Discrimination of Light Precipitation from COWVR + TEMPEST, GMI, and ATMS Passive Microwave Radiometer Observations
1. Introduction
In the near future, the number of passive microwave (MW) radiometric observations available from low-Earth orbiting (LEO) is expected to increase substantially. As depicted in the timeline in Fig. 1, the majority of these observations will originate from passive MW cross-track scanning high-frequency (HF) sounders (channel bands near or exceeding 90 GHz), with many expected contributions from SmallSat- or CubeSat-sized sensors. This increase will reduce overall revisit time, enabling advancements for applications in the tracking of convective cloud tops (Brogniez et al. 2022). For quantitative precipitation estimation, the passive MW sensors that include the preferred constant-incidence angle, dual-polarization low-frequency (LF) channels (bands near or below 37 GHz) such as the lengthy (1987–current) Special Sensor Microwave Imager (SSM/I) and SSM/I Sounder (SSM/IS) have remained steady or declined. Frequencies near 10 GHz are particularly underrepresented (Kidd et al. 2021a).
Timeline from 1978 to 2023, showing the total number of passive MW sensors, broken apart by those with channel sets ranging 6–37 (dark blue), 6–89 (light blue), 10–183 (green), 89–183 (gold), and 183 GHz only (red).
Citation: Journal of Atmospheric and Oceanic Technology 42, 6; 10.1175/JTECH-D-24-0061.1
Physically, the use of HF-only sensors for precipitation skews the derived precipitation estimates toward indirect, ice-based (scattering) retrievals rather than the precipitation closer to the surface. Over certain Earth surfaces, a significant portion of light (e.g., under 2 mm h−1 intensity) precipitation may go undetected (Kidd et al. 2021b). This has ramifications for constellation-based global precipitation products (Behrangi and Song 2020), such as the Japanese Aerospace Exploration Agency (JAXA) Global Satellite Mapping of Precipitation (GSMaP) (Ushio et al. 2009) and the National Aeronautics and Space Administration (NASA) Integrated Multi-satellitE Retrievals for Global Precipitation Measurement (GPM) (IMERG) (Huffman et al. 2020). This need to represent LF passive MW sensing capabilities in the global precipitation constellation and a benchmarking activity to properly assess the impact of the mixture of LF + HF sensors on global precipitation products are key recommendations from the recent 11th International Precipitation Working Group (IPWG) workshop (Kubota et al. 2025). This study investigates precipitation detection capabilities when the HF-only observations are augmented by SmallSat-sized, conically scanning passive MW radiometers providing LF observations.
The Compact Ocean Wind Vector Radiometer (COWVR) and the Temporal Experiment for Storms and Tropical Systems (TEMPEST) SmallSat duo were delivered to the International Space Station (ISS) in December 2021 and situated onboard the Japanese Experiment Module (JEM) for 3-yr operations. COWVR is a conically scanning passive MW polarimeter providing fore- and aft imaging at 18.7, 23.8, and 33.9 GHz (full Stokes parameters), similar to the WindSat polarimeter (Brown et al. 2017). The COWVR design and primary role are to estimate ocean surface wind speed and direction. COWVR is based on instrumentation originally developed for the Advanced Microwave Radiometer for Climate (AMR-C) used for water vapor corrections to sea surface height estimates (Maiwald et al. 2020). TEMPEST provides near-simultaneous cross-track MW sounding at five channels between 89 and 182 GHz (Schulte et al. 2020), similar to channels from the existing Advanced Technology Microwave Sounder (ATMS) currently orbiting on the National Oceanic and Atmospheric Administration (NOAA) operational platforms. Although not identical, this sensor duo provides a channel and viewing angle diversity that would be approximated if several ATMS channels were cojoined with selected channels from the GPM Microwave Imager (GMI), developed and operated jointly by NASA and JAXA. Table 1 provides a cross reference.
Cross-reference of similar channels from COWVR, GMI, TEMPEST, and ATMS. All channels in units of gigahertz. On-Earth resolution in kilometers is provided in (cross track) × (along track) direction from ISS, GPM, and NPP altitudes of 400, 407, and 825 km, respectively. V (H) = constant vertical (horizontal) polarization; QV (QH) = quasi-V (H) polarization skews toward V (H) away from nadir. ATMS channels 8–15 are not included. All six COWVR channels provide both a fore and aft view of the same scene. The 10 cells in bold font indicate the GMI and ATMS channels that have a similar corresponding channel on COWVR or TEMPEST.
Precipitation detection is the initial step for passive MW precipitation estimates provided to techniques such as GSMaP and IMERG, which use motion-based techniques (i.e., Kalman filtering) to transport the precipitation pattern at time 1 (t1) to the passive MW precipitation pattern at time 2 (t2). The time interval (t2 − t1) is on the order of 30 min, based upon current geostationary satellite scan schedules. Since precipitation is a discontinuous quantity, the detection step introduces uncertainty. Detection of precipitation in the MW regime involves discriminating its observed equivalent blackbody brightness temperature (TB) against similar TB from nonprecipitating scenes (Ferraro et al. 1998), owing to the variable surface emissivity. An analogy is made to the detection of clouds from multispectral visible-to-infrared (IR) satellite imagers, when reflected sunlight (or emitted thermal radiation) signatures appear similar to bright (or thermally cold) Earth surfaces. However, over many surfaces, the variable surface emissivity is generally not known at the time of the satellite observation. In reality, the surface emissivity is rarely homogeneous; rather, it encompasses smaller scale variability, such as mixed water/land or different soil and vegetation conditions. Over certain types of vegetation, the surface emissivity within the skin depth in the 10–37-GHz range (upper 2 cm or less) is dependent upon antecedent precipitation conditions (You et al. 2015; Ringerud et al. 2021). Numerous studies since SSM/I data were made available have investigated techniques to estimate and map global emissivity patterns (Prigent et al. 2006). A recent survey article contributed by the GPM Land Surface Working Group (LSWG) (Turk et al. 2021) provides background and cross references to several recent studies of the microwave surface emissivity.
In this study, the capability of the GMI, COWVR + TEMPEST, and ATMS observations is presented for purposes of discriminating precipitation over various Earth surface types, including mixed water/land (snow-covered surfaces are excluded owing to limited data). Precipitation is discriminated using a surface emissivity–based transformation applied to selected channel combinations from each of these sensors, described in the appendix. With its ±65° latitude global coverage, the GPM dual-frequency precipitation radar (DPR) vertically resolved precipitation structure from the DPR-only and DPR + GMI [Combined Radar–Radiometer Algorithm (CORRA)] algorithms is used as reference near-surface precipitation rates to quantify detection metrics such as false alarm rate (FAR)1 and probability of detection (POD). These metrics are calculated at three threshold rates (0.5, 1, and 2 mm h−1) and different LF and HF channel combinations for each sensor. From each of these, relative operating characteristic (ROC) curves are derived that determine how effectively the precipitating and nonprecipitating observations can be discriminated over different surface types.
While this study examines only the discrimination of precipitation, a forthcoming investigation will present results from the COWVR + TEMPEST precipitation profile estimation, relating the precipitation profile to the ocean surface wind divergence in the surrounding environment.
2. Description of datasets
a. Passive MW datasets
COWVR and TEMPEST temperature sensor data records (TSDRs) data were obtained from the NASA Physical Oceanography Distributed Active Archive Center (PO.DAAC), containing level 1 geolocated and calibrated equivalent blackbody TBs. For this investigation, 22 months of data (January 2022–October 2024) were used (prior to the GPM orbit boost in early November 2024), whenever COWVR and TEMPEST were operating simultaneously. COWVR acquires data on a portion of its forward and aft view of its conical scan at an ∼50° Earth incidence angle, whereas TEMPEST scans cross track with an approximate 1400-km swath, both from the ∼400-km ISS altitude. To facilitate their use, these data were coregistered using a nearest-neighbor approach and merged into a format similar to the level 1C format used for the GPM X-Cal effort (Berg et al. 2016). The forward scan of COWVR is used as the common basis for scene matching of the TEMPEST data, and the associated COWVR aft-scan data occur anywhere from 2 to 5 min later, depending upon the location relative to the satellite subpoint. This results in a common 17-channel (six channels for each of the COWVR fore- and aft view, and five TEMPEST channels) dataset covering an ∼900-km swath. ISS operations periodically obstruct COWVR and/or TEMPEST views, and the two sensors are not synchronized nor designed to be operated simultaneously. Only data that pass the sensor quality flag are used, which leaves occasional gaps within the sensor swath and at the swath edge. While only the fore-view COWVR observations were assessed in this study, fore- and aft-viewing COWVR observations were included for future investigations. Hereafter, the term “C+T” is used as a shorthand for the COWVR + TEMPEST sensor duo.
GPM GMI and National Polar-orbiting Parternship (NPP) ATMS level 1C (level 1C-R for GMI) TB products and precipitation profiles from the GPM level 2A DPR and level 2B combined (DPR + GMI) radar-radiometer (CORRA, often denoted as CMB) (all version 7) data products were obtained from the NASA Precipitation Processing System (PPS). While ATMS is a 22-channel sensor, the level 1C dataset provided by PPS contains only nine channels. Five 50-GHz band temperature sounding channels were appended from the PPS level 1 Base file that are sensitive to surface and precipitation (between 50.3 and 54.4 GHz), for a total of 14 ATMS channels. Since the ISS orbit inclination (51.6°) restricts the coverage of the C+T observations, analysis is restricted to ±55° latitude to equalize the environmental conditions sampled by all three sensors. Table 1 summarizes the sensor resolution and polarization capabilities and highlights similarities between the C+T channels and corresponding channels on GMI and ATMS.
b. DPR precipitation reference
Conditional statistics are stated relative to a reference precipitation dataset. A common reference is precipitation estimated from ground radar or rain gauges, which generally cover land surfaces and are available for limited regions. Due to the global nature of this study, the reference sources are chosen to be the DPR-only Ku-band precipitation product (Seto et al. 2021), and the combined CMB Ku-band precipitation product (Grecu et al. 2016), to provide a common global reference covering all seasons and surface conditions in the GPM coverage area. Each of these algorithms produces a similarly structured 250-m vertical resolution precipitation profile, but using different assumptions on the drop size distribution (Liao et al. 2020). Hamada and Takayabu (2016) determined that the low-end sensitivity of the Ku-band radar of near 18 dBZ. For this study, a low-end precipitation sensitivity of 0.5 mm h−1 is assumed for both of these products.
For GPM, 6 months (approximately 2500 orbits) randomly chosen between 2015 and 2022 are used. For C+T and ATMS, the same procedure was carried out for all near-time coincident observations (maximum allowable time separation of ±15 min) with the GPM DPR. ATMS data from 2019 to 2022 were used, providing ∼24 M coincident DPR + ATMS pixels. While only a 22-month period for C + T was available, this period did supply ∼16 M coincident DPR and C+T pixels.
While each channel has different resolutions, there is no single resolution of a multichannel radiometer-based precipitation estimate. In general, for shallow precipitation or precipitation lacking a developed ice phase, the precipitation signal is mainly carried in the lower (coarser resolution) frequencies near or below 37 GHz. For precipitation with a developed ice phase, the precipitation signal is mainly carried in the higher (finer resolution) frequencies near and above 90 GHz. For each GMI pixel, the DPR profiles and near-surface precipitation rate are averaged to the approximate resolution of the 89-GHz channel (a 3 × 3 set of DPR pixels centered at the location of the 89-GHz channel). For ATMS and C+T, the DPR was averaged to the resolution of the 90-GHz channel (Table 1). Since these sensors scan cross track, the number of DPR pixels averaged increases as the scan position [and associated on-Earth field of view (FOV)] increases away from nadir.
3. Surface emissivity–based discrimination
A description and flow diagram of the emissivity principal component (EPC) precipitation discrimination and profile estimation are previously reported in Turk et al. (2018) and Utsumi et al. (2021). For ease of review and understanding, the discrimination step is summarized in the appendix. At the center of this method is a basis transformation which first transforms the desired set of TB observations into a set of orthogonal normalized EPC terms. A discriminant analysis separates precipitating or nonprecipitating observations at a desired precipitation rate threshold and desired FAR and/or POD levels. In the material to follow, unless otherwise indicated, the optimal discriminant is defined as the value of the discriminant that maximizes the threshold skill score (TSS),2 where TSS = POD − FAR. A value of TSS = 1 indicates a perfect score; a value of 0 indicates no skill (Manzato 2007).
For practical use with passive MW algorithms, the discrimination technique is illustrated with a C+T example using two channel combinations: the full 11-channel C+T and the five-channel TEMPEST only. The top row of Fig. 2 shows a segment of the C+T sensor swath, offshore of Japan at 0331 UTC 30 September 2022. The ISS and the GPM spacecraft orbited over this area within an ∼6-min time offset, enabling the precipitation-discriminated region to be compared to the DPR reference precipitation (within the overlapping regions of the 245-km DPR swath). Figure 2d shows the near-surface precipitation estimated by the Ku-band DPR algorithm.
Example from the C+T radiometer for an overpass just offshore of Japan at 0330 UTC 30 Sep 2022. (a) COWVR 33.4-GHz H-polarized TB, (b) TEMPEST 165-GHz TB, (c) corresponding GMI 166-GHz H-polarized TB, and (d) near-surface precipitation produced by the DPR Ku-only precipitation data product. (e) The value of the discriminant function for the 11-channel C+T combination (cowvr6F_tempest5) at a 0.5 mm h−1 precipitation rate threshold; (f) as in (e), but for the 5-channel TEMPEST-only combination (cowvr0_tempest5); (g) precipitation screen (blue = precipitation; red = no precipitation) that results for a desired FAR < 0.05 discrimination for the 11-channel C+T combination; (h) as in (g), but for the five channel TEMPEST-only combination.
Citation: Journal of Atmospheric and Oceanic Technology 42, 6; 10.1175/JTECH-D-24-0061.1
Figures 2e and 2f depict the discriminant function at the 0.5 mm h−1 threshold. Each color table is normalized (center white color) to the value of the optimal discriminant for this threshold rate. Increasingly blue (red) implies that the observation is further to the left (right) of the optimal discriminant (black vertical line in Fig. A6), implying higher (lower) certainty of precipitation exceeding this level. At the optimal discriminant setting, the 11-channel combination would result in values of POD = 0.882 and FAR = 0.113. The TEMPEST-only combination would result in a slightly worse POD = 0.792 and FAR = 0.182.
To intercompare the detected precipitation area for these two channel combinations on an equal basis, the discriminant operating point is set to the FAR < 0.05 level for both. Pixels that are flagged as precipitating will have a FAR that is lower than this threshold. Figures 2g and 2h depict the associated precipitation screen (blue = precipitation; red = no precipitation) for these two different channel combinations. In this example, both screens capture the main heavy precipitation regions, but much of the surrounding light rain is missed by the HF-only five-channel combination (Fig. 2h). This example highlights the area of light rain that is missed when HF-only passive MW sounders are used to detect precipitation overocean surfaces.
4. Discrimination using selected sensor channels
This section expands upon the above example with overall skill scores from different sensors, covering the entire dataset period. The precipitation discrimination performance from C+T is first presented, after which these same characteristics are compared to those from GMI and ATMS.
a. COWVR + TEMPEST observations
For the full 11-channel C+T channel set, there are seven channels at or below 89 GHz (the aft-viewing channels are not used), which is considered the highest window channel frequency where the surface is still partially transparent and from which the surface emissivity can be estimated from the nonprecipitating observations. For C+T, nine channel combinations were examined. For example, combination 0 uses all 11 channels, 7 of which are <90 GHz. From Eq. (A1), therefore, the emissivity vector for this combination is length 7 + 3 = 10.
The EPC state vector does not use any surface-type characterization; it inherently clusters around specific types of Earth surface conditions (Figs. 9–12 in Turk et al. 2021). Therefore, these analyses are presented for four surface classes determined by the value of the Tool to Estimate Land Surface Emissivities at Microwave frequencies (TELSEM) index (Aires et al. 2011), which is used and included in the GPROF precipitation data product used for GPM (GPROF 2022). GPROF is designed around an 18-class version of TELSEM. For this investigation, ocean is class 1; high (low) vegetation land are classes 3–4 (5–7); and mixed coasts are classes 13–15. Snow-covered surfaces (classes 8–11) are not included in this investigation due to the limited number of events observed. In general, at frequencies < 37 GHz, heavily vegetated surfaces have a more stable surface emissivity in the MW range near 0.9 with little polarization signature, relative to the emissivity of lightly vegetated surfaces which can be lower, more variable, sensitive to soil moisture, vegetation water content, and more polarized (Prigent et al. 2006; Turk et al. 2014; Ringerud et al. 2014).
Figure 3 depicts the ROC curves for eight different channel combinations at the CMB 1.0 mm h−1 threshold level. Each combination is assigned a unique line color, corresponding to a row of the color index table at the bottom of Fig. 3. In each panel, the channel combinations are listed in the descending order of their respective TSS.
ROCs for precipitation discrimination using nine different combinations of the 11 C+T channels at the 1.0 mm h−1 threshold level. Each combination is assigned a unique line color, indexed according to the key below. The legend ranks the channel combinations in decreasing order of TSS performance. The horizontal and vertical dotted lines indicate the FAR and POD values of the top-ranked performer. Each panel represents the observations separated by the TELSEM index: (a) overwater, (b) mixed water/land, (c) lightly vegetated land, and (d) heavily vegetated land.
Citation: Journal of Atmospheric and Oceanic Technology 42, 6; 10.1175/JTECH-D-24-0061.1
For ocean surfaces, nearly all channel combinations perform with TSS scores near 0.85 or higher, with the lowest score assigned to TEMPEST only (combination 6) and the highest to the C+T combination with H-polarized LF channels (combination 5). Over lightly vegetated surfaces, the most substantial performance degradation occurs for the five GMI-like COWVR channels plus 89 GHz (combination 3) or when the (lower emissivity) H-polarized COWVR channels are removed (combination 4). Overall detection performance is worse over heavily vegetated and mixed coastal surfaces, most degraded for combination 3. In general, for precipitation detection over ocean and even heavy vegetation surfaces, the value of a joint LF + HF capability is evident. Over heavy vegetation, one explanation is that vegetation canopies can often be wet with a high vegetation water content. Water content implies an absorption/emission process, which implies that LF channels (whose signal arises deeper into the canopy) might aid the HF channels in discriminating precipitation over wet dense vegetation.
Over mixed coast and lightly vegetated surfaces, TEMPEST only (combination 6) is competitive with many of the LF + HF channel combinations. One possible explanation is the difference in viewing geometry (constant incidence for COWVR, variable for TEMPEST) and different azimuthal viewing directions across each scan. These may mask the impact of the LF channels across these types of surface conditions whose polarized emissivity signature responds to the onset of precipitation (Turk et al. 2014; Ringerud et al. 2014). However, further investigation is needed before drawing either of these conclusions.
Figure 4 summarizes the optimal TSS scores for each of the nine channel combinations and each of the surface types (horizontal axis) using the CMB reference. For each of these combinations, there are three adjacent vertical lines representing (left to right) the optimal TSS for the discrimination at a 0.5, 1, and 2 mm h−1 precipitation rate threshold level. As expected, the TSS increases as the precipitation threshold level increases. For overocean and for heavy vegetation surfaces, the various combinations begin to converge to similar performance, with the exception of the TEMPEST only (combination 6) over ocean. Over light vegetation and mixed water/land (labeled as coast), the value of the HF channels is evident. The TEMPEST-only combination performs slightly better than the C+T combinations where either the 18- or 23-GHz COWVR channels are removed (combinations 7 and 8).
(top) Summary of optimal TSS scores for the C+T observations across the four surface conditions using the CMB-only precipitation as a reference. The channel combination index is defined by the color index key in Fig. 3. For each channel combination, the three adjacent vertical lines correspond to a CMB precipitation rate threshold of 0.5, 1, and 2 mm h−1, respectively. (bottom) Difference between the optimal TSS score using the CMB and the DPR-only precipitation as the reference source (negative value implies that DPR-only provided larger TSS).
Citation: Journal of Atmospheric and Oceanic Technology 42, 6; 10.1175/JTECH-D-24-0061.1
Over heavy vegetated surfaces, the performance increase with threshold rate is most evident, owing to gradually increasing scattering-induced TB depressions that provide more contrast with the higher TB under nonprecipitating observations. There is relatively little difference whether V- and/or H-polarized channels are used (combinations 4 and 5), owing to the largely unpolarized surface emissivity of heavy vegetation.
These results have used the CMB precipitation dataset as a reference. The bottom panel of Fig. 4 shows that these TSS scores would be slightly higher if the DPR-only precipitation dataset was used as the reference, notably over vegetated land surfaces. Since the procedure to build the DPR + MW radiometer datasets (described above) was done identically for both data products, this suggests a systematic difference between the DPR and CMB products at the light precipitation end. The implication to this study is that differences in the TSS of ∼0.05 arise from the choice of reference data alone. While the reason for this discrepancy is beyond the scope of this investigation, it is consistent with earlier findings (i.e., DPR outperforming CMB in precipitation detection) reported by Li et al. (2023) using ground-based GV-MRMS (Kirstetter et al. 2012) data over the continental United States.
b. GMI observations
For the full 13-channel GMI channel set, there are nine channels at or below 90 GHz, for which the surface emissivity can be estimated from the nonprecipitating observations. Figure 5 shows the relative operating characteristics for each of the seven surface classifications in a format similar to Fig. 3. For ocean, channel combinations 0, 1, and 6 (combination 6 approximates the channel set on the future AMSR-3 sensor) perform with TSS scores near 0.9, implying limited benefit for detecting precipitation at this level with the added 166 GHz and higher channels. For GMI, each channel views at nearly the same Earth incidence angle (EIA), whereas the C+T channel set is partially conical scanning (like GMI) and partially cross-track scanning (like ATMS). However, comparisons between Figs. 3 and 5 suggest that even with this effect, the C+T duo can perform similarly to GMI at this level of precipitation detection.
As in Fig. 3, but for seven different combinations of the 13 GMI channels at the 1.0 mm h−1 threshold level.
Citation: Journal of Atmospheric and Oceanic Technology 42, 6; 10.1175/JTECH-D-24-0061.1
Over the other surfaces, GMI discrimination at this level drops off considerably when only the COWVR-like channels (combination 5) or only the TEMPEST-like channels (combination 4) are used. Over light vegetation, the substantially lower performance of the HF-only (89 GHz and higher) combination is in contrast to C+T, where the TEMPEST-only combination exhibited the highest performance. One possible reason is that the dual-polarized LF channels and a conical view angle for LF and HF channels better capture the polarized emissivity signature over many lightly vegetated surfaces and how this changes after precipitation events. In general, for GMI, the overall discrimination is improved when LF + HF channels are used together.
When the DPR precipitation reference is used instead of CMB (bottom of Fig. 6), the same trend as noted in the COWVR analysis (bottom of Fig. 4) was noted, with the resulting negative trend in the TSS score more equally distributed across the surface types.
As in Fig. 4, but for the seven GMI channel combinations.
Citation: Journal of Atmospheric and Oceanic Technology 42, 6; 10.1175/JTECH-D-24-0061.1
c. ATMS observations
For the 14-channel ATMS, there are four channels at or below 89 GHz from which the surface emissivity was estimated from nonprecipitating observations. These are at 23.8, 31.4, 50.3, and 88.2 GHz. For ATMS, eight channel combinations were examined. For example, combination 0 represents the nine channels used by the GPROF GPM precipitation data product, three of which are <90 GHz. From Eq. (A1), therefore, the emissivity vector for this combination is length 3 + 3= 6. Combinations 3–7 gradually increase the number of sounding channels from the 50-GHz band, whose associated weighting function gradually peaks at an increasingly higher altitude. Figure 7 shows the relative operating characteristics for each of the six surface classifications in a format similar to Fig. 3. For ocean, coast, and lightly vegetated scenes, the most notable feature is the substantial performance degradation when the LF (23 and 31 GHz) channels are omitted and the five TEMPEST-like channels are retained (combination 2). This characteristic is similar to what was noted with the conically scanning GMI.
As in Fig. 3, but for eight different combinations of the 14 ATMS channels at the 1 mm h−1 threshold level.
Citation: Journal of Atmospheric and Oceanic Technology 42, 6; 10.1175/JTECH-D-24-0061.1
For heavily vegetated surfaces, combination 7 (all 14 channels) substantially improves the detection at this 1 mm h−1 level, relative to combination 0. This feature is likely related to the scattering signature that is more notable at the (higher weighting function altitude) 52.8-GHz channel, when convective precipitation develops within a warm, moist air mass. To provide an example, Fig. 8 shows a NOAA-20 overpass on 18 May 2020, where the ATMS 50.3-, 51.76-, and 52.8-GHz channels gradually pick up the precipitation ice scattering signal (TB depressions against the radiometrically warm surface) as the surface gradually becomes more opaque. For lightly vegetated surfaces, the 50.3-GHz surface channel provides added diversity to the surface emissivity state, being midway between the 31.4- and 88.2-GHz window channel bands. Over mixed water–land scenes, a slight improvement occurs when the three TEMPEST-like channels are used in the 183.31-GHz water vapor band (combination 1) rather than all five (combination 0). This same result shows up for all three precipitation rate thresholds plotted in Fig. 9. This could be related to water vapor gradients near coastlines and how these are manifested in the spectral signature within the 183-GHz absorption complex, but it requires further investigation to state conclusively. As noted with COWVR and GMI (bottom of Figs. 4 and 6, respectively), for ATMS, the negative trend in the TSS score (bottom of Fig. 9) is present when the DPR precipitation reference is used instead of CMB.
NOAA-20 overpass at 1815 UTC 18 May 2020 during a period of flooding in lower Michigan that occurred after a 4-day period of steady rain. ATMS channels at 31.4, 50.3, 51.76, and 52.8 are shown from left to right.
Citation: Journal of Atmospheric and Oceanic Technology 42, 6; 10.1175/JTECH-D-24-0061.1
As in Fig. 4, but for the eight ATMS channel combinations.
Citation: Journal of Atmospheric and Oceanic Technology 42, 6; 10.1175/JTECH-D-24-0061.1
5. Practical use for discrimination across land–water boundaries
This analysis has indicated that a joint LF + HF capability is desirable for discriminating precipitation across ocean surfaces. The impact over mixed water–land boundaries is demonstrated with the example in Fig. 10.
As in Fig. 2, for a C+T overpass just offshore of eastern Canada at 1808 UTC 2 Jun 2023. In this case, the second discriminant is the 9-channel C+T combination 7 (18-GHz channels are removed) denoted by cowvr4F_no18_tempest5.
Citation: Journal of Atmospheric and Oceanic Technology 42, 6; 10.1175/JTECH-D-24-0061.1
While the lowest frequency (e.g., 10 GHz for GMI, 18 GHz for C+T) is beneficial for extending the precipitation rate (the 10-GHz emission signal saturates at a higher precipitation rate than at 18 GHz), its horizontal field-of-view resolution is also the coarsest. As the sensor scans across a mixed land–water scene such as a coastline or an inland water body, the lowest frequency footprint encompasses a partially mixed land–water area, but the associated finer resolution HF channels may sense a smaller region covering only the overwater or overland portion. As noted in Fig. 10e, this difference in resolution leads to a nonphysical artifact (blue “ribboning” along coastlines) in the discriminant along partially mixed land–water pixels, which has the effect of flagging precipitation where none exists (Fig. 10g).
One way to mitigate this is to remove the lowest frequency channel (in this case, 18 H and 18 V) from the precipitation discrimination (combination 7 in Fig. 3), which employs 9 C+T channels. (The postdetection precipitation estimation algorithm could still use these channels; they are only removed for discrimination purposes.) Figure 4 shows that this combination 7 only slightly degrades the overall TSS performance over ocean, relative to the full 11-channel C+T (combination 0). Effectively, this combination desensitizes the discrimination to the presence of the coarsest resolution channel in the resulting discriminant (Fig. 10f). Comparison of the precipitation screen to the full 11-channel screen is provided in Figs. 10g and 10h at the same FAR < 0.05 performance level as used in Fig. 2. Figure 10h shows that coastal artifacts are largely removed at the expense of missing some of the light rain from the pixels that are located fully overwater. This implies a trade-off between an all-surface discrimination method (not requiring any prior knowledge of where water–land boundaries exist) and detection performance. If this reduced detection of precipitation over water is undesirable, a high-resolution water mask (e.g., Mikelsons et al. 2021) could be used to calculate the water fraction within each 18-GHz FOV. The discrimination would remove the lowest frequency channel when the water fraction is within some range (e.g., 10%–90%).
6. Conclusions
This study examined a microwave surface emissivity–based discrimination methodology to intercompare the precipitation detection performance of the COWVR + TEMPEST sensor duo with similar channels from the GMI and ATMS sensors. In general, a means to complement the large number of CubeSat-sized high-frequency (HF) observations with dual-polarization low-frequency (LF) passive MW channel capability is desirable for producing global precipitation estimates covering the range of light to high intensity. The main conclusions are as follows:
-
Over water surfaces, the addition of the conically scanning COWVR (especially the H polarized) channels to the cross-track TEMPEST exhibited precipitation detection performance similar to GMI at the three (0.5, 1, and 2 mm h−1) threshold rates.
-
Not surprisingly, the 13-channel GMI exhibited superior detection performance under the surface conditions investigated.
-
The inclusion of dual-polarized LF channels, especially H polarized, improves the discrimination of light precipitation over water, mixed coasts, and heavy vegetation. The benefit of an LF + HF combination was noted over lightly vegetated surfaces, with the exception of C+T, where the HF-only channel exhibited a slightly higher skill score than any of the LF + HF channel combinations. One possible explanation is the difference in viewing geometry (constant incidence for COWVR, variable for TEMPEST) and different azimuthal viewing directions. These may mask the impact of the LF channels across the emissivity associated with these surface conditions. However, further investigation is needed before drawing conclusions.
-
Incorrect false detections are introduced along nonprecipitating mixed land–water scenes when all COWVR channels were included for detection. These artifacts were largely mitigated by removing the coarsest resolution (30-km FOV or larger) channel for discrimination purposes.
-
Inclusion of ATMS temperature sounding channels in the 50.3–54.4-GHz range improves discrimination of precipitation over heavily vegetated surfaces, by virtue of scattering depressions at 52.8 GHz over warm land surfaces, and additional surface emissivity diversity at the 50.3-GHz window channel. While not investigated further, previous studies (Camplani et al. 2024) have demonstrated the benefit of the 50-GHz channels for detecting frozen precipitation.
-
A slightly higher optimal threshold skill score (TSS) was noted when the DPR-only Ku-band dataset was used as the source of the reference precipitation compared to the use of the DPR Ku-only combined radar + radiometer (CMB) dataset. This suggests a systematic difference between these two products that may need to be further investigated but emphasizes the importance of examining multiple sources for the precipitation reference.
This study was unable to consider the benefit of the 118-GHz temperature sounding channels that are included with the current (CubeSat sized) tropics microwave sounders (TMS) (You et al. 2023) and which will be common on similar CubeSats and operational MW-sounding-based satellites. This study did not consider the (postdetection) capability of the C+T channel combination to evaluate the precipitation structure and intensity. A separate study (in progress) analyzes the EPC-estimated precipitation profile in relation to the surrounding ocean surface convergence derived from global COWVR ocean wind products.
The acronym FAR is typically used to denote the false alarm ratio (a function of false alarms and hits only). It is not to be confused with the false alarm rate (probability of false detection) that is sensitive to false alarms and correct negatives. However, in this study the acronym FAR is used to denote false alarm rate.
Also known as the Hanssen and Kuipers discriminant, and the Peirce skill score.
Acknowledgments.
The first author acknowledges support from JAXA under the Earth Observation Research Announcement collaborative research Agreement Grant JX-PSPC-549967 and from the Air–Sea Interface and Atmospheric Profile Observatory Science Working Group (ASAP-SWG) under the Weather Focus Area of the NASA Earth Science Division. The authors acknowledge the efforts from Dr. Shannon Brown of JPL, Principal Investigator for the COWVR instrument, and the COWVR and TEMPEST technical staff. Part of this work was carried out at the Jet Propulsion Laboratory, California Institute of Technology, under a contract with the National Aeronautics and Space Administration. © 2024. All rights reserved.
Data availability statement.
The GPM DPR, GMI, and ATMS passive MW radiometer data products used in this study are openly available online from the NASA Precipitation Processing System (PPS), hosted at NASA Goddard Space Flight Center. COWVR and TEMPEST TSDR datasets are available via the Early Adopter program, through the Physical Oceanography Distributed Active Archive Center (PO.DAAC).
APPENDIX
Discrimination Methodology
Figure A1 depicts the flow diagram of the precipitation discrimination procedure described in this appendix. This procedure is carried out one time for each desired passive MW sensor and channel set combination. For purposes of explanation, the COWVR + TEMPEST 11-channel set (6 COWVR fore-view + 5 TEMPEST channels, denoted by C+T) is assumed, but the procedure is identical for other sensors and channel combinations.
Flow diagram of the process to separate precipitating and nonprecipitating TB observations for each passive MW sensor channel combination, in this example at a 0.5 mm h−1 precipitation rate threshold. The flow on the left side is carried out to train the coefficients to estimate the EPC (denoted by u) from the observed TB. Once this is done, the flow on the right side separates u from precipitating and nonprecipitating TB observations producing a discriminant (denoted by t) that can be applied to subsequent observed TB from the same sensor. In this example, the optimal discriminant topt is set to maximize the TSS.
Citation: Journal of Atmospheric and Oceanic Technology 42, 6; 10.1175/JTECH-D-24-0061.1
The discrimination procedure is based upon a transformation applied to the observed brightness temperatures (TBs) and the associated environmental conditions at the time of the observation. Earlier works have shown that under nonprecipitating conditions, the surface emissivity below 90 GHz is highly correlated (Aires et al. 2011). Therefore, a principal component (PC) analysis reduces the overall dimensionality of the data. The resulting orthogonal normalized emissivity PCs (EPCs) are derived from the physical conditions that most influence the measured TB. In addition to the surface emissivity at the frequency and viewing angle of interest, under nonprecipitating conditions, the upwelling TB largely is influenced by the presence of absorbing constituents (water vapor and oxygen molecules) along the path between the surface and the satellite. The absorbing properties of these atmospheric gases are a known function of pressure and temperature. Forecast or reanalysis models such as MERRA2 (Gelaro et al. 2017) provide a reasonable estimate of the surface temperature and water vapor profile. However, the surface emissivity is generally not known (or highly variable) at the time of the satellite overpass.
Figure A2 shows histograms of the first three EPCs and the percent variance that is explained. As expected, over 70% of the variance is explained by the first EPC, which largely represents the surface type (the peaks near 0 and 5 being associated with overwater and overland pixels, respectively).
Histograms of (from left to right) the first three EPCs for the 11-channel C+T sensor. The number in each panel indicates the percent variance that is explained by that EPC.
Citation: Journal of Atmospheric and Oceanic Technology 42, 6; 10.1175/JTECH-D-24-0061.1
Using this estimation, Fig. A3 shows the 2D density histograms for each of the first three EPCs. The correlation is plotted in the upper value, and the accumulated percentage of the explained variance is printed below that. For the leading EPC terms that explain over 93% of the variance, the correlation exceeds 0.98.
2D density histograms for (from left to right) the first three EPCs showing the estimated [Eq. (A2)] value vs the actual value. The top number in each panel indicates the correlation coefficient; the bottom number is the cumulative variance that is explained.
Citation: Journal of Atmospheric and Oceanic Technology 42, 6; 10.1175/JTECH-D-24-0061.1
Figure A4 shows the 2D density histograms of the first three emissivity vector terms (for this C + T channel combination, these are the emissivity at 18.7 H, 18.7 V, and 23.8 V GHz) when these estimated EPCs are transformed back into the emissivity vector. The correlation exceeds 0.99. It is reiterated that this process does not utilize prior surface information or take geographical coordinates into consideration.
2D density histograms of the first three emissivity vector terms [(from left to right) the emissivity at 18.7 H, 18.7V, and 23.8V GHz] when the estimated EPC is transformed back into the emissivity vector. The top number in each panel indicates the correlation coefficient.
Citation: Journal of Atmospheric and Oceanic Technology 42, 6; 10.1175/JTECH-D-24-0061.1
This analysis shows that the factors that influence the nonprecipitating scenes can be well described by the observed TB. As clouds and any associated precipitation begin to influence the TB, these relationships begin to deviate from the range encompassed by the precipitation-free analysis. The degree to which these precipitating EPC terms shift and overlap with the precipitation-free range determines the efficacy of the precipitation discrimination. For example, Fig. A5 shows the histograms of the first three EPCs (of the M = 10 total) where the DPR-estimated near-surface rain rate exceeded 0.5 mm h−1 (blue), overlaid on the top of the nonprecipitating observations from Fig. A2 (green). In this example, the precipitation-related observations for EPC 1 cluster somewhere between the overwater and overland peaks, but for EPC-3, they move significantly away from the nonprecipitating range.
Histograms of the first three EPCs where the DPR-estimated near-surface rain rate exceeded 0.5 mm h−1 (blue), overlaid on top of the EPC from the nonprecipitating observations from Fig. A1 (green).
Citation: Journal of Atmospheric and Oceanic Technology 42, 6; 10.1175/JTECH-D-24-0061.1
(left) Histograms of the discriminant function for the precipitating observations (blue) and the nonprecipitating observations (green), for the 11-channel C+T combination. The black vertical line denotes the value of the optimal discriminant (maximum TSS). The red dashed line to the left (right) represents the value of the discriminant to operate at a FAR < 0.05 (POD > 0.95) rate. (right) Associated ROC. The black operating point denotes the value of the optimal discriminant, corresponding to the black vertical line in the left panel. The red operating points correspond to the discrimination at the FAR < 0.05 and POD > 0.95 levels.
Citation: Journal of Atmospheric and Oceanic Technology 42, 6; 10.1175/JTECH-D-24-0061.1
This analysis is carried out one time for each passive MW sensor channel combination and can be efficiently applied to subsequent TB observations. While this discriminant threshold was selected to maximize the TSS, the user can select other operating characteristic performance levels. For example, to maintain a maximum allowable FAR, the left-most red line in Fig. A6 separates the population at the FAR < 0.05 level (at the expense of lower POD). On the other hand, to maintain a minimum allowable POD, the right-most red line in Fig. A6 separates the population at the POD > 0.95 level (at the expense of higher FAR).
REFERENCES
Aires, F., C. Prigent, F. Bernardo, C. Jiménez, R. Saunders, and P. Brunel, 2011: A Tool to Estimate Land‐Surface Emissivities at Microwave frequencies (TELSEM) for use in numerical weather prediction. Quart. J. Royal Meteoro. Soc., 137, 690–699, https://doi.org/10.1002/qj.803.
Behrangi, A., and Y. Song, 2020: A new estimate for oceanic precipitation amount and distribution using complementary precipitation observations from space and comparison with GPCP. Environ. Res. Lett., 15, 124042, https://doi.org/10.1088/1748-9326/abc6d1.
Berg, W., and Coauthors, 2016: Intercalibration of the GPM microwave radiometer constellation. J. Atmos. Oceanic Technol., 33, 2639–2654, https://doi.org/10.1175/JTECH-D-16-0100.1.
Brogniez, H., and Coauthors, 2022: Time-delayed tandem microwave observations of tropical deep convection: Overview of the C2OMODO mission. Front. Remote Sens., 3, 854735, https://doi.org/10.3389/frsen.2022.854735.
Brown, S., and Coauthors, 2017: The COWVR mission: Demonstrating the capability of a new generation of small satellite weather sensors. 2017 IEEE Aerospace Conf., Big Sky, MT, Institute of Electrical and Electronics Engineers, 1–7, https://doi.org/10.1109/AERO.2017.7943884.
Camplani, A., D. Casella, P. Sanò, and G. Panegrossi, 2024: The High lAtitude sNowfall Detection and Estimation aLgorithm for ATMS (HANDEL-ATMS): A new algorithm for snowfall retrieval at high latitudes. Atmos. Meas. Tech., 17, 2195–2217, https://doi.org/10.5194/amt-17-2195-2024.
Favrichon, S., C. Prigent, C. Jimenez, and R. Vogt, 2023: Global mapping of microwave emissivity and emitting depth in arid areas using GMI observations. Earth Space Sci., 10, e2022EA002756, https://doi.org/10.1029/2022EA002756.
Ferraro, R. R., E. A. Smith, W. Berg, and G. J. Huffman, 1998: A screening methodology for passive microwave precipitation retrieval algorithms. J. Atmos. Sci., 55, 1583–1600, https://doi.org/10.1175/1520-0469(1998)055<1583:ASMFPM>2.0.CO;2.
Gelaro, R., and Coauthors, 2017: The Modern-Era Retrospective Analysis for Research and Applications, version 2 (MERRA-2). J. Climate, 30, 5419–5454, https://doi.org/10.1175/JCLI-D-16-0758.1.
GPROF, 2022: Algorithm Theoretical Basis Document GPROF2021, version 1. Algorithm Theoretical Basis Doc., 62 pp., https://gpm.nasa.gov/sites/default/files/2022-06/ATBD_GPM_V7_GPROF.pdf.
Grecu, M., W. S. Olson, S. J. Munchak, S. Ringerud, L. Liao, Z. Haddad, B. L. Kelley, and S. F. McLaughlin, 2016: The GPM combined algorithm. J. Atmos. Oceanic Technol., 33, 2225–2245, https://doi.org/10.1175/JTECH-D-16-0019.1.
Hamada, A., and Y. N. Takayabu, 2016: Improvements in detection of light precipitation with the Global Precipitation Measurement Dual-frequency Precipitation Radar (GPM DPR). J. Atmos. Oceanic Technol., 33, 653–667, https://doi.org/10.1175/JTECH-D-15-0097.1.
Huffman, G. J., and Coauthors, 2020: Integrated Multi-satellitE Retrievals for the Global Precipitation Measurement (GPM) mission (IMERG). Satellite Precipitation Measurement: Volume 1, V. Levizzani et al., Eds., Advances in Global Change Research, Vol. 67, Springer, 343–353, https://doi.org/10.1007/978-3-030-24568-9_19.
Kidd, C., G. Huffman, V. Maggioni, P. Chambon, and R. Oki, 2021a: The global satellite precipitation constellation: Current status and future requirements. Bull. Amer. Meteor. Soc., 102, E1844–E1861, https://doi.org/10.1175/BAMS-D-20-0299.1.
Kidd, C., E. Graham, T. Smyth, and M. Gill, 2021b: Assessing the impact of light/shallow precipitation retrievals from satellite-based observations using surface radar and micro rain radar observations. Remote Sens., 13, 1708, https://doi.org/10.3390/rs13091708.
Kirstetter, P.-E., and Coauthors, 2012: Toward a framework for systematic error modeling of spaceborne precipitation radar with NOAA/NSSL ground radar–based national mosaic QPE. J. Hydrometeor., 13, 1285–1300, https://doi.org/10.1175/JHM-D-11-0139.1.
Kubota, T., C. D. Kummerow, G. Panegrossi, S. Ringerud, N. Utsumi, and F. J. Turk, 2025: The 11th Workshop of the International Precipitation Working Group (IPWG). Bull. Amer. Meteor. Soc., https://doi.org/10.1175/BAMS-D-25-0011.1.
Li, Z., D. B. Wright, S. H. Hartke, D. B. Kirschbaum, S. Khan, V. Maggioni, and P.-E. Kirstetter, 2023: Toward a globally-applicable uncertainty quantification framework for satellite multisensor precipitation products based on GPM DPR. IEEE Trans. Geosci. Remote Sens., 61, 1–15, https://doi.org/10.1109/TGRS.2023.3235270.
Liao, L., R. Meneghini, T. Iguchi, and A. Tokay, 2020: Characteristics of DSD bulk parameters: Implication for radar rain retrieval. Atmosphere, 11, 670, https://doi.org/10.3390/atmos11060670.
Maiwald, F., and Coauthors, 2020: Completion of the AMR-C instrument for Sentinel-6. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens., 13, 1811–1818, https://doi.org/10.1109/JSTARS.2020.2991175.
Manzato, A., 2007: A note on the maximum Peirce skill score. Wea. Forecasting, 22, 1148–1154, https://doi.org/10.1175/WAF1041.1.
Mikelsons, K., M. Wang, X.-L. Wang, and L. Jiang, 2021: Global land mask for satellite ocean color remote sensing. Remote Sens. Environ., 257, 112356, https://doi.org/10.1016/j.rse.2021.112356.
Prigent, C., F. Aires, and W. B. Rossow, 2006: Land surface microwave emissivities over the globe for a decade. Bull Amer. Meteor. Soc., 87, 1573–1584, https://doi.org/10.1175/BAMS-87-11-1573.
Ringerud, S., C. Kummerow, C. Peters-Lidard, Y. Tian, and K. Harrison, 2014: A comparison of microwave window channel retrieved and forward-modeled emissivities over the U.S. Southern Great Plains. IEEE Trans. Geosci. Remote Sens., 52, 2395–2412, https://doi.org/10.1109/TGRS.2013.2260759.
Ringerud, S., C. Peters-Lidard, J. Munchak, and Y. You, 2021: Applications of dynamic land surface information for passive microwave precipitation retrieval. J. Atmos. Oceanic Technol., 38, 167–180, https://doi.org/10.1175/jtech-d-20-0048.1.
Schulte, R. M., C. D. Kummerow, W. Berg, S. C. Reising, S. T. Brown, T. C. Gaier, B. H. Lim, and S. Padmanabhan, 2020: A passive microwave retrieval algorithm with minimal view-angle bias: Application to the TEMPEST-D CubeSat mission. J. Atmos. Oceanic Technol., 37, 197–210, https://doi.org/10.1175/JTECH-D-19-0163.1.
Seto, S., T. Iguchi, R. Meneghini, J. Awaka, T. Kubota, T. Masaki, and N. Takahashi, 2021: The precipitation rate retrieval algorithms for the GPM Dual-frequency Precipitation Radar. J. Meteor. Soc. Japan, 99, 205–237, https://doi.org/10.2151/jmsj.2021-011.
Turk, F. J., L. Li, and Z. S. Haddad, 2014: A physically based soil moisture and microwave emissivity data set for Global Precipitation Measurement (GPM) applications. IEEE Trans. Geosci. Remote Sens., 52, 7637–7650, https://doi.org/10.1109/TGRS.2014.2315809.
Turk, F. J., Z. S. Haddad, P.-E. Kirstetter, Y. You, and S. Ringerud, 2018: An observationally based method for stratifying a priori passive microwave observations in a Bayesian‐based precipitation retrieval framework. Quart. J. Roy. Meteor. Soc., 144, 145–164, https://doi.org/10.1002/qj.3203.
Turk, F. J., and Coauthors, 2021: Adapting passive microwave-based precipitation algorithms to variable microwave land surface emissivity to improve precipitation estimation from the GPM constellation. J. Hydrometeor., 1755–1781, https://doi.org/10.1175/JHM-D-20-0296.1.
Ushio, T., and Coauthors, 2009: A Kalman filter approach to the Global Satellite Mapping of Precipitation (GSMaP) from combined passive microwave and infrared radiometric data. J. Meteor. Soc. Japan, 87A, 137–151, https://doi.org/10.2151/jmsj.87A.137.
Utsumi, N., F. J. Turk, Z. S. Haddad, P.-E. Kirstetter, and H. Kim, 2021: Evaluation of precipitation vertical profiles estimated by GPM-Era satellite-based passive microwave retrievals. J. Hydrometeor., 22, 95–112, https://doi.org/10.1175/jhm-d-20-0160.1.
You, Y., N.-Y. Wang, and R. Ferraro, 2015: A prototype precipitation retrieval algorithm over land using passive microwave observations stratified by surface condition and precipitation vertical structure. J. Geophys. Res. Atmos., 120, 5295–5315, https://doi.org/10.1002/2014JD022534.
You, Y., G. Huffman, C. Kidd, S. Braun, W. Blackwell, J. X. Yang, and C. Da, 2023: Evaluating and improving TROPICS millimeter‐wave sounder’s precipitation estimate over ocean. J. Geophys. Res. Atmos., 128, e2023JD038697, https://doi.org/10.1029/2023JD038697.
