The objective of this paper is to describe the development and evaluate the performance of a completely new version of the Passive microwave Neural network Precipitation Retrieval (PNPR v2), an algorithm based on a neural network approach, designed to retrieve the instantaneous surface precipitation rate using the cross-track Advanced Technology Microwave Sounder (ATMS) radiometer measurements. This algorithm, developed within the EUMETSAT H-SAF program, represents an evolution of the previous version (PNPR v1), developed for AMSU/MHS radiometers (and used and distributed operationally within H-SAF), with improvements aimed at exploiting the new precipitation-sensing capabilities of ATMS with respect to AMSU/MHS. In the design of the neural network the new ATMS channels compared to AMSU/MHS, and their combinations, including the brightness temperature differences in the water vapor absorption band, around 183 GHz, are considered. The algorithm is based on a single neural network, for all types of surface background, trained using a large database based on 94 cloud-resolving model simulations over the European and the African areas.
The performance of PNPR v2 has been evaluated through an intercomparison of
the instantaneous precipitation estimates with co-located estimates from the
TRMM Precipitation Radar (TRMM-PR) and from the GPM Core Observatory Ku-band
Precipitation Radar (GPM-KuPR). In the comparison with TRMM-PR, over the
African area the statistical analysis was carried out for a 2-year
(2013–2014) dataset of coincident observations over a regular grid at
0.5
The availability of data from the Advanced Technology Microwave Sounder (ATMS), a cross-track scanning radiometer currently onboard the Suomi National Polar-orbiting Partnership (Suomi-NPP) satellite (and on the Joint Polar Satellite System (JPSS) series starting in 2017), represents an important step in short- and long-term weather forecasting and environmental monitoring. Combining the capabilities of its predecessor sounders such as the Advanced Microwave Sounding Unit-A (AMSU-A) and the Microwave Humidity Sounder (MHS) aboard NOAA-18 and NOAA-19 and the ESA MetOp-A and MetOp-B satellites, ATMS provides sounding observations with improved resolution, sampling, and coverage for retrieving atmospheric vertical temperature and humidity profiles. Moreover, this new-generation instrument provides more information about surface, vertical distribution of hydrometeors, precipitation, and other key environmental variables (Chen et al., 2007; Boukabara et al., 2013; Zou et al., 2013; Kongoli et al., 2015).
With regard to precipitation it should be mentioned that, although the reliable knowledge of its intensity and accumulation is essential for understanding the global hydrological and energy cycles, precipitation estimate (from satellite and from the surface) is complicated by several factors: the large variability of the precipitation in time and space, the conversion of satellite measurements into quantitative precipitation estimates, uncertainties associated to rain gauges (and to their spatial distribution) and radar measurements (i.e., attenuation, beam-blocking), and their unavailability in several regions in the world and over ocean (Mugnai et al., 1993; Iturbide-Sanchez et al., 2011; Bennartz and Petty, 2001; Tian et al., 2009; Kirstetter et al., 2012).
An important step forward towards the improvement of global precipitation
monitoring is represented by the Global Precipitation Measurement (GPM)
mission launched on 27 February 2014. GPM is expected to provide accurate
precipitation estimates thanks to the availability of the NASA/JAXA GPM Core
Observatory (GPM-CO) (equipped with the GPM Microwave Imager (GMI) and the
Dual-frequency Precipitation Radar (DPR)), a common, global observatory of
3-D precipitation structure at 5 km resolution, and to the
exploitation of a constellation of international Low Earth Orbit (LEO)
satellites equipped with microwave radiometers for precipitation
observation, providing frequent measurements over most of the globe
(3-hourly coverage between 65
In Europe, the EUMETSAT “Satellite Application Facility on Support to Operational Hydrology and Water Management” (H-SAF; Mugnai et al., 2013a) has been called upon to participate in and to contribute towards the GPM by providing its own precipitation products and simultaneously be a user of GPM data and a direct collaborator of GPM on two main aspects: development and refinement of retrieval techniques through the exploitation of all available radiometers in the GPM constellation, and validation activity. In this context, operational passive microwave (PMW) precipitation products for the different radiometers are being released within H-SAF as new radiometers become available, and they are based on two approaches (Mugnai et al., 2013b): the physically based Bayesian Cloud Dynamics and Radiation Database (CDRD) algorithm (Casella et al., 2013, Sanò et al., 2013) for conically scanning radiometers and the Passive microwave Neural network Precipitation Retrieval algorithm (PNPR) for cross-track scanning radiometers, originally developed for AMSU/MHS and fully described in Sanò et al. (2015) (PNPR-AMSU/MHS, hereafter PNPR v1).
The objective of this paper is to describe the development and evaluate the performance of a newly developed version of PNPR designed to retrieve the instantaneous surface precipitation using the ATMS radiometer data. This algorithm (PNPR-ATMS, hereafter PNPR v2) represents an evolution of PNPR v1 (used operationally within the EUMETSAT H-SAF) with improvements aimed at exploiting the new precipitation-sensing capabilities of ATMS with respect to AMSU/MHS.
Neural networks (NNs) represent a highly flexible tool alternative to regression and classification techniques, widely applied in an increasing field of meteorological research for their capability to approximate complex nonlinear and imperfectly known functions (e.g., Liou et al., 1999; Del Frate and Schiavon, 1999; Shi, 2001; Marzban, 2003; Blackwell and Chen, 2005; Chen et al., 2006; Krasnopolsky et al., 2008; Shank et al., 2008; Haupt et al., 2009; Aires et al., 2012).
NNs have been used in precipitation retrieval – precipitation being one of the most difficult of all atmospheric variables to retrieve – because of the opportunities offered by their ability to learn and generalize (Hsu et al., 1997; Hall et al., 1999; Staelin et al., 1999; Sorooshian et al., 2000; Chen and Staelin, 2003; Hong et al., 2004; Surussavadee and Staelin, 2007, 2008a, b, 2009, 2010; Bellerby, 2007; Krasnopolsky et al., 2008; Leslie et al., 2008; Mahesh et al., 2011). However, it should be mentioned that the use of NNs involves the training phase with a large representative database, often obtained from cloud-resolving model simulations. Consequently, the performance of the network is largely dependent on the completeness and the representativeness of the database and on its consistency with the observations.
Retrieval algorithms based on NNs, proposed for precipitation estimation
from remotely sensed information, using MW or VIS/IR measurements, are
different from each other in the different approaches used, in the design of
the network architecture, in the selection of type and number of input
variables, in the determination of the number of networks used, in the
implementation of the training database (e.g., the cloud-resolving model),
and in the training process. With regard to the input variables, when MW
radiometers are used their choice is normally based on physical
considerations on the radiometric signatures (or brightness temperatures,
TBs) of different microwave channels and on the direct or indirect
relationship of these signatures with environmental, meteorological, and
microphysical variables (e.g., atmospheric temperature and humidity, surface
conditions, hydrometeor types, sizes, and shapes) involved in the
precipitation retrieval process. The TBs at the MW channels so identified
are selected as part of the input variables. However, some techniques such
as principal component analysis (PCA) are applied to the selected channels
in order to reduce the number of inputs, reduce the complexity of the
NN, and to reduce the noise (e.g., to filter out the signal due to the
background surface) (Chen and Staelin, 2003; Surussavadee and Staelin,
2008a, 2010; Blackwell and Chen, 2005). Special functions of TBs already
proposed for rainfall retrieval (Kidd, 1998; Ferraro and Marks, 1995; Grody,
1991), such as the polarization corrected temperature (PCT
In the VIS/IR-based NN algorithms the input selection is based on different
considerations due to the indirect relationship between cloud-top radiances
and surface rainfall and the lack of information on the precipitation
structure within the cloud. Additional inputs are then considered in
addition to TBs, i.e., cloud texture information (TB mean and variance
for 3
The number of NNs used in the precipitation retrieval algorithms is defined so as to optimize the network performance under different operating conditions. In PMW precipitation retrieval separate NN algorithms are usually proposed depending on the type of surface (i.e., land or sea) to discriminate between the different precipitation emission signatures relative to background (e.g., Surussavadee and Staelin, 2008a). Separate NN algorithms are also proposed to deal separately with stratiform and convective precipitation (e.g Sarma et al., 2008).
In the design of PNPR v2 important aspects in relation to the topics mentioned above, concerning the choice of the inputs, the number of networks used by the algorithm, and the database used in the training phase, have been thoroughly analyzed and will be presented in this paper.
Another important issue to consider is that PNPR v2 has been designed in the perspective of the full exploitation of the MW radiometers in the GPM constellation of satellites, and of the achievement of consistency (besides accuracy) of the retrievals from the different sensors. These goals are considered priorities in the international GPM mission community because their achievement leads to a significant reduction of the errors, also associated with the inadequate sampling of precipitation, with positive impact on precipitation monitoring (see also Panegrossi et al., 2015, 2016), hydrological applications, and climate studies. This is also true when higher spatial/temporal-resolution products based on MW/IR combined techniques are used, such as IMERG (GPM) and TMPA (Tropical Rainfall Measuring Mission – TRMM; see Huffman et al., 2007, 2015); within the EUMETSAT H-SAF program these aspects have also become a priority. Therefore, PNPR v2 for ATMS, as well as PNPR v1 for AMSU/MHS, and all other H-SAF products for conically scanning radiometers represent an important contribution towards the exploitation of the current and future constellation of PMW radiometers for global precipitation monitoring.
In this paper the PNPR v2 algorithm is described in detail, and the methodology and the results of an intercomparison of the PNPR v2 instantaneous precipitation estimates with co-located spaceborne radar estimates from the TRMM Precipitation Radar (TRMM-PR) and from the GPM-CO Ku-band Precipitation Radar (GPM-KuPR) are presented.
Section 2 presents a brief description of the characteristics of ATMS. In Sect. 3 a description of the PNPR v2 algorithm is presented, with reference to the design of the neural network, the main characteristics of the algorithm, and the relevant features of the ATMS training database. The verification study is presented in Sect. 4, which includes a brief description of the characteristics of PR and DPR, of the methodology used to create the co-located observation dataset used in the study, the analysis of the performance of PNPR v2 compared to TRMM-PR and to GPM-KuPR, and a comparison with PNPR v1 using TRMM-PR rainfall estimates as reference. Section 5 contains the conclusive remarks about the performance of PNPR v2 and future perspectives.
ATMS is a total power cross-track scanning microwave radiometer on board the
NPP satellite (and JPSS
satellites scheduled for early 2017), with a swath of 2600 km, angular span
of
Compared with its predecessors AMSU and MHS, ATMS has improved resolution
(31.6 km at nadir in the 54 GHz band, vs. 48.6 km for AMSU) and angular
sampling (1.11
PNPR v2 represents an evolution for ATMS applications, of the previous PNPR v1 algorithm based on a NN approach, developed at ISAC-CNR for precipitation rate estimation using AMSU/MHS observations. The full description of PNPR v1 is provided in Sanò et al. (2015), while some important aspects are reviewed in this paper for completeness.
Both versions of PNPR are designed to work over the full Meteosat Second
Generation (MSG) disk area (60
Another significant aspect in the design of PNPR v1 was the use of the TB
differences in the water vapor absorption band channels at 183 GHz as input
to the neural network. Opaque channels around 183 GHz were originally
designed to retrieve water vapor profiles due to their different sensitivity
to specific layers of the atmosphere (Wang et al., 1997; Staelin and Chen,
2000; Blackwell and Chen, 2005). However, these channels have shown great
potential for precipitating cloud characterization and for precipitation
retrieval. The different penetration ability of these channels in the
atmosphere can be exploited to analyze the vertical distribution of
hydrometeors (Wang et al., 1989, 1997; Burns et al., 1997; Staelin and Chen,
2000; Ferraro et al., 2005; Hong et al., 2005, 2008; Funatsu et al., 2007,
2009; Laviola and Levizzani, 2011) and to obtain some criteria for the
characterization of precipitation as weak, moderate, strong convective, or
stratiform using the TB differences
The flow diagram of the PNPR v2 algorithm is basically the same as that
of PNPR v1, described in detail in Sanò et al. (2015), except for the
use of one unique network trained on a database representative of MSG full
disk area (see Sect. 3.2) and changes in the input selection in the
design of the network (described in Sect. 3.4). Furthermore, in the
preprocessing of the brightness temperatures, in addition to the decoding of
the file format and the quality control of the input data, the removal of
the three outmost pixels along the scan is carried out. Other processing
steps of the algorithm, such as the screening procedure of no-rain pixels and the
quality index map providing indications on areas or conditions where the retrieval is more or less reliable,
are unchanged with respect to those used for the algorithm PNPR v1 (Sanò et al., 2015). In a similar way,
the new algorithm also provides at its output, in addition to the precipitation rate
(mm h
The training of PNPR v2 was performed using a large cloud–radiation database representative of the MSG full disk area, built from 94 cloud-resolving model (CRM) simulations of different precipitation events including 60 simulations over the European/Mediterranean area (Casella et al., 2013) and 34 simulations over Africa and Southern Atlantic (Panegrossi et al, 2014). The simulations were carried out using the University of Wisconsin Nonhydrostatic Modeling System (UW-NMS) (Tripoli, 1992; Tripoli and Smith, 2014a, b) coupled to a radiative transfer model (RTM) relating CRM environments to expected top-of-atmosphere PMW TBs of the ATMS radiometer (see Smith et al., 2013, and Casella et al., 2013, for the details about the cloud model configuration setup and Sanò et al., 2015, for AMSU/MHS RTM simulations). Figure 1 shows the geographical location of the inner domain of the 94 simulations. Simulated events were selected in order to cover the different seasons and different meteorological situations and precipitation regimes. The selection of the simulations in terms of season, typology of event, and geographical location was performed in order to optimize the completeness and representativeness of the database for the area of interest (see Casella et al., 2013). In detail, over the European/Mediterranean area we have considered 15 different meteorological events for each season over different geographical areas. Simulations over African and Southern Atlantic area were chosen also on the basis of the TRMM-PR observations (in particular the rain type flag and the freezing level height) and different climatic regions in order to cover as much as possible the climatic variability in the area of interest with a limited number of simulations.
Geographical location of the inner domain of the 94 NMS simulations over European and African areas.
The simulated TBs were calculated considering the different ATMS viewing angles and channel frequencies using the same approach used for AMSU/MHS and described in Sanò et al. (2015). For the European/African regions, the database contains more than 70 million entries. Each entry is a vector composed of the simulated ATMS TBs, surface precipitation rate, and the corresponding ancillary parameters, associated with one cloud-resolving model microphysical realization and one ATMS viewing angle (and corresponding IFOV). It is worth noting that 45 different ATMS viewing angles (discarding the three outmost pixels due to the low resolution) are considered to build the database.
A detailed description of the NN is provided in Sanò et al. (2015), but some basic aspects are presented for completeness.
The neural network scheme, shown in Fig. 2 in Sanò et al. (2015), is
characterized by
Relative sensitivity (
The first objective in the new NN design was the selection of the inputs
based on the evaluation of their impact on the performance of the NN or on
their sensitivity to precipitation. Consistently with PNPR v1 and on the
basis of the results obtained for AMSU/MHS (Sanò et al., 2015), for the
new NN we have initially imposed the use of the three inputs
Results of the tests for the selection of the inputs to the NN.
Input combinations are listed in the first column (
In the table the various possible differences considered as input to the NN
in this analysis are shown in the first column;
The contribution of
Another difference between PNPR v2 and PNPR v1 algorithms is the result of the canonical correlation analysis (CCA) applied to the training database to find the linear combination of TBs (LCT) of selected channels best correlated with surface precipitation rate, to be used as additional input to the network (see Sanò et al., 2015). The resulting linear combination for ATMS is composed of the window channels 31.4, 88.2, and 165.5 GHz, showing the highest CCs in the CCA analysis (with respect to the surface rain rate) for all types of background surfaces (in PNPR v1 for AMSU/MHS the 50.3, 89, and 150 GHz were selected for LCT).
With regard to other inputs to the network, in PNPR v2 the same ancillary data used in PNPR v1 were maintained (surface height, background surface type, month, and secant of the zenith angle along the ATMS cross-track scan). An additional auxiliary input was added to drive NN in the transition between the European and African area, i.e., the monthly mean total precipitable water (TPW) obtained from ECMWF Era Interim reanalysis in the 2011–2014 period. It should be mentioned that geographical and environmental/meteorological parameters (including TPW) in PMW precipitation retrieval are utilized to reduce the ambiguity intrinsic to the PMW precipitation retrieval process (for example in the NASA GPM Bayesian algorithms – see Kummerow et al., 2011, 2015; Kidd et al., 2016).
During the phase of network design and the training process, more than 400 architectures have been tested and an “optimal” NN has been obtained.
In summary, 10 input variables (five TBs derived and five ancillary inputs)
are used in the NN for ATMS:
an LCT at 31.4, 88.2, and 165.5 GHz; surface type (land, sea, coast); monthly mean TPW; month; surface height (altitude); secant of the zenith angle.
The network architecture is similar to that of PNPR v1, with one input layer (with number of nodes equal to the number of inputs) and two hidden layers with 23 and 10 nodes in the first and in the second layer, respectively (the number of nodes differs from PNPR v1). The tan-sigmoid transfer function is used for the input and the hidden layers, while a linear transfer function is used for the output node.
During the training procedure, an assessment of the sensitivity of the NN output to variations of the inputs was carried out. Sensitivity analysis provides an estimation of the relative importance of the inputs (Coulibaly et al., 2005). The knowledge of the NN behavior, in relation to input perturbation, helps to assess the relevance of the individual contributions to the output and to verify the correct training of the NN (i.e., the weights remain stable) that is achieved when there is no significant changes of the sensitivity during the last training iterations (epochs).
The sensitivity analysis, limited to the TBs derived variables that are more
related to the rain rate estimate and not to the ancillary variables, was
applied to the “optimal” NN (i.e., defined by the listed inputs and the
architecture described in the previous section) and was carried out during
the final phase of the training (see Sanò et al., 2015). The final phase
was reached when the two parameters indicating the quality of the learning
process, i.e., the CC (
The relative sensitivity (
Number of co-located pixels from TRMM-PR and the Suomi-NPP ATMS coincident overpasses over the African area in the 24-month period 2013–2014 (left panel) and from GPM-Ku-NS and Suomi-NPP ATMS coincident overpasses over European and African areas in the 15-month period (March 2014–May 2015) (right panel).
The results show a similar behavior of the sensitivity for the three
different surface backgrounds considered. It is evident the higher
sensitivity of NN with respect to the LCT in comparison with the other
inputs; this is due to the contribution of window channels used in LCT,
selected by maximizing the correlation with the surface precipitation rate.
Another important aspect is the relative contribution of the other inputs
(TBs difference in the 183 GHz band channels) quite similar among the three
types of surface, with a slightly higher contribution of the input
This section presents the verification study carried out for the PNPR v2
algorithm, using as reference the data provided by the TRMM and GPM
spaceborne radars. The TRMM-PR is a 13.8 GHz radar with a swath width of 247 km
(after the satellite was boosted to higher orbit in 2001). Its coverage
allows regional intercomparison of convective–stratiform contributions to
precipitation across the tropics, with data available since the launch of the
satellite in November 1997 until October 2014. It is considered the precursor
to GPM DPR and has represented, during this time interval, the best
available remote-sensing instrument for precipitation (Schumacher and Houze Jr.,
2003). The TRMM PR2A25 product (Iguchi et al., 2000) provides rainfall rates
based on the reflectivity–rainfall rate relationships, along with a raindrop
size distribution (DSD) model, attenuation correction, and a non-uniform beam-filling correction. Even though issues have been raised about the accuracy
of PR2A25, related to surface properties, variations of the DSD, or impact of
incidence angles (i.e., Iguchi et al., 2009; Hirose et al., 2012; Kirstetter
et al., 2013), during its operational period this radar has provided accurate
estimates of instantaneous rain rate, as well as calibration for other
precipitation-relevant sensors in sun-synchronous orbits (Bellerby et al.,
2000; Heymsfield et al., 2000; Liao et al., 2001; Schumacher and Houze Jr., 2003;
Lin and Hou, 2008). The GPM DPR (on board the GPM-CO) is composed of two
precipitation radars, the GPM-KuPR at 13.6 GHz (an updated version of the
TRMM-PR) and the Ka-band precipitation radar (GPM-KaPR) at 35.5 GHz. The
simultaneous use of the two radars was designed to obtain a greater dynamic
range in the measurements, more detailed information on the microphysical
rain structure (such as raindrop size distribution), and a consequent better
accuracy in the rainfall retrieval (Le and Chandrasekar, 2013a, b; Hou et
al., 2014; Chandrasekar et al., 2014). KuPR and KaPR have the same space
resolution at nadir, equal to 5.2 km, the same beamwidth, equal to
0.71
It is worth considering also that in spite of the similarities between the two radars, the GPM-KuPR has higher sensitivity (with minimum detectable reflectivity between 12 and 14 dBZ, outperforming the original instrumental design of 18 dBZ) (Toyoshima et al., 2015; Hamada and Takayabu, 2016) than the TRMM-PR radar (18 dBZ minimum detectable reflectivity).
Two datasets have been created, one composed of 2 years (2013–2014) of
coincident Suomi-NPP ATMS and TRMM-PR overpasses over the African area
(36
It should be pointed out that the results obtained from the ATMS-DPR-Ku coincidence dataset are not as robust as the results obtained from the ATMS-PR dataset because of the limited size of the dataset and because of some uncertainties in the less consolidated day-1 V03 DPR products, linked to factors such as the DSD parameterization (Liao et al., 2014), the evaluation of the path-integrated attenuation, the surface reference technique, and the non-uniform beam-filling effect (Shimozuma and Seto, 2015).
Figure 3 (left panel) shows the geographical distribution (on the ATMS grid) of about 1.8 milions coincident pixels ATMS-PR found over the African area in the 2-year time frame 2013–2014. The figure shows a rather good coverage of the entire area, with a number of coincident pixels between 30 and 150 on Central Africa, increasing moving to the north and to the south.
In the right panel of the figure, the distribution of the coincident pixels ATMS-DPR-Ku over the European and African areas, between March 2014 and May 2015, is shown. In contrast to the left panel, the coverage is not as good with a lower number of coincident pixels, and with some uncovered areas. The number on coincident pixels increases over northern Europe at the high latitudes, reaching a maximum value around 200. In the southern part of Europe and Africa, the number of coincidences is significantly reduced (maximum values around 50).
To obtain co-located vectors of rainfall estimates of ATMS and TRMM-PR, and
of ATMS and GPM-KuPR, the radar precipitation rate at the surface was
downscaled to the PNPR v2 product nominal resolution (variable along the
scan line, see Sect. 3.1), by averaging the rainfall rate of all radar
pixels falling within each PNPR v2 pixel. In order to reduce the geolocation
and synchronization errors, due to the different viewing geometry of ATMS
and the spaceborne radar, and to the time lag between the observations,
statistical analysis was carried out over a regular grid at 0.5
Figure 4 shows the geographical distribution of the values of three
statistical indexes (hit bias, CC, and root mean
squared error (RMSE); see Tian et al., 2016, for the definition of these
scores), obtained for the ATMS-PR dataset. The scores are computed
considering all coincident ATMS-PR pixels within each 0.5
Hit bias (top panel), correlation coefficient (CC, middle
panel),
and root mean squared error (RMSE, bottom panel) resulting from the
comparison between PNPR v2 and TRMM-PR retrievals over the African area
(using a 0.5
The top panel shows a rather uniform distribution of low bias (between
In Fig. 5 the density scatter plots for all 0.5
Density scatter plots of the PNPR v2 and TRMM-PR mean rainfall
rates (over a 0.5
Table 2 presents the contingency table for the ATMS-PR dataset, based on the
mean rainfall rate from ATMS and TRMM-PR within each
0.5
Contingency table of PNPR v2 retrievals relative to TRMM-PR
measurements (at 0.5
Performance indexes for the different background surfaces.
Table 3 shows the performance index calculated for the different background
surfaces, defined as
As mentioned previously, a verification of PNPR v2 algorithm has been made also using precipitation rate estimates from the GPM-KuPR, available at mid-high latitudes. This was initially intended for the European area only, where a larger number of coincident overpasses are available during the time frame considered (March 2014–May 2015) (see Fig. 3). However, results are shown also for the African area, despite the lower number of coincidences available, in order to assess the degree of consistency of the results obtained over the same area with the two Ku-band spaceborne radars.
As for the comparison with the TRMM-PR, all co-located ATMS and GPM-KuPR
retrievals were regridded at a 0.5
Statistical indexes obtained in the comparisons of PNPR v2 retrievals with GPM-KuPR and TRMM-PR products.
Density scatter plots of the PNPR v2 and GPM-KuPR mean rainfall
rates (over a 0.5
The table shows a good agreement between the scores obtained with two datasets, with very low bias (slightly positive/negative over land/ocean for the ATMS-DPR-Ku dataset, while the reverse is valid for the ATMS-PR dataset), low RMSE (lower for the ATMS-DPR-Ku dataset), and good correlation.
The right panels of Fig. 6 show the scatter plots obtained in the comparison
of PNPR v2 and GPM-KuPR over the European area. Pixels with likely presence
of ice or snow on the ground have been eliminated from the dataset in order
to exclude from the verification study cases of snowfall (or precipitation
over frozen background) whose precipitation rate estimate is affected by
larger uncertainty (both in the GPM-DPR-Ku V03 product and in PNPR v2). For
the identification of these pixels the “Snow Depth” and “Sea Ice Cover”
products from the ECMWF Era Interim re-analysis (at 0.5
The scatter plots in Fig. 6 show a similar behavior for vegetated land for
the two areas, while over ocean in the European area there is a general
tendency of PNPR v2 to overestimate the precipitation with respect to the
GPM-KuPR. The total bias has very low values, negative for vegetated land
(
In order to better interpret the results in Fig. 6, the geographical
distribution of bias, CC, and RMSE over the European area is shown in Fig. 7
(similarly to Fig. 4). In these maps the statistical indexes are evaluated
including pixels with snow or ice on the ground. There is a prevalence of a
positive bias (although mostly below 0.3 mm h
Hit bias (top panel), CC (middle panel), and RMSE (bottom panel)
resulting from the comparison between PNPR v2 retrievals and GPM-KuPR
measurements over the European area (using a 0.5
In the second part of the verification study we have compared the performances over the African area of the PNPR v2 with the PNPR v1 to evaluate whether the use of the new ATMS channels and the newly designed NN have led to improvements in the retrievals. The performance of the PNPR v1 algorithm has been tested on the same 2-year period (2013–2014) used for PNPR v2, considering coincident observations of AMSU/MHS radiometers, on board the NOAA-18, NOAA-19, MetOp-A, and MetOp-B satellites, with TRMM-PR. The PNPR v1 and TRMM-PR coincidence dataset is made of about 3 million pixels. The procedure used to evaluate the PNPR v1 performance is the same as that adopted for the PNPR v2 algorithm, described in Sect. 4.1.
Table 5 presents the values of the statistical indexes hit bias, CC, and RMSE
obtained in the comparison of PNPR v1 and PNPR v2 with TRMM-PR precipitation
retrievals, over a 0.5
Statistical indexes of the comparison of PNPR v1 and PNPR v2 vs. TRMM-PR retrievals.
A further analysis of the performance of the two algorithms has been
performed through the study of the relative bias percentage (RB
These variables are defined as
Considering the RB
Relative bias percentage (top panel) and AFSE percentage (bottom panel) of PNPR v1 and PNPR v2 retrievals with respect to the TRMM-PR measurements.
Over land both the algorithms present similar performances, with a slightly
better result for PNPR v2 (solid black line) for low rain rates
(0–3 mm h
In the bottom panel of Fig. 8 AFSE
Over land, PNPR v2 shows lower AFSE
It is worth noting that the main improvement of PNPR v2 with respect to PNPR
v1 is the reduction of the relative bias (RB
This paper describes the design of a new algorithm, PNPR v2, for estimation of precipitation on the ground for the cross-track ATMS radiometer and presents the results of a verification study where the instantaneous precipitation rate estimates available from TRMM and GPM spaceborne radars are used as reference.
PNPR v2 has been designed for retrieval of precipitation in the MSG full disk area. The algorithm, based on a neural network approach, represents an evolution of the previous version PNPR v1, designed for the AMSU/MHS radiometer, with some changes made to take advantage of the improvements of ATMS with respect to AMSU/MHS. Similarly to the previous algorithm it is based on a single neural network for all types of surface background, trained using a large database based on 94 cloud-resolving model simulations over the European and the African areas.
The verification study carried out through a comparison with co-located
observations of ATMS with the NASA/JAXA TRMM-PR and GPM-KuPR spaceborne
radars analyzed on a 0.5
For reference, it is useful to compare these results with those found by
other authors carrying out validation studies using ground-based radar data.
Tang et al. (2014) investigated the performance of PMW precipitation
products from 12 passive microwave radiometers, including AMSU-B (NOAA 15,
16, 17) and MHS (NOAA 18, 19 and MetOp-A) rainfall rate estimates based on
Ferraro et al. (2005) over a 3-year period. They found values of CC around
0.55 (on an annual scale, at 0.25
In the comparison of PNPR v2 and PNPR v1 retrievals, performed over the African area and based on a 2-year period of coincident observations of ATMS and AMSU/MHS radiometers with TRMM-PR, an appreciably better performance of PNPR v2 has been evidenced by statistical indexes (e.g., CC equal to 0.71 for PNPR v2, vs. 0.68 for PNPR v1 over vegetated land, and equal to 0.69 for PNPR v2, vs. 0.61 for PNPR v1 over ocean) and by a general improvement of the estimate of low precipitation, mostly over ocean. The resulting differences can likely be attributed to improvements in the design of the neural network and also to the best technical features of ATMS compared to AMSU/MHS.
Overall, the two versions of PNPR algorithm have shown a general consistency in the results, as expected considering that both are based on the same physical basis (the training databases are based on the same cloud-resolving model and to the same radiative transfer model). It is worth noting that the achievement of consistency between products derived from different sensors is very relevant in the current GPM mission era, with constellation satellites (equipped with cross-track or conical scanning microwave radiometers) contributing to global coverage and higher temporal sampling of precipitation. This aspect has become very important also within the EUMETSAT H-SAF program and represents a guideline for the development of PMW precipitation products. PNPR v2 and PNPR v1 for ATMS and AMSU/MHS, as well as other products for conically scanning radiometers (e.g., CDRD for SSMIS – Casella et al., 2013, Sanò et al., 2013), and new products for the other constellation radiometers are developed within H-SAF in this direction, with foreseen improvements of derived MW/IR products used in operational hydrology and near-real-time precipitation monitoring applications.
The results, however, have revealed a slight tendency of PNPR v2 to underestimate moderate to high precipitation, mostly over land, and overestimate moderate to light precipitation over the ocean, especially compared to GPM-KuPR product over the North Atlantic Ocean. Besides well-known issues affecting PMW precipitation retrieval, such as non-uniform beam-filling effects related to small-scale rainfall structures associated with local convection and difficulties in the retrieval of warm or shallow rain processes, in addition to the lack of low-frequency channels very useful for precipitation retrieval over ocean, other issues might be related to the use of spaceborne radar products as reference. The impact of sample volume discrepancies between radiometers and spaceborne radars, and uncertainties in the spaceborne radar estimates (due to attenuation correction, sensitivity thresholds, non-uniform beam-filling effect), needs to be evaluated when using spaceborne radar precipitation estimates as reference. PNPR v2 will undergo thorough extensive validation within the EUMETSAT H-SAF program carried out by the H-SAF Precipitation Products Validation Service (Puca et al., 2014), using ground-based radars and rain gauges over Europe and, in limited areas, over Africa, which will be useful to clarify some of these issues.
In spite of the above mentioned limitations, this study shows that the TRMM and GPM spaceborne radars can be very useful for an extensive verification, over long time periods, of consistency and accuracy of instantaneous precipitation rate estimates from different sensors. The use of spaceborne radars as reference overcomes some of the limitations in the use of ground-based data (such as inhomogeneity in their technical characteristics and data treatment, limited coverage, and beam blocking), providing consistent measurements around the globe, including remote areas where ground-based data are scarce or not available and oceans.
The data used in the research concerning the ATMS radiometer brightness
temperature are provided by the National Oceanic and Atmospheric
Administration (NOAA), Comprehensive Large Array-data Stewardship System
(CLASS), and are available at
The authors would like to thank the National Aeronautics and Space
Administration (NASA,