanced signal reference measurement at channel i is given by
where V(i)sig(NDon) is number of counts in channel i from the FFT in the signal position with the noise diode turned on, and V(i)ref(NDoff) is the number of counts in the reference position with the noise diode turned off.
 In order to convert the FFT counts to temperature we need to establish Tcal, the effective temperature of the noise diode. The noise diodes are themselves calibrated periodically (approximately every 2 weeks and whenever any changes are made to the instrument) with a standard hot-cold load method using a liquid nitrogen load and an ambient temperature load [Deuber and Kämpfer, 2004]. The hot load is a sheet of Eccosorb CV-3 absorber placed above the mirror and covering the entire beam. The temperature of the load is monitored in two places. The cold load is Eccosorb CV-3 submerged in liquid nitrogen in a stainless steel cylinder 10 cm deep and 33 cm in diameter. The size of the load is minimized to conserve liquid nitrogen. The cold load is placed below the mirror. The calibration then consists of comparing these two measurements of known temperature (Thot and Tcold below) to the signal strength of the noise diode. Because of concerns about standing wave reflections in these calibration measurements, this primary calibration is applied only as a single frequency-independent measurement counts to temperature conversion based on an average of the calibrations for channels near line center:
In Figure 6b we show emission spectra calibrated using (1) and (2). It is important to note that the spectra shown in Figure 6 have not had any baseline components removed.
 This calibration method does not require any estimate of either sky or receiver temperature. It does, however, depend on the stability and absence of frequency-dependent structure in the emission from the noise diode. In Figure 7a we show the value of Tcal as determined by this method at Mauna Loa for the two noise diodes on the WVMS3 system from January 2005 through March 2010. Each point represents the median value of the ∼10 hot-cold calibration measurements performed on each noise diode over a ∼30 min period. Changes to the instrument in March 2007 affected the effective temperatures of the noise diodes. In Figure 7b these values have been normalized by dividing Tcal by the median value of Tcal for that noise diode over the periods January 2005 to March 2007 and March 2007 to March 2010 separately. In Figure 7c we show (in green) the ratio of the normalized Tcal values shown in Figure 7b.
 While liquid nitrogen calibrations provide the basis of the calibration system, individual calibrations can be noisy. The ratios of the Tcal values in Figure 7c are less variable than the calibrations themselves, emphasizing that much of this variability comes not from variations in the strengths of the noise diodes, but from the difficulty of performing a consistent calibration. Therefore, by comparing the two diodes and assuming that the strengths of the diodes do not generally vary together, it is possible to distinguish variations in the calibration procedure from variations in the strengths of the diodes.
 In addition to the approximately biweekly calibration procedure with liquid nitrogen, we also compare the secondary noise diode with the primary noise diode after every two scans. This comparison consists of six full power measurements with the mirror positioned at 30°, two with both diodes off, two with the primary noise diode on, and two with the secondary noise diode on. The daily median of the ratio of the strengths of the noise diodes calculated in this comparison is also shown (in blue) in Figure 7c. The ratios of the strengths of the noise diodes calculated according to this procedure are in good agreement with the ratios calculated during the liquid nitrogen calibrations. Since noise diode comparisons are performed every two scans (as opposed to the approximately biweekly calibrations) this noise diode comparison process helps to pinpoint the exact period during which one of the noise diodes has drifted.
 Over the period January 2005 to March 2007 both noise diodes appear to be stable. Almost all of the normalized Tcal ratios fall between 0.99 and 1.01, and there is no clear trend. From March 2007 to March 2010 there are temporal variations evident in Figure 7c. First, there is a slight trend in the noise diode ratio, with ratios at the beginning of the period generally falling between 1.00 and 1.01, while those at the end generally falling between 0.99 and 1.00. Second, there are two periods in 2009 when the ratio increases from slightly below 1.00 to as high as ∼1.02 for several weeks. Comparisons between Figures 7c and 7b show that both of these variations can be attributed primarily to variability in the primary noise diode.
 Based on this comparison of over 5 years we conclude that, if this system were operated as a single noise diode system using only the primary noise diode, and if that noise diode were presumed constant, then there would be temporal calibration uncertainties of as much 2% over several weeks, and drift of as much as 1% per year from 2007 to 2010. However, since it appears from the external calibrations that the secondary noise diode is less variable then the primary noise diode, we can adjust the applied calibration (Tcal in (2)) to reduce calibration uncertainties in these measurements to within ∼1% for the entire period. The high temporal density of the noise diode comparisons is clearly crucial in order to ensure the application of an accurate correction of the more variable noise diode. We note that all of these uncertainties apply only to variations in the calibration. Systematic uncertainties from calibrations have been estimated at ∼5% [Nedoluha et al., 1995] but are probably best determined by comparisons with other measurements [Nedoluha et al., 1997, 2007].
 Since the early 1990s we have retrieved water vapor profiles from 40 to 80 km using a spectrum with 40 MHz bandwidth and 40 channels (WVMS1) or 60 MHz bandwidth and 60 channels (WVMS2 and WVMS3). The channels are of varying width, with narrow channels placed near line center as described by Nedoluha et al. . The Agilent U1080 (formerly Acqiris AC240) signal analyzer with FFT firmware produces 16,384 channels and a total bandwidth of 500 MHz resulting in a resolution of 30.52 kHz. Because of the spurious artifacts at the edges of the FFT spectrometer and dropoffs at the edges of the band-pass filters, we do not use the first and last 50 MHz of the data in the retrievals. This results in a 400 MHz retrieval bandwidth. The FFT channels used in the retrieval are averaged into bins as shown in Table 1.
Table 1. The 16384 Channels of the Spectrometer Separated Into Bins to Save Computer Timea
Number of Sets
Number of Channels
Full-resolution 30.5 kHz single channels are used in the center, where the most sensitivity is required. The outer channels are divided into sets of 67 channels, reducing the resolution to ∼2 MHz.
 The additional high-resolution bandwidth provided by the FFT spectrometer makes it possible to identify some features in the spectrum which are clearly instrumental baseline artifacts. In order to reduce these artifacts, we improved and simplified RF interconnections, cabling, and connectors. As a result the instruments now produce spectra with fewer instrumental artifacts than before, which should result in a better retrieval. With these improved spectra it is possible to detect otherwise imperceptible differences in antennas and to characterize them. One specific change which was made to reduce instrumental baseline waves was the elimination of the feed horn cover.
 While the reduction in instrumental baseline structure improves retrievals at all altitudes, it is particularly important for making use of the additional bandwidth in order to retrieve profiles below ∼40 km. The 400 MHz bandwidth used in the retrievals does allow for the possibility of retrievals down to ∼26 km, but this is only possible if the baseline structure is correctly characterized. Although the high altitude (2300 m) and generally dry conditions of JPL's Table Mountain Facility provide the low-humidity troposphere ideal for middle atmospheric measurements, stable and accurate retrievals of water vapor at altitudes below 40 km remain difficult. Lower-stratospheric retrievals are sensitive to changes in the troposphere from ∼6 km to the tropopause since the variability of water vapor in the troposphere is orders of magnitude higher than that of the stratosphere. Therefore the ability to correctly measure and characterize the tropospheric water vapor separately from instrumental baselines becomes increasingly important as retrievals are extended into the lower stratosphere.
 The tropospheric optical depth is determined by a tipping measurement taken as part of the regular measurement cycle. The tropospheric optical depth is then combined with the spectral measurement, and a water vapor profile is retrieved using optimal estimation as described by Nedoluha et al. . The retrieval is performed from the ground, in this case the altitude of the Table Mountain site (2.3 km), to 100 km. Although the retrieved tropospheric profile shapes are difficult to validate, they are not physically unrealistic, and they are retrieved along with the stratospheric profile. Nonphysical tropospheric profiles would be indicative of a possible instrumental baseline which could also affect retrievals in the lower stratosphere. Daily measurements in the upper mesosphere are limited by signal to noise, however by averaging measurements from several days it is possible to get useful retrievals up to ∼80 km. Above ∼80 km Doppler broadening becomes important, and profile information becomes limited.
 The result of a 24 h integrated retrieval is shown in Figure 8. Figure 8 (top left) shows the measured and modeled spectrum for the 400 MHz bandwidth used in the retrieval. This spectrum differs from that shown in Figure 6 only in that a linear slope has been removed from the spectrum. Also shown in Figure 8 (top right) are the measurement-model residuals in solid lines. The increase in the magnitude of the residuals near line center is a result of the larger bins into which channels further from line center are averaged (Table 1). An expanded view of the residuals near line center is shown in Figure 8 (bottom left). Included in this retrieval is the fit to a single sine wave baseline term. The baseline term shown in Figure 8 maintained a similar shape for the several months when the instrument was in this particular configuration. Analysis of the residual allows us to monitor spectral features that cannot be seen in the raw spectra and is extremely valuable as a tool for identifying reflections in cabling, problems with waveguide placement, connector failures, and antenna characterization. The residual shown in Figure 8 represents a significant improvement over that shown for 28 January 2009 by Nedoluha et al. . This improvement is a result of the replacement of a failing connector between the first local oscillator and the first IF mixer. Finding the sources of baseline artifacts and stabilizing or eliminating them is the key to accurate long-term measurements. By reducing the sources of baseline artifacts that are clearly instrumental, we can hopefully also minimize those baseline terms which can affect the retrieval. Although fits for instrumental baselines can be, and generally are, included in the retrieval process, it is often difficult to distinguish between atmospheric and instrumental baseline components, and this adds uncertainty to the retrievals, particularly in the lower stratosphere. It should also be emphasized that, if baseline fits are included as part of a retrieval process, they should be included as part of the atmospheric retrieval (and not as part of a preprocessing step) in order to characterize their effect on the sensitivity of the retrieval.
Figure 8 (bottom middle and bottom right) also shows the a priori and retrieved water vapor profiles. The a priori tropospheric profile is determined by matching the optical depth of water vapor calculated from the tipping measurement with a 2 km–scale height water vapor profile. As is shown in Figure 8 (bottom middle), the retrieved water vapor profile in the troposphere seems to be physically realistic, and hence there is no indication that an instrumental baseline term is being matched by a nonphysical tropospheric retrieval.
 The fourth-generation instrument presented here will allow us to extend long-term trend measurements down to ∼26 km. Although every effort has been made to reduce and eliminate baseline artifacts, the stability of the remaining instrumental baselines is critical to making long-term measurements. The experience gained from our previous generations of long-term instruments has allowed us to create a high-performance, high-stability instrument on a proven platform. By minimizing the number of parts and using commercially available components we have reduced maintenance, down time, and failure rate. This reduces physical interaction with the instrument and improves stability. Using not only the data spectra but also the retrievals as tools in developing the hardware to optimize the performance of the instrument has resulted in an extremely stable and robust instrument for long-term measurements.
 An additional fourth-generation instrument (WVMS5) with a new antenna design [Teniente et al., 2011] has been built and was deployed alongside WVMS3 on Mauna Loa Observatory for validation. Measurements from these two instruments and the Microwave Limb Sounder (MLS) were extremely helpful to properly calibrate the transition as WVMS3 was replaced by a fourth-generation instrument (WVMS6). After a lengthy period of side-by-side measurements at Mauna Loa, we hope to eventually deploy this instrument to another high-altitude site in the Southern Hemisphere. WVMS1 in Lauder, New Zealand has also been retired and replaced with a fourth-generation instrument (WVMS7) which is currently deployed.
 The transition between third- and fourth-generation instruments represents an unusually large step; however there is more that can be done to further improve these instruments. As the baselines resulting from other aspects of the instrumentation were reduced, the differences in our antenna inventory became more apparent. Further investigation has shown asymmetrical near field effects resulting from having an absorber bar in only part of the antenna beam at the reference angle. These effects present themselves as waves in the spectrum which can affect the retrieval. The current deployment of two fourth-generation instruments (WVMS5 and WVMS6) colocated at Mauna Loa Observatory has provided a unique opportunity to investigate methods to mitigate this issue, to characterize antennas, and to investigate any instrumentation issues that may affect long-term measurements. The lessons learned will be applied to future WVMS instruments.
 We wish to thank the staff at the Table Mountain Facility and the Mauna Loa Observatory for their assistance and support. This work is sponsored by NASA under the Upper Atmosphere Research Program and the Naval Research Laboratory.