Contributors: H. Konrad (DWD), T. Sikorski (DWD), M. Schröder (DWD), O. Bobryshev (DWD), L. Bagaglini (CNR), P. Sanò (CNR), G. Panegrossi (CNR), E. Cattani (CNR)

History of modifications

List of datasets covered by this document

Related documents

Acronyms

Scope of the document

This Product Quality Assessment Report summarizes the results of the product quality assessment of the Copernicus micrOwave-based gloBal pRecipitAtion (COBRA) Climate Data Record. The methodology has been outlined in the respective Product Quality Assurance Document [D1]. COBRA daily and monthly precipitation is based on merged instantaneous precipitation rate estimates derived from MW radiometers. Estimates from conically scanning MW imagers are obtained by applying methodologies of the Hamburg Ocean Atmosphere Parameters and Fluxes from Satellite (HOAPS) produced in the scope of the EUMETSAT CM SAF, while estimates from cross-track scanning MW sounders are obtained through the newly developed Passive microwave Neural network Precipitation Retrieval for Climate Applications (PNPR-CLIM). HOAPS and PNPR-CLIM sub-datasets are evaluated separately against various reference datasets, and compared with each other. The merged gridded daily and monthly COBRA datasets are evaluated against various reference datasets.

Executive summary

COBRA global gridded daily and monthly precipitation is derived from two source datasets, PNPR-CLIM and HOAPS v4. HOAPS v4 has been assessed on its own elsewhere and it is only referenced here. PNPR-CLIM has been developed in C3S. The performance of PNPR-CLIM Level 2 instantaneous precipitation rates from 2015–2017 is assessed here in detail. When assessed against MRMS, a radar-based precipitation dataset over the Contiguous United States (CONUS), the mean error and RMSE are -0.007 mm/h and 0.6 mm/h, respectively, with a correlation coefficient of ~0.7. At low latitudes, COBRA agrees better with GPCP, a calibrated satellite-based climate data record, than the ERA5 reanalysis product. At high latitudes, the situation is reversed.

PNPR-CLIM and HOAPS v4 precipitation rates are assessed against each other in terms of Level 2 instantaneous precipitation rates and Level 3, hourly gridded precipitation rates; the latter is an intermediate step in the processing towards the daily and monthly gridded data. The instantaneous comparison yields an overall hit rate of ~90%, a mean error of ~-0.04 mm/h, an RMSE of ~0.24 mm/h, and a correlation coefficient of ~0.6 between the two sub-datasets. The comparison of hourly gridded values is improved by the bias correction implemented in the processing, so that overall the hit rate is at 79%, with a mean error of -0.01 mm/h, a RMSE of 0.32 mm/h, and a correlation coefficient of 0.74. The bias correction proves most effective in straightening the effects of high-latitude NOAA17 observations that initially introduced higher deviations.

Global averages of daily and monthly COBRA precipitation rates are mostly within the targeted limits, both in terms of deviations from respective GPCP averages, and a sufficiently low overall trend in the differences.

COBRA reproduces low-latitude precipitation over oceans and continent quite well, compared to GPCP and ERA5. Over high latitudes, GPCP and ERA5 see generally higher precipitation. A comparison with the raingauge-based dataset GPCC shows that COBRA tends to underestimate precipitation over land (median difference of 0.3 mm/d). The comparison with MRMS and NIMROD radar-based precipitation datasets, over CONUS and Europe, respectively, at daily temporal resolution shows that while COBRA mean errors might be higher than those of GPCP and ERA5, COBRA often outperforms GPCP in terms of RMSE and correlation coefficients. A different behaviour over open water and land is visible in the comparison with NIMROD.

Finally, for the sake of completeness and because COBRA daily and monthly precipitation rates are computed differently in terms of aggregating the instantaneous observations, respective monthly mean values of the daily product are compared to the actual monthly product. Differences exist, but they are small compared to previously reported differences to other datasets.

1 Product validation methodology

1.1 Validation of the COBRA datasets and its components

During COBRA production, the precipitation rate estimates obtained from PNPR-CLIM are complemented by precipitation rates estimates obtained from the HOAPS v4 (see ATBD [D2]) (over ocean only). Due to the different status of the components of the COBRA datasets, we assess different instances of the product.

With PNPR-CLIM being a newly developed and previously untested retrieval algorithm for the MW sounders AMSU-B and MHS (see ATBD [D2]), section 2.1 is dedicated to assessing the respective retrieved precipitation rates regionally and globally. HOAPS v4 data on their own have been assessed in the framework of CM SAF activities, see the respective validation report [D4].

In order to assess the level of consistency between PNPR-CLIM and HOAPS v4, they are compared to each other based on collocated data pairs. Section 2.2.1 covers the comparison of instantaneous, spatially irregularly sampled Level 2 (L2) estimates. In section 2.2.2, we compare the respective precipitation rate estimates averaged at hourly time scale on a regular 1° × 1° grid, an intermediate step in the production of the COBRA daily data product. Here, we are also able to assess the impact of the bias correction which is carried out on this instance of the COBRA dataset production.

The final daily and monthly resolved COBRA data products on the regular 1° × 1° grid are assessed in section 2.3 at various levels of detail in both spatial and temporal dimensions. A brief overview of deviations between COBRA data at daily and monthly resolution due to different processing choices is included in section 2.3.4.

1.2 Validating datasets

The validating datasets are listed and explained in detail in the respective PQAD [D1]. We include here a brief introduction with respective citations.

Precipitation rates from the current version of the ECMWF reanalysis (ERA5) product (reference C3S, 2017) are compared to both the PNPR-CLIM-only precipitation rates as well as to the final COBRA gridded datasets at global scale. For a comparison at daily time scale, the respective hourly ERA5 product is accumulated to obtain daily precipitation. The spatial fields of both daily and monthly ERA5 products are averaged over 4 × 4 grid cells to match the coarser 1° × 1° grid of the COBRA data.

The Global Precipitation Climatology Project (GPCP) provides satellite-based global estimates of precipitation at monthly resolution on a 2.5° × 2.5° spatial grid since 1979 (GPCP monthly v2.3, Adler et al. 2016, 2018) and at daily resolution on a 1° × 1° spatial grid since 1996 (GPCP daily v1.3, Adler et al. 2017, Huffman et al. 2001). The datasets are calibrated against the GPCC rain gauge dataset (see below). As for ERA5, the comparison with GPCP is carried out both for the PNPR-CLIM-only dataset and for the final COBRA datasets.

The GPCC “monitoring” dataset v2020 available on a global 1° × 1° grid at monthly resolution (Schneider et al., 2020) is compared to the final monthly gridded COBRA dataset. A comparison at daily resolution is not carried out because of the mismatch in the definition of one day between COBRA and GPCC (start and end times of a day in local times in GPCC and in UTC in COBRA).

The precipitation rate estimates provided by the MRMS over CONUS, and based on various ground-based observations and predictions (Zhang et al., 2016), are compared to PNPR-CLIM precipitation rate estimates and to the final COBRA daily product. For parts of the analysis, the dataset is regridded in space and time to match the respective resolution of PNPR-CLIM / COBRA datasets. The MRMS data are filtered with respect to their quality index.

The United Kingdom’s Met Office provides precipitation estimates across Europe based on observations at weather radar stations on a 5 km grid (NIMROD; reference Met Office, 2003). These are regridded to match the coarser COBRA 1° × 1° daily grid, and compared to the final COBRA daily product.

1.3 Overview of the validation methodology

The datasets listed in section 1.1 are validated against the datasets listed in sections 1.2 (and section 1.1, in the case of PNPR-CLIM vs. HOAPS v4 and COBRA daily vs. COBRA monthly analyses) in terms of statistical scores and metrics and distributions (of differences) in the case of collocated data pairs, spatial maps of differences of climatological means (also error components, see PQAD [D1]), and time series of global averages. In the case of the PNPR-CLIM-only analysis, single scenes are also compared for a qualitative visual assessment.

The methodological details as well as formal definitions of the evaluated statistical scores and measures are described in the PQAD [D1].

2 Validation results

2.1 Stand-alone verification of PNPR-CLIM

2.1.1 Regional verification

In this section instantaneous PNPR-CLIM L2 precipitation rates from 2015–2017 are compared with coincident precipitation estimates from MRMS over the CONUS area. MRMS data are averaged over a regular 15 km × 15 km grid.

Coincident pixels are identified using a nearest-neighbour approach in both time and space. Only the best available observations are considered in the analysis, identified by RadarQualityIndex_015cdeg > 0.9 (RQI). The average RQI is shown in figure 1. In addition, for the binary classification analysis, an MRMS pixel is classified as precipitating if PrecipRate_015cdeg > 0.1 mm/h and as non-precipitating if PrecipRate < 0.1 mm/h only for less than 10% of the original cells within the fixed 15 km × 15 km grid box.

Figure 1: One-year average MRMS Radar Quality Index at 15 km × 15 km resolution over CONUS.

HSS, POD and FAR are computed at different values of the detection threshold δ. This means that the scores are evaluated over the reduced population of pixels with MRMS precipitation rate either equal to 0 mm/h or greater than δ. The results are shown in Figure 2. HSS and POD remain above 0.6 and 0.8 between 0 mm/h and 0.3 mm/h, whereas FAR decreases from about 0.5 to 0.33 in the same range. It is worth noting that Bagaglini et al. (2021) proved these results to be consistently better than those achieved by other passive microwave precipitation retrieval algorithms based on the same radiometers (MHS).

Figure 2: HSS (left panel), POD (right panel) and FAR (bottom panel) achieved by PNPR-CLIM against MRMS for different detection thresholds. The Precipitation/No-precipitation convention is specified in the text.

In figure 3 the joint normalized densities of PNPR-CLIM and MRMS are drawn and in table 1 the values of the three statistical indices ME, RMSE and CC are reported. A substantial agreement between the products is visible over the entire range. In detail, above 10 mm/h PNPR-CLIM slightly underestimates MRMS while below 10 mm/h the opposite is true. The same behaviour, related to other passive microwave precipitation products, was also observed by Bagaglini et al. (2021), where, in addition, they found similar values for the statistical indices shown in table 1.

Figure 3: Joint normalized densities of PNPR-CLIM and MRMS. The black line denotes the bisector.

Table 1: ME, RMSE and CC between PNPR-CLIM and MRMS using the entire dataset.

In the following, two case studies of MHS/AMSU-B overpasses over the CONUS area are discussed. Both the PNPR-CLIM and MRMS instantaneous precipitation rates (the latter regridded to the MHS original grid) are displayed.

Figure 4: Comparisons between PNPR-CLIM instantaneous precipitation rate retrieval (left panels) and the radar-based MRMS precipitation field regridded to the MHS original grid considering the antenna pattern (right panels) for two different scenes (upper and lower panels). The scene in the upper panels refers to the MHS, on-board MetOp-B, overpass at 02:37 UTC on 2017-04-03. The scene in the lower panels refers to the MHS, on-board NOAA19. overpass at 12:15 UTC on 2017-03-10.

Note that, in this case, no additional filtering has been applied to the MRMS data. The top panels in figure 4 show the MHS overpass at 02:37 UTC on 2017-04-03. The scene is characterized by several fine precipitating structures all across the central US with a wider and more intense precipitating area over the Louisiana and Mississippi states. In the images, the spatial coherence between the satellite and radar products is remarkable. The thin northern structures captured by the radar observations are also resolved by the satellite product. It should be stressed that the MRMS product has been regridded to the MHS grid (considering the antenna pattern) and therefore the radar rates have been smoothed. The most appreciable difference between the two products appears in the magnitude of the precipitation peak, characterizing the southern precipitating structure. There, the PNPR-CLIM estimates above 10 mm/h extend over a large area, whereas the MRMS ones are occasionally below 10 mm/h.

The bottom panels in figure 4, show a second case study on 2017-03-10 at 02:37 UTC on the US East coast. An extended precipitation front elongates, almost continuously, from Texas to Massachusetts. It is worth noting that all the precipitating areas with rates above 0.2 mm/h are coherently identified by both products, with only small differences between the two products. Some high values (above 10 mm/h) estimated by PNPR-CLIM over the Gulf of Mexico are weakly represented by MRMS (rates below 1 mm/h). In this area, however, the MRMS quality is quite low (see figure 1).

2.1.2 Global verification

The daily gridded ERA5 and GPCP data from 2015 to 2017 are selected for the global verification of PNPR-CLIM. However, since neither ERA5 nor GPCP can be considered as reference (truth) datasets, this verification is intended as an inter-comparison for consistency verification on global scale of PNPR-CLIM with other widely used global datasets. PNPR-CLIM provides instantaneous L2 precipitation rates based on MHS measurements only and as such, its daily estimates are necessarily affected by sampling errors, which are not present in ERA5 and partially overcome in GPCP. Bagaglini et al. (2021) carried out an extended comparative analysis between GPCP and ERA5 daily precipitation and the daily estimates derived from PNPR-CLIM and another MHS L2 instantaneous precipitation rate product, sharing the same temporal sampling issues.

In figure 5 the overall mean errors (bottom panels) of the two daily precipitation products PNPR-CLIM and ERA5 compared to GPCP are shown. In the same figure, the mean error components given by the hit error, missed precipitation and false precipitation, computed with respect to a detection threshold of 1 mm/d, are also displayed.

Figure 5: Mean errors of PNPR-CLIM (first column) and ERA5 (second column) with respect to GPCP over the period 2015–2017. Mean Errors (fourth row) are decomposed into mean Hit Errors (first row), Missed Precipitations (second row) and False Precipitations (third row) by applying a detection threshold of 1 mm/d.

The overall errors of PNPR-CLIM are characterized by significant underestimates over high latitude ocean areas and sparse overestimates over land (central Africa, southern America). In contrast, ERA5 errors are more uniformly positive, particularly over the ocean. However, some underestimates (relative to GPCP) are seen over land: in central Africa, southern America, and Northern Europe, weak underestimations are visible.

False and missed precipitation maps of PNPR-CLIM and ERA5 show opposite patterns: ERA5 retrieves much more precipitation than GPCP, resulting in higher (lower) values of false precipitation (missed precipitation). PNPR-CLIM, in contrast, misses more precipitation, especially over ocean and at high latitudes.

Finally, the hit errors seem to be more pronounced for PNPR-CLIM than for ERA5. The ERA5 hit errors, in particular, are uniformly lower except in certain highly localised, regions in the tropics (positive over ocean and negative over land).

The PNPR-CLIM and ERA5 RMSE with respect to GPCP shown in figure 6 (top panels), highlight the high variability of the overall errors. It turns out that all the products show very high RMSE values (up to 10 mm/d) in the wettest regions of the globe (see the GPCP overall mean precipitation in figure 7 for context). There, the high precipitation activity, often characterized by intense convection and extremely pronounced diurnal and seasonal cycles, leads to notable error accumulations. It is worth noting that the ERA5 RMSE peaks along the mean inter-tropical convergence zone. In the same region, PNPR-CLIM shows a smoother RMSE.

The CC global patterns (figure 6, bottom panels) show that, below 40° N/S, PNPR-CLIM has the best agreement with GPCP, both over ocean and land. Above 40° N/S, instead, the CC values decrease below those of ERA5. This behaviour was discussed by Bagaglini et al. (2021) and it is mainly due to the PNPR-CLIM limitations in snowfall detection and quantification, and to the lack of microwave channels in MHS (or AMSU-B) needed to characterize the cold surface background (mostly occurring in the winter season) at the time of the satellite overpass (Turk et al., 2021, Panegrossi et al., 2021).

Figure 6: RMSE (first row) and CC (second row) of PNPR-CLIM and ERA5 (first and second column respectively) with respect to GPCP over the period 2015–2017. 

Figure 7: GPCP 2015–2017 mean precipitation.

2.2 Mutual assessment of PNPR-CLIM and HOAPS v4.0

2.2.1 Comparison of PNPR-CLIM and HOAPS v4.0 instantaneous observations

Co-located pairs of instantaneous L2 precipitation rate estimates in PNPR-CLIM and HOAPS v4 have been identified. Only pairs within 15 km and 15 minutes spatial/temporal distance and from inner scan positions, as indicated in table 2, are retained. The filter on the scan positions ensures a similar spatial resolution and viewing geometry through local atmospheric conditions in these scan positions¹ Identified data pairs cover only the years 2000–2009, 2012 and 2015, with only partial coverage in certain years. For an overview of spatiotemporal coverage of co-located data pairs, figure 9 later in the analysis shows the time series of monthly ME and RMSE between PNPR-CLIM and HOAPS v4. Most of the months present in this graph are processed completely, i.e. all data pairs that meet the above criteria have been identified. The choice of processed months ensures that each platform in both datasets is present in the comparison.

Table 2: Scan positions per instrument of the inner scan lines around the swath centre. As the TMI field of view is much smaller than the others (see the ATBD [D2]), we reject respective data here².

Table 3 contains the numbers of collocated data pairs, the respective detection statistics as well as mean error (ME) and RMSE and the correlation coefficient (CC) for various scenarios. Here, HOAPS v4 has been chosen as the reference dataset for PNPR-CLIM. All statistics refer to the entire timeline of collocated data pairs.

The “Swath Centre” scenario comprises all available data pairs as per the above requirements. HOAPS v4 sees on average 0.04 mm/h higher precipitation rates than PNPR-CLIM (ME). The RMSE lies at 0.24 mm/h, and the CC is above 0.6. The “Latitude” scenarios filter the “Swath Centre” pairs with respect to their zonal position in three latitude bands. Most statistics are best in low latitudes, but the RMSE is higher there, most likely due to heavy precipitation being represented differently in the two datasets. Finally, the “NOAA15” scenario uses the “Swath Centre” data pairs and retains only those for which NOAA15 is the respective PNPR-CLIM platform. This subset performs significantly worse than the “Swath Centre” set, in terms of HSS, ME, and CC. The deterioration of observations by NOAA15, whose identification led to NOAA15 data being phased out as soon as NOAA16 data were available, has already been discussed in the PUGS [D3] and will be discussed, for example, in section 2.3.2.1 of this document. The results here illustrate that the deterioration is already present in the instantaneous precipitation rate estimates.

Figure 8 shows the two-dimensional histogram of PNPR-CLIM and HOAPS v4 precipitation rates and the histogram of differences between the two datasets for the “Swath Centre” scenario. The best-fit line in the two-dimensional histogram and the negatively skewed histogram of differences confirm the negative bias in table 3, implying that HOAPS v4 sees on average higher precipitation than PNPR-CLIM, despite PNPR-CLIM in principle seeing higher precipitation rates (figure 8, left panel). One source of this overestimation by HOAPS v4 is the NOAA17 platform (see also section 2.2.2).

Table 3: Number of collocated data pairs, mean error (ME), root mean square error (RMSE), correlation coefficient (CC) and skill scores of PNPR-CLIM L2 data with respect to HOAPS L2 data. The number of non-precipitation events implies both datasets see zero precipitation in the respective pair.

Figure 8: Scatter plot PNPR-CLIM vs. HOAPS (left panel) and associated histogram of differences PNPR-CLIM minus HOAPS (right panel) in the "Swath Centre" scenario. In the left panel, the grey solid line is the identity, the line of best linear fit is displayed as black dash-dotted line. 

Figure 9: Time series of monthly ME (blue) and RMSE (orange) for PNPR-CLIM minus HOAPS L2 data in the "Swath Centre" scenario, with each marker representing one month. 

Figure 9 shows the temporal evolution of monthly ME and RMSE between PNPR-CLIM and HOAPS v4. Note in particular that RMSE peaks in the years 2002–2010. This is most likely related to features in NOAA17 observations, see section 2.2.2 for a more comprehensive analysis.

2.2.2 Comparison of PNPR-CLIM and HOAPS v4.0 hourly gridded observations

During the production of COBRA, the L2 instantaneous precipitation rate estimates are aggregated to hourly values on a 1° × 1° spatial grid, referenced as 1DH, see the ATBD [D2]. An assessment of these 1DH values similar to the one above offers the main advantage that the impact of collocation errors is reduced (for example discrepancies related to the different viewing geometry of the radiometers, or to the different time of the satellite overpass over the observed scene). We are also able to assess the effect of the bias correction via quantile mapping, which has been applied to the 1DH data, see the ATBD [D2]. Finally, the 1DH database is much smaller than the original L2 instantaneous database, which allows a full screening of all possible data pairs. On the other hand, the main disadvantage is that the actual distance in space and time between the underlying L2 observations that went into the gridded averages cannot be retrieved, i.e. could stem from opposite areas of the respective grid cell and represent very different atmospheric situations.

Here, we compare PNPR-CLIM and HOAPS v4 1DH collocated data pairs (i.e., occurring in the same hourly interval and in the same 1° × 1° grid cell). We exclude data pairs at latitudes above ±75°, due to HOAPS v4 data not being available over land or ice-covered ocean. We consider spatial variations in the consistency of PNPR-CLIM and HOAPS v4 databases by filtering the data pairs for three different sub-regions of the Earth, defined by latitudinal bounds at -75°, -25°, +25°, and +75°. The temporal variation is assessed by computing monthly mean error and RMSE. In addition, we also evaluate collocated data pairs internally for both the PNPR-CLIM and the HOAPS v4 databases, i.e. compare collocated 1DH precipitation rate estimates by each two PNPR-CLIM platforms, and likewise for HOAPS v4. It should be noted here that the internal PNPR-CLIM comparison includes data pairs over land whereas the other comparisons do not, due to HOAPS v4 only being available over ice-free ocean.

Table 4 summarizes the results of the intercomparison (PNPR-CLIM vs. HOAPS v4) and the internal comparisons over the entire time period (2000–2017) in terms of various statistical scores and measures. The slightly negative mean error between PNPR-CLIM and HOAPS v4 implies that HOAPS sees higher precipitation rates on average, see also the L2 comparison in section 2.2.1. As can be expected when different types of radiometers and algorithms are used, the datasets compare better internally (PNPR-CLIM vs. PNPR-CLIM, HOAPS v4 vs. HOAPS v4) than against each other (PNPR-CIM and HOAPS v4) both in terms of the statistical scores that mostly reflect detection (hit rate, HSS) and the ME and RMSE as well as the correlation. The PNPR-CLIM and HOAPS v4 data agree best at low latitudes, in every aspect. The comparison in northern hemisphere outside the tropics mostly performs worst (likely due to the larger uncertainties in winter/high latitude precipitation (snowfall) retrieval, see also section 2.1.2).

As envisaged, the bias correction reduces the ME and RMSE between the datasets. A quite significant improvement by a factor 3–5 is achieved outside the tropics. At low latitudes, the small improvement in the RMSE is accompanied by a modest deterioration in the ME. Also, the bias correction greatly improves the correlation between PNPR-CLIM and HOAPS v4 and harmonizes the ME and RMSE over the different latitudinal ranges. However, it does not bring ME, RMSE and the correlation coefficient to the level of the intra-dataset comparisons (HOAPS vs. HOAPS, PNPR-CLIM vs. PNPR-CLIM) where ME values are almost zero, RMSE is 2/3 or 1/3 of the inter-dataset level, and correlations are close to 0.9.

Table 4: Statistical results of the comparison of hourly gridded precipitation rate estimates in PNPR-CLIM (P) and HOAPS v4 (H). Collocated data pairs inside the specified latitudinal bands over the entire time period (2000–2017) are evaluated with the “validating dataset” as reference dataset for the “validated dataset”. ME and RMSE refer to the differences between the validated and validating datasets. Note that the scores (hit rate and HSS) vary only a little between uncorrected and bias-corrected data pairs, because only zero-precipitation events are mapped to zero precipitation during the bias correction. Small variations result from discarding certain data, see the ATBD [D2], thus reducing the database. Here, only the values for the bias-corrected data are given.

Figure 10 shows the distributions of differences in the 1DH PNPR-CLIM vs. HOAPS v4 comparison over the entire time period and the full latitudinal range, for uncorrected and bias-corrected data pairs. The distribution of differences of uncorrected values is slightly yet visibly skewed to the left (PNPR-CLIM underestimates HOAPS v4), which is not the case in the bias-corrected version, at least for small deviations from zero (≤ 0.25 mm/h).

Figure 10: Histograms of the distributions of differences between collocated data pairs from PNPR-CLIM and HOAPS over the entire time period (2000 and 2017) between -75° and +75° latitude. Data pairs which both see zero precipitation (difference = zero in these cases) have been removed for a better visualisation. The left panel represents the differences without bias-correction. The differences in the panel of the right include the bias correction. The numbers in the y-axis labels give the respective total number of collocated data pairs which is not exactly the same due to data filtering during the bias correction procedure. 

Figure 11 shows the temporal evolution of monthly ME and RMSE between 1DH PNPR-CLIM and HOAPS v4 collocated data pairs. Uncorrected and bias-corrected mean are almost identical outside the window from Jan 2003 to Dec 2010, although very slightly higher RMSE values exist in the uncorrected data. In the specified window, ME show a minimum of about -0.5 mm/h and RMSE go up to almost 3 mm/h. An annual cycle is visible in the comparison. The bias correction eliminates both the annual cycle and the increase in mean and RMS differences.
The window fits the time period in which NOAA17 data processed through PNPR-CLIM are included in COBRA. The problem is restricted to mid-to-high latitudes, mostly in the northern hemisphere, as the bias correction change the comparison only to a small extent in low latitudes and the comparison of uncorrected data performs worst in northern mid-to-high latitudes (table 4).

Figure 11: Time series of the ME (thick lines) and RMSE (thin lines) in monthly ensembles of collocated data pairs from PNPR-CLIM and HOAPS v4 with latitudes between -75° and +75°, see also the overall mean and RMS differences in table 4. Black lines show the respective quantities for uncorrected data; red lines for differences of bias-corrected data. 

Interestingly, the statistics of the PNPR-CLIM/HOAPS v4.0 1DH comparison are partially worse than for the comparison of instantaneous data pairs (first row in table 3 vs. first row in table 4). It should be noted, however, that no filter on scan positions or spatiotemporal proximity has been applied in the collocation of 1DH data, except for comparing data in the same 1° × 1° grid cells in the same hourly interval. In addition, hit rates of 1DH data pairs may suffer from the mixing of zero- and non-zero- precipitation events in the processing towards 1DH values and possibly higher spatiotemporal distance between actual measurements. Another factor may be the effect of NOAA17 data on ME and RMSE. NOAA17 data are fully present in the 1DH comparison but only partially in the comparison of L2 instantaneous data.
In summary, the above findings indicate that PNPR-CLIM and HOAPS v4 compare reasonably well, especially in terms of bias-corrected 1DH gridded data pairs. The fitness for purpose of the bias correction procedures is clearly demonstrated by the improvements in the 1DH comparison.

2.3 Assessment of merged, gridded datasets

2.3.1 Assessment of spatiotemporal completeness

For each temporal instance (month for the monthly product, day for daily product) in the 2000-2017 data record, the respective COBRA file exists. The monthly data are complete in the sense that every grid cell at every temporal instance contains a valid value. This is not entirely the case for the daily data. The number of gaps in the 1° × 1° grid per day are displayed in figure 12. It is obvious that the situation improves over the time. With the inclusion of NOAA18 in late 2005, there are always three platforms available in the PNPR-CLIM subset of input data of COBRA (figure 1 in the COBRA PUGS [D3]), which is the only source of information over land in COBRA. This coverage is apparently necessary to go to virtually zero gaps per day. The previous step from about 100 to mostly below 10 gaps per day is related to the inclusion of NOAA17 in PNPR-CLIM, which leads to two platforms being available in this subset.

Figure 12: Number of grid cells in daily data that are not filled with a valid value. The overall amount of 1° × 1° grid cells is of course 64800.

2.3.2 Comparison with global datasets

2.3.2.1 Time series of global averages

Here we evaluate the temporal evolution of global mean precipitation rates of COBRA against different reference datasets. figure 13 illustrates the respective monthly global mean precipitation rates for COBRA, GPCP, and ERA5, as well as respective differences between COBRA and the other two datasets. Figure 26 in the Appendix shows the respective daily comparison. The main feature in the daily time series is the expected increased temporal variability (compared with the monthly data). figure 27 in the Appendix shows the respective comparison of monthly data with GPCC. In this case, the averages are only computed over land, where GPCC is in principle available, and for monthly data, due to the different start and end times of a day in GPCC (local times) and COBRA (UTC). Statistics on the time series of differences for the various comparisons are summarized in table 5.

Figure 13: Upper panel: Time series of monthly global mean precipitation in COBRA, GPCP v2.3 (monthly), and ERA5. Lower panel: Time series of differences (solid lines) of monthly global mean precipitation, i.e., COBRA minus the respective reference datasets GPCP and ERA5. The respective trends, based on the data from 2001-04-01 (NOAA15 phase-out, see main text), are displayed as dashed lines. The black dash-dotted lines mark the ±0.3 mm/d accuracy limits. 

Table 5: Basic statistics for global mean differences of daily and monthly COBRA precipitation estimates with GPCP, ERA5 and GPCC datasets as references. GPCC covers land only, so COBRA data have been filtered respectively for this comparison. Minimum, maximum, mean, median, root mean square deviation and both quantiles are given in mm/d. The slope's unit is mm/d/decade. The second-to-last column indicates the fraction of temporal instances at which the target requirement of absolute differences between COBRA and a reference dataset staying below 0.3 mm/d is met. The last column ("slope") contains the linear trend in the respective time series of differences, see main text.

Requirements for the accuracy and stability of the data have been formulated with respect to global mean precipitation rates, see the PUGS [D3]. The accuracy, i.e., absolute difference between the COBRA data and an appropriate reference dataset, shall remain below 0.3 mm/d at all times. The stability, i.e., the trend in the respective differences, shall remain below 0.034 mm/d/decade.

The COBRA data see the lowest global mean values (total or over land only), compared to the other three datasets, manifesting for example in the negative mean and median values in table 5. ERA5 sees the highest offset, such that the accuracy requirement of 0.3 mm/d for differences between COBRA and ERA5 is violated most of the time. In contrast, the comparison of COBRA with GPCP and GPCC data meets this requirement most of the time. The use of the low-quality data from NOAA15 in the PNPR-CLIM database until 2001-04-01 leads to a significant bias towards lower values (figure 13). When restricting the comparison to after 2001-04-01, the requirement is met almost at any time in the comparison of monthly data, see the PUGS [D3]. The comparison of daily data is affected by a higher temporal (sub-monthly) variability, which would in principle require a more relaxed target, see the PUGS [D3]. The seasonal cycle visible in COBRA and ERA5 is not present in GPCP. In principle, the comparison with GPCP produces the lowest amount of violations of the accuracy requirement. As the GPCP dataset is influenced by the GPCC dataset over land, we hypothesize that mismatches between COBRA and GPCC are moderated over the oceans when comparing COBRA to GPCP.

The trends in the differences of COBRA and ERA5 are negligible. In contrast, the trend in the monthly differences with GPCP just meet the formulated stability requirement. In the comparison with GPCC (land only), the requirement (which had been formulated for a global comparison, averaging over both land and ocean) is clearly violated. Differences in the trends between daily and monthly resolutions originate from the higher level of temporal variability in the daily data.

Please note that the differences in global means are explained by the findings presented in sections 2.3.2.2 to 2.3.2.6

2.3.2.2 Climatological means

In figure 14 the climatological mean of the COBRA 1DD product together with those of GPCP and ERA5 are shown, as well as the inter-product mean differences. A general agreement in the global precipitation distribution is observed among the products. The most appreciable differences are the estimates in central Africa, central America, north-eastern Russia and northern Europe and, finally, the north-eastern Pacific. Over central Africa, COBRA estimates more precipitation than GPCP and ERA5, the latter giving the smallest estimate. In central America, ERA5 estimates are larger than those of both COBRA and GPCP, with least precipitation provided by COBRA. Over north-eastern Russia and northern Europe, the COBRA estimates are smaller than both GPCP and ERA5. Finally, over the north-eastern Pacific, GPCP gives lowest estimates. The mean differences highlight these patterns and, in addition, the different behavior of GPCP over the central Pacific with respect to the other two products. COBRA data underestimate precipitation over the Southern Ocean and they display an anomalous precipitation area over the Antarctica plateau. The latter, as discussed in section 2.3.2.3 and additionally illustrated in appendix 5.2 (figure 28, period 2000-2005), is mainly related to an issue with the NOAA15 data in 2000.

Figure 14: Mean precipitation of COBRA, GPCP and ERA5 (first row) together with the relative mean differences (second row) over the entire period 2000–2017. 

Bagaglini et al. (2021) found that the underestimates of PNPR-CLIM at high latitudes (specifically over northern Europe and Russia as well as over the Southern Ocean), are common issues experienced by other passive microwave precipitation products. The HOAPS estimates, which are derived from MW radiometers equipped with additional channels, partially correct this feature (see section 2.2.2), although it is nonetheless still present in the COBRA dataset.

In summary, there is a certain coherence among the various products, with relevant but manageable (and well understood) differences which turn out to be related to the various and different methodologies that are behind each product, and to intrinsic limitation of the COBRA MW-based approach, namely, radiometer channels assortment and spatial resolution, and temporal sampling issues.

2.3.2.3 Error components

The error decomposition of the COBRA, GPCP and ERA5 1DD products, shown in figure 15, confirms the results outlined in section 2.3.2.2. ERA5 underestimates the convective precipitation in central Africa whereas it overestimates the precipitation over the central Pacific. The low estimates of COBRA in high latitudes, instead, are mainly due to missed precipitation, which is likely due to sensor limitations (see the final paragraph in section 2.1.2). Finally, the GPCP estimates turn out to be much more conservative with respect to the other products, manifesting in high false precipitation values of COBRA and, more severely, ERA5. Finally, the anomalous feature observed in Antarctica for COBRA, as shown in appendix 5.3 (figure 30 – figure 32), is mainly limited to the year 2000 estimates, uniquely based on the NOAA15 AMSU-B measurements affected by large uncertainties. In conclusion, the three products show peculiar but comparable uncertainties between themselves.

Figure 15: Error components (from top to bottom: hit error, missed precipitation and false precipitation) between the various daily gridded precipitation datasets (from left to right: COBRA referred to GPCP, COBRA referred to ERA5 and ERA5 referred to GPCP) over the period 2000–2017. A detection threshold of 1 mm/d has been applied.

Figure 16: Zonal means of COBRA, GPCP and ERA5 over the period 2000–2017 and computed globally (first panel), over ocean (second panel) and over land (third panel).

2.3.2.4 Zonal means

The zonal means, shown in figure 16, offer a more concise picture of the relationships among the three 1DD products: COBRA, ERA5 and GPCP. COBRA underestimates the precipitation outside the tropics (above/below 45° N/S), both over land and ocean. This is mainly due to the light precipitation in cold-dry conditions (mainly snowfall) occurring during the winter season (see also the seasonal zonal means in appendix 5.4, figure 33), which are hardly detected by passive MW sensors (especially over snow-covered land and sea ice), as highlighted by several studies (e.g. Panegrossi et al., 2021).

In contrast, in the tropics the COBRA precipitation estimates are close to those of the other products. Over ocean, COBRA estimates track the higher ERA5 rates closely, while, over land, they are lower than both GPCP and ERA5 in a narrow band at about 4 °N.

2.3.2.5 Analysis of spatiotemporally collocated grid cells

We analyze the ensembles in each of the three global 1DD products COBRA, GPCP, and ERA5 in terms of their probability accumulation densities. Accumulation densities (or probability densities by volume) are directly related to probability densities (or probability densities by occurrence) and are generally used to analyze precipitation datasets (Kirstetter et al., 2012). Raw precipitation distributions are commonly skewed toward zero. In contrast, the accumulations help to highlight the precipitation amounts that contribute most to the overall precipitation volume. Accumulation densities are probability densities in which each occurrence has been weighted by the product of its precipitation value, and by the cosine of the latitude. Therefore, in any given precipitation bin, the accumulation density  is proportional to the weighted sum of the estimates  within the fixed bin and, being a density, has unit L1 norm.

Formally:

where the proportionality constant  turns out to be overall observed volume.

In figure 17 the accumulation densities (together with the observed volumes) of COBRA, ERA5 and GPCP 1DD products are shown. The statistics are categorized by latitude band and surface type. The previously mentioned high-latitude underestimates by COBRA are again visible, in terms of both total volumes and densities. Indeed, GPCP and ERA5 show larger volumes and their densities have higher values between 4 and 20 mm/d. Nonetheless, it is worth noting that GPCP density rapidly collapses below 4 mm/d, whereas both COBRA and ERA5 still show a relevant contribution in this range.

In the tropics, the behaviour is different. Looking at the ocean distributions, ERA5 estimates the highest volume, followed by COBRA and then by GPCP. The three density functions show very different shapes, with GPCP peaking at 10 mm/d, and the other two peaking at about 2 mm/d. COBRA density appears to be smoother, with a relevant contribution coming from a wide range of rates. In contrast, ERA5 exhibits a narrower density with a prominent peak at 2 mm/d. Also, its volume is greater; this is likely due to the overestimates showed by ERA5 in the central Pacific (see section 2.3.2.3). Over land, in contrast, GPCP and COBRA show similar patterns, while ERA5 seems to estimate more moderate regimes (below 10 mm/d) than high regimes (above 10 mm/d). Nevertheless, the three volumes appear to be similar.

Figure 17: Precipitation distribution (via accumulation densities, bars, and overall volumes, hatched bars) of COBRA, GPCP and ERA5 over the period 2000–2017. Histograms are categorized by latitude band (rows) and surface type (columns).

Figure 18: ME (mm/d, first column), RMSE (mm/d, second column) and CC (third column) between the products COBRA, GPCP and ERA5 categorized by sub-periods (ordinate) and surface type (abscissa). From top to bottom: COBRA referred to GPCP, COBRA referred to ERA5 and GPCP referred to ERA5. The different shades of blue give a visible impression of the respective number (light blue = low values, dark blue = high numbers).

The evolution of the ME, RMSE and CC between COBRA, ERA5 and GPCP over time is shown in figure 18, where the scores are also categorized by surface type (ocean, land or both). The same statistical indices are also computed over the entire time period but for distinct latitude bands and can be found in figure 19.

The analysis shows that a significant reduction of the land biases between COBRA and GPCP, as well as between COBRA and ERA5, occurs over time. Indeed, the highest absolute values of ME and RMSE occur in the 2000–2005 period. By contrast, in 2012–2017 the same scores are about 25% lower. Also the CC increases over time.

Over ocean the behaviour is less clear. The ME against GPCP remains constant and that referred to ERA5 slightly varies. Nonetheless, both the two RMSEs show moderate reductions over time.

In the tropics, over the full time period (figure 19), COBRA compares better to GPCP and ERA5 than these two against each other in terms of ME and RMSE. The same behavior is observed in the CC values: higher correlations for COBRA and ERA5, followed by those of COBRA and GPCP and, finally, those of ERA5 and GPCP.

Outside the tropics, however, the COBRA-ERA5 and COBRA-GPCP MEs are larger than the GPCP-ERA5 analogous. Nonetheless, the COBRA-ERA5 correlations are still higher than those between ERA5 and GPCP (and also of COBRA and GPCP). The same relation is observed in the RMSE, which are appreciably smaller for COBRA vs. ERA5.

In summary, in the tropics, COBRA produces estimates between those of ERA5 and GPCP. Outside 25° N/S, in contrast, COBRA shows the highest biases towards smaller rates of precipitation, although the RMSE and CC are better for the COBRA-ERA5 comparison than in the COBRA-GPCP or ERA5-GPCP counterparts. The latter implies that these products show a better temporal coherence. In contrast, comparing GPCP and ERA5, the high RMSE and small ME suggest an error compensation rather than a better agreement, as also confirmed by the moderate correlations.

Figure 19: ME (mm/d, first column), RMSE (mm/d, second column) and CC (third column) between the products COBRA, GPCP and ERA5, over the period 2000–2017, categorized by latitude band (ordinate) and surface type (abscissa). From top to bottom: COBRA referred to GPCP, COBRA referred to ERA5 and GPCP referred to ERA5. The different shades of blue give a visible impression of the respective number (light blue = low values, dark blue = high numbers).

2.3.2.6 Comparison with GPCC

The spatially-resolved comparison of COBRA with the station-based data product GPCC is carried out in this section separately, because here the analysis needs to be based on the monthly products instead of the daily ones, due to differences in the start and end times of a day in GPCC (local times) and COBRA (UTC). It should be noted that due to GPCC only being available over land, the respective grid cells in the COBRA dataset will mostly be populated only by precipitation rate estimates by PNPR-CLIM with only marginal influence from the HOAPS v4 database over small islands and to an even smaller extent, if at all, along coastlines.

Figure 20: Top left panel: Map of climatological mean over the entire COBRA time period (2000–2017) of the COBRA monthly dataset. Top right: Same for GPCC monthly. Middle left: Differences of the climatological means (COBRA minus GPCC). Middle right: Absolute values of relative differences ("COBRA minus GPCC, divided by GPCC"). Bottom centre: maximum number of stations available over all evaluated GPCC months. For convenience, grid cells that are not covered by the station- and thus land-based GPCC dataset are also not shown for the COBRA dataset. 

We restrict the analysis to a comparison of the maps of climatological means. These are shown for COBRA and GPCC in the top panels of figure 20. The main features visible in these maps – high precipitation in the tropics and a decline towards higher latitudes – are the same for the two datasets. The middle left panel shows differences between COBRA and GPCC; the middle right panel puts the absolute differences in relation to the GPCC climatological means. The bottom centre panel shows the maximum number of stations in GPCC over all evaluated months (2000–2017). The underlying data are used as a quality filter on the GPCC data below.

The differences of the climatological means on the 1° × 1° grid have a relatively wide maximum range from -7.7 mm/d to +9.4 mm/d. However, with a median of -0.3 mm/d and 80% of the populated grid cells between -1.2 mm/d and 0.6 mm/d, the spread is mostly modest. The median value indicates a general underestimation of GPCC by COBRA, see also section 2.3.2.1. This underestimation occurs over much of the northern hemisphere and in South America, along and north of the Amazon river. Conversely, COBRA overestimates GPCC in much of Africa and Australia, south of the Amazon river in South America, central North America, and high-mountain Asia. Island settings mostly see an underestimation. The differences reach peak values here in both directions, possibly due to the smaller degree of spatial smoothing of single-station data in the GPCC processing in these situations.

The absolute values of relative differences peak mostly over areas characterized by low precipitation and often little to zero coverage in GPCC (Greenland, Sahara, high-mountain Asia, Australia). In these regions, the relative difference often exceeds 100%, due to the small reference values. In contrast, relative differences in regions with usually high precipitation such as sub-Saharan Africa and especially South America and the Southeast Asian islands are much smaller. Apart from the extreme relative differences in Greenland, Arctic and sub-Arctic areas generally stand out with a relatively high relative difference, a circumstance that might be mostly related to the detection and quantification of snowfall in COBRA. The median in this quantity over the globally populated grid cells is 37%; 80% of grid cells see a relative difference between 7% and 91%. When considering only grid cells that had at least one station in one month in the GPCC dataset (see figure 20, bottom centre panel), the median is at 27.6% and 80% of the remaining grid cells see an improved relative difference between 5% and 63%.

2.3.3 Comparison with high-resolution regional datasets

2.3.3.1 MRMS

The regional comparison with MRMS is carried out over the period 2016–2017. For comparison, also the GPCP and ERA5 estimates are taken into account here. The analysis is limited to those 1° × 1° grid cells with daily average RQI greater than 0.8. Note also that, in order to extract the seasonal radar quality information and smooth out the higher temporal frequencies, the daily average RQI has been further smoothed over a moving time window of 30 days.

Figure 21: ME (first row), RMSE (second row) and CC (third row) of COBRA (first column), GPCP (second column) and ERA5 (third column) referred to MRMS over the period 2016–2017. Only pixels with daily average RQI greater than 0.8 have been considered. 

In figure 21 the spatial distributions of ME, RMSE and CC (rows) of COBRA, GPCP and ERA5 (columns) with respect to MRMS are shown. The COBRA ME shows the highest spatial variability, with positive biases in the central US and negative biases along the coastline, whereas the MEs of both ERA5 and GPCP are almost uniformly negative and at lower absolute deviations from the MRMS data. The COBRA RMSE appears to be similar to that of GPCP, with higher peaks in the central US. In contrast, the ERA5 RMSE is lower.

The CC maps show that the best agreement with MRMS is reached by ERA5 (values higher than 0.7 almost everywhere), followed by COBRA (values above 0.6) and finally by GPCP which, in southern California, shows extremely low correlation values.

Some characteristics of the daily differences with MRMS are reported in table 6 (upper block), whereas the actual distributions (through frequencies and cumulative frequencies) are shown in figure 22. The GPCP error distribution has the widest shape, manifesting also in its extreme 5^th and 95^th percentiles. In contrast, COBRA and ERA5 distributions of errors are more concentrated around zero. In particular, 50% of the ERA5 errors are between 0 mm/d and 1 mm/d. For COBRA, this value is 40%, and for GPCP, it is 35%. Both GPCP and COBRA have 25% of their errors between -1 and 0 mm/d, while ERA5 counts less than 20% of instances in this range. Absolute errors above 4 mm/d stem from less than 20% of the entire population in each dataset. Despite the highlighted differences, all the products show similar ME and RMSE (-0.34 mm/d and 5.74 mm/d for COBRA, -0.28 mm/d and 5.83 mm/d for GPCP, -0.33 mm/d and 4.80 mm/d for ERA5). The CC, instead, are slightly different: 0.73 for COBRA, 0.67 for GPCP and 0.77 for ERA5.

Table 6: Error 5^th and 95^th percentiles, ME, RMSE and CC of COBRA, GPCP and ERA5 against MRMS over the period 2016–2017 (upper block) and against NIMROD over the period 2002–2017 (lower block). Only pixels with daily average RQI greater than 0.8 have been considered with MRMS as reference.

Figure 22: Daily Errors distribution of COBRA, GPCP and ERA5, with reference MRMS, over the period 2016–2017. Colored bars and dashed lines denote frequencies (left y-axis) and cumulative frequencies (right y-axis) respectively. Only pixels with daily average RQI greater than 0.8 have been considered.

2.3.3.2 NIMROD

The regional comparison with the ground-based radar dataset NIMROD is carried out over the period 04/2002–12/2017. As above in the analysis against MRMS, the GPCP and ERA5 estimates are considered as well. As there is no radar quality index provided with NIMROD, in contrast to MRMS, no quality filter has been employed here. Likewise, we did not smooth the data in the temporal dimension

Figure 23 shows the ME, RMSE and CC (rows) of COBRA, GPCP and ERA5 (columns) with respect to NIMROD. All three datasets see a bias towards higher precipitation over open water and towards lower precipitation over land, with certain exceptions over the Alps (first row). ERA5 sees the lowest deviations from NIMROD overall, see also table 6 (lower block). In COBRA, the negative bias over land dominates, probably due to it not being tied to rain gauge data as GPCP is. In GPCP, the positive bias over the sea is more prominent. The overall mean error is highest for COBRA (table 6, lower block). For the RMSE and CC, in turn, COBRA compares similarly well to NIMROD as ERA5 does, whereas GPCP data reproduce NIMROD data to only a lesser degree (figure 23 and table 6, lower block). COBRA and ERA5 see highest correlation over Great Britain and Ireland where NIMROD has the longest record. This indicates that over the entire record, including continental areas, the quality of the comparison may also suffer from data availability. In general, the correlation of the three tested datasets with NIMROD is less than with MRMS (table 6). This might be largely due to a quality filter and a temporal smoothing being applied in the analysis against MRMS.

Figure 23: ME (first row), RMSE (second row) and CC (third row) of COBRA (first column), GPCP (second column) and ERA5 (third column) referred to NIMROD over the period 2002–2017. Only grid cells with more than 1000 days of mutual coverage are displayed. In contrast, the overall statistics in table 6 and the graphs in figure 24 have all data included. 

Distributions of differences are shown in figure 24. Similar to the comparison with MRMS above, the GPCP error distribution has the widest range, also visible in the percentiles given in table 6. COBRA sees the highest percentage of error instances in the ±1 mm/d range compared to GPCP and ERA5.

Figure 24: Daily Errors distribution of COBRA, GPCP and ERA5, with reference NIMROD, over the period 2002–2017. Colored bars and dashed lines denote frequencies (left y-axis) and cumulative frequencies (right y-axis) respectively.

2.3.4 Internal validation of monthly mean precipitation

Here we compare the COBRA monthly dataset and monthly means of the COBRA daily dataset (“monthly COBRA daily dataset”) in order to assess deviations that arise from the different processing. The latter data are computed specifically for this exercise. As a reminder, the COBRA daily dataset has been produced by accumulating hourly precipitation averaged over all platforms; gaps in the hourly database have been filled through nearest-neighbour interpolation in the temporal dimension. The COBRA monthly dataset is not simply based on monthly means (or sums) of the daily values, but as averages over the initial instantaneous precipitation-rate estimates, first platform-by-platform, and then one final average over the available platforms. See the ATBD [D2] and PQAD [D1] for details.

During times when NOAA15 has not been used (04/2001 onwards), the monthly global mean values in the COBRA monthly dataset deviate from the respective global mean values in the monthly COBRA daily dataset between -0.01 mm/d and +0.05 mm/d, with a median deviation of 0.01 mm/d, i.e., overall slightly higher values in the COBRA monthly dataset. With COBRA global means in the order of 2.6 mm/d (see section 2.3.2.1, figure 13), this implies a maximum relative deviation of ~1.9%, with 86% of all monthly instances below 1% relative deviation. The deviations during times when NOAA15 contributes are slightly larger, amounting to negative peak deviation of -0.06 mm/d or 2.7%.

There are spatial patterns visible in the geographically resolved differences of climatological means of COBRA monthly and monthly COBRA daily (figure 25; see section 2.3.2.2 for a general assessment of the spatially resolved COBRA climatological mean). These differences vary between ±0.6 mm/d. At low latitudes, COBRA monthly generally sees higher values than monthly COBRA daily, which is mostly reversed at mid latitudes. At high latitudes, especially in the Southern Ocean, COBRA monthly again exceeds monthly COBRA daily. The relative differences are in general small. Over all available 1° × 1° grid cells, the median relative absolute difference is 2.2%; 90% of the grid cells see differences of less than 10.2%. The maximum relative differences (referenced to the COBRA monthly climatology) occur in Greenland, peaking at ~123%, and less so in Antarctica, including sea-ice areas in the Southern Ocean. In all these areas, the precipitation rates have to be used with caution in general, due to the involved frozen background surfaces (sea ice and snow cover), see also the discussion in section 2.1.2. The discussed differences occur mostly because of the reverse order of temporal and inter-platform averaging, which leads to different weighting of the platforms. Also, the hourly gap filling in the production of the COBRA daily dataset weighs certain observations higher than others, depending on the gap to the next satellite overpass, whereas observations in the monthly processing are treated equal.

It is assumed that the discussed differences are small compared to the precision of the data. For example, the climatological comparison with GPCP and ERA5 (section 2.3.2.2, figure 14) sees differences higher by at least a factor 5. However, as both versions of monthly data could be used for climatological applications, such as the analysis of water cycle components, we have documented these differences here so that users are aware of inherent deviations.

Figure 25: The left panel shows differences in the maps of climatological means over the entire COBRA time period (2000–2017) of the COBRA monthly and the monthly COBRA daily datasets. In areas of positive differences (red), monthly COBRA daily exceeds COBRA monthly, and vice versa for negative differences (blue). The right panel shows the absolute differences relative to the COBRA monthly climatological map.

3 Application(s) specific assessments

n/a

4 Compliance with user requirements

Section 2.3.2.1 and the PUGS [D3] discuss the compliance with the formulated requirements for accuracy and stability of global mean precipitation rates, see also the KPI document [D5] and the respective TRGAD [D6]. As mentioned in the PUGS [D3], the data originating from NOAA15 platform show some issues affecting the datasets accuracy. When including times where low-quality data by NOAA15 produce a prominent negative bias (2000-01-01 to 2001-03-31), the comparison with GPCP v2.3 monthly and v1.3 daily, respectively, shows compliance with the accuracy target of 0.3 mm/d in 91.7% of all months and 84.7% of all days (table 5).

The percentage of COBRA data fitting the accuracy target rise up to 98.5% for monthly and 87.5% for daily values if COBRA is been evaluated after the NOAA15 platform is excluded (see above for the timing) with GPCP as the reference.

In the case of the daily comparison, a more relaxed target would be required, to allow for sub-monthly temporal variability in the datasets, see the PUGS [D3].

The stability target of 0.034 mm/d/dec is met by both the monthly and the daily resolved datasets with GPCP v2.3 monthly / v1.3 daily as reference datasets. The evaluation of COBRA global mean values against ERA5 shows a more negative bias (accuracy), which has been discussed extensively in sections 2.1.2 and 2.3.2.

On the scale of single grid cells, COBRA sees a negative bias (mean error) of about 0.1 mm/d with respect to GPCP (section 2.3.2.5). The respective spread (RMSE) is, of course, much larger than in the comparison of global mean values above (the latter staying mostly below 0.3 mm/d).

Other requirements have clearly been met by the dataset: COBRA is based on stable, high-quality FCDRs. It is available over a time span of 18 years, allowing an analysis on near-climatic time scales. And finally, the spatial coverage extends to the full globe, albeit ice-covered areas are problematic, see for example section 2.1.2.

5 Appendix

5.1 Additional time series of spatial averages

Figure 26: Upper panel: Time series of daily global mean precipitation in COBRA, GPCP v1.3 (daily), and ERA5. Lower panel: Time series of differences (solid lines) of daily global mean precipitation, i.e., COBRA minus the respective reference datasets GPCP and ERA5. GPCP data have been filtered for anomalous, invalid values > 3.3 mm/d or < 2.2 mm/d) prior to the difference calculations. The respective trends, starting on 2001-04-01, see main text, are displayed as dashed lines. The black dash-dotted lines mark the ±0.3 mm/d accuracy limits.

Figure 27: Time series of monthly global mean precipitation (upper panel) and monthly global mean differences between COBRA data and GPCC (lower panel). GPCC is only available over Land, so COBRA data have been filtered for land covering grid cells, hence the depicted time series is not the same as in figure 13. For the time series of differences a trend is displayed as dashed. That trend has been computed using a limited time series (data before 2001-04-01 are discarded). The black dash-dotted lines mark the ±0.3 mm/d accuracy limits.

5.2 Temporal means and differences over three time periods

Figure 28: Mean precipitation of COBRA (first column), GPCP (second column) and ERA (third columns) over three periods: 2000–2005 (first row), 2006–2011 (second row) and 2012–2017 (third row). 

Figure 29: Mean differences between COBRA and GPCP (first column), COBRA and ERA5 (second column) and GPCP and ERA5 (third column) over three periods: 2000–2005 (first row), 2006–2011 (second row) and 2012–2017 (third row).

5.3 Error decompositions over three time periods

Figure 30: Error components (from top to bottom: hit error, missed precipitation and false precipitation) between the various daily gridded precipitation datasets (from left to right: COBRA referred to GPCP, COBRA referred to ERA5 and ERA5 referred to GPCP) over the period 2000–2005. A detection threshold of 1 mm/d has been applied. 

Figure 31: Error components (from top to bottom: hit error, missed precipitation and false precipitation) between the various daily gridded precipitation datasets (from left to right: COBRA referred to GPCP, COBRA referred to ERA5 and ERA5 referred to GPCP) over the period 2006–2011. A detection threshold of 1 mm/d has been applied.

Figure 32: Error components (from top to bottom: hit error, missed precipitation and false precipitation) between the various daily gridded precipitation datasets (from left to right: COBRA referred to GPCP, COBRA referred to ERA5 and ERA5 referred to GPCP) over the period 2012–2017. A detection threshold of 1 mm/d has been applied.

5.4 Seasonal zonal means

Figure 33: COBRA, GPCP and ERA5 seasonal zonal means for September-October-November (first row) December-January-February (second row), March-April-May (third row) and June-July-August (fourth row) over the globe (first column), the ocean (second column) and the land (third column). The means are computed over the full time period 2000–2017.

References

Adler, R., et al. (2016). Global Precipitation Climatology Project (GPCP) Climate Data Record (CDR), Version 2.3 (Monthly) [Data set]. National Centers for Environmental Information. https://doi.org/10.7289/V56971M6

Adler, R., et al. (2017). Global Precipitation Climatology Project (GPCP) Climate Data Record (CDR), Version 1.3 (Daily) [Data set]. NOAA National Centers for Environmental Information. https://doi.org/10.7289/V5RX998Z

Adler, R. F., et al. (2018). The Global Precipitation Climatology Project (GPCP) Monthly Analysis (New Version 2.3) and a Review of 2017 Global Precipitation. Atmosphere 9(4), 138. https://doi.org/10.3390/atmos9040138

Bagaglini, L., et al. (2021). The Passive Microwave Neural Network Precipitation Retrieval Algorithm for Climate Applications (PNPR-CLIM): Design and Verification. Remote Sens. 13, 1701. https://doi.org/10.3390/rs13091701

Copernicus Climate Change Service (C3S) (2017). ERA5: Fifth generation of ECMWF atmospheric reanalyses of the global climate [Data set]. Copernicus Climate Change Service Climate Data Store (CDS), 21/12/2018. https://cds.climate.copernicus.eu/cdsapp#!/home

Kirstetter, P.-E., et al. (2012). Toward a Framework for Systematic Error Modeling of Spaceborne Precipitation Radar with NOAA/NSSL Ground Radar–Based National Mosaic QPE. J. of Hydrometerol., 13(4), 1285–1300. https://doi.org/10.1175/JHM-D-11-0139.1

Huffman, G. J., et al. (2001). Global Precipitation at One-Degree Daily Resolution from Multi-Satellite Observations. J. Hydrometeor., 2(1), 36–50.

Met Office (2003). Met Office Rain Radar Data from the NIMROD System. NCAS British Atmospheric Data Centre. http://catalogue.ceda.ac.uk/uuid/82adec1f896af6169112d09cc1174499

Panegrossi, G., et al. (2021). Recent Advances and Challenges in Snowfall detection and Estimation. In S. Michaelides (Ed.), Precipitation Science. Elsevier, in press.

Schneider, U., et al. (2020). GPCC Monitoring Product: Near Real-Time Monthly Land-Surface Precipitation from Rain-Gauges based on SYNOP and CLIMAT data [Data set]. Global Precipitation Climatology Centre (GPCC, http://gpcc.dwd.de/) at Deutscher Wetterdienst. https://doi.org/10.5676/DWD_GPCC/MP_M_V2020_10

Turk, J., et al (2021). Adapting Passive Microwave-Based Precipitation Algorithms to Variable Microwave Land Surface Emissivity to Improve Precipitation Estimation from the GPM Constellation. J. of Hydrometeorol., 22(7), 1755–1781. https://doi.org/10.1175/JHM-D-20-0296.1

Zhang, J., et al. (2016). Multi-Radar Multi-Sensor (MRMS) quantitative precipitation estimation: initial operating capabilities. Bulletin of the American Meteorological Society, 97, 621–638. https://doi.org/10.1175/BAMS-D-14-00174.1.

Related articles

• Page:
  ERB TSI: Product User Guide and Specification (PUGS)
• Page:
  ERB TSI: Product Quality Assurance Document (PQAD)
• Page:
  ERB TSI: Product Quality Assessment Report (PQAR)
• Page:
  ERB TSI: System Quality Assurance Document (SQAD)
• Page:
  ERB TSI: Algorithm Theoretical Basis Document (ATBD)