1. Introduction
With the introduction of a precipitation radar on board the Tropical Rainfall Measuring Mission (TRMM) satellite, meteorological radar applications could be expanded to the oceans and lands between ±35° latitude on Earth. The Global Precipitation Measurement (GPM) mission
Core Observatory satellite, launched on 27 February 2014 [
1], has extended the TRMM coverage to ±65° latitude, hence including the Alps. Because of the roughly 5 km diameter footprint of the GPM dual-frequency precipitation radar (DPR), clutter contamination will typically limit GPM measurements to areas above the mountains, except where very broad valleys are present. The DPR adds significant new coverage to orographically complex areas, especially in poorly instrumented regions of the world.
Precipitation in orographically complex regions like the Alps is inherently complex because of the interactions between the mountainous terrain and the atmosphere. The resultant effects have been the subject of extensive research (e.g., [
2]). Additionally, the higher altitudes (and hence lower temperatures) mean that a large proportion of precipitation falls in solid phase, resulting in significant amounts of water being stored in the Alps as snow. This makes the mountains a vital water source in many parts of the world. Better knowledge of precipitation fields in the Alps is necessary in weather, nowcasting, and climate research, and more accurate precipitation measurements, especially in orographically complex catchments, are essential to develop reliable warnings. However, the same complex orography and colder temperatures that make the Alpine environment an important water reservoir in Europe, also make the measurement of precipitation challenging not only for spaceborne radars, but also for ground-based weather radars. The rough and complex orography together with cold and windy conditions make the installation of large radar antennas difficult [
3]; the terrain may cause partial or almost total shielding of the radar beam, restricting measurements to high altitudes (e.g., [
4,
5,
6]). Unfortunately, the well-known sources of error that affect radar estimations (e.g., [
7,
8,
9,
10]) become particularly severe in mountainous terrain (e.g., [
11,
12]).
Rain gauges are usually considered to provide accurate “point” measurements. In this context, “point” means a gauge cross section that is usually 200 cm
2 (and a sampling volume of the order of ~60 m
3 every 10 min, assuming, for instance, an average terminal fall velocity of 4 m/s). Consequently, it is well known that gauge observations represent local effects and not areal quantities. For hydrological applications, areal precipitation estimates are required, and, therefore, their main drawback is undersampling [
13], i.e., there will not be enough observations to describe the variability of the field. Because of the spatial variability of rainfall (especially during intense, convective events), it is obvious that point measurements would undersample the precipitation fields, even though the measurements themselves were correct. Radar measurements can add the desired information on the areal distribution of precipitation. There are enormous differences between remotely sensed radar observations and in situ point measurements by the rain gauges. On the one hand, the small-scale variability of precipitation is such that even three decades ago Austin [
14] wrote that “the spatial variability of rainfall is so pronounced that a single gauge often does not sample representatively over any area, even as small as 4 km
2”. On the other hand, spatial smoothing by the main Gaussian lobe of the radar antenna radiation pattern, as well as post-detection integration over many pulses reduce the variability of the distribution of precipitation rate in space. Even in the liquid phase, when radar data are compared with “instantaneous” point rainfall rate, a random error and a bias are introduced by the smoothing [
15]. This explains some of the difficulties in the use of radars for quantitative precipitation estimations. When gauge data are also smoothed in time, the random error and the bias are reduced.
Various deterministic or statistical methods for radar–gauge merging have been proposed. Many applications of these tools have appeared in the literature, all of which provide valuable contributions to the common knowledge regarding the problem at hand (e.g., [
16,
17,
18,
19,
20,
21,
22]). In Switzerland, the merging of radar and rain gauge measurements is based on spatio-temporal kriging with external drift [
22]; it was implemented at MeteoSwiss five years ago and it has been run operationally since then. This product is called CombiPrecip.
To summarize, in the Swiss Alps there are currently three different ground-measured precipitation products available: (1) a product based on a network of automatic telemetered rain gauges (GAUGE); (2) a product based on a state-of-the art, C-band, dual-polarization Doppler, ground-based weather radar network (RADAR); and (3) a product consisting in a real-time merging of observations from these two networks, implemented and operational at MeteoSwiss (CombiPrecip). In this paper, we compare these products with another product based on the GPM spaceborne radar, which should provide the best space-based instantanous rainfall rate estimations.
Several authors (e.g., [
13,
15,
23]) illustrate the problems of the two types of instruments (radars and gauges) concerning the large scatter of “short time-scale” values for a given location, which are caused by the huge differences in the sampling modes. In many cases, and certainly for our evaluation purposes, it is wise to integrate the measurements over time. An integration over long times also reduces the differences in the retrieved estimates between the spaceborne and the ground-based radars. Although these two sensors certainly have more similarities than a radar and a rain gauge, there are nevertheless enormous differences between the two. In this regard, mention can be made of the different operating frequencies, sampling volumes, geometrical viewing angles, attenuation, sensitivity, and temporal resolution (see
Section 2.1 and
Section 2.2).
Consequently, in this short note aimed at evaluating the performance of three Swiss precipitation products (RADAR, GAUGE, and their combination, i.e., CombiPrecip) in comparison with GPM, a 2 year integration period is used to minimize the large differences in the sampling modes between the various instruments. During this period, the same number of rain gauges (221) and ground-based radars (four) were used. Obviously, only RADAR, GAUGE, and CombiPrecip observations that were closest in time to GPM overpasses are used to derive the total precipitation amounts in every 25 km
2 reference pixel (see
Section 2.1 and
Section 2.2.3 for more details). The GPM precipitation product underestimates, on average, the precipitation in Switzerland by a factor of 2 (−3 dB) with respect to RADAR, and by a factor of −3.4 dB with respect to GAUGE. However, GPM estimates are used to independently evaluate the three Swiss products in terms of dispersion of the error around the mean (as well as of correlation among the various precipitation products), rather than in terms of exact average amount of 2 year precipitation over the country (mean error). To the best of the authors’ knowledge, this is the first attempt to compare in complex terrain four precipitation products that use three very different types of sensors. Using GPM as a reference, the best precipitation product results to be the combination, through a geostatistical approach, of ground-based radar and rain-gauge measurements.
This fact is thoroughly described in
Section 3, which presents the results. The findings and their implications are discussed
Section 4, which presents also future research directions.
Section 2 gives an overview of the instruments and precipitation products used, and describes the study area and methods. A conclusion is given in
Section 5.
4. Discussion
In this Section, we further discuss the quantitative agreement among the four precipitation products in the 221 sites where gauge measurements are available; then, we analyze the qualitative agreement among GPM, RADAR, and CombiPrecip in the described study area (360 × 240 km2), with special focus on the Swiss territory.
We first complete the visualization of the scatterplots associated with the correlation coefficient values that are shown in
Table 1 and discussed in
Section 3. In this respect,
Figure 2 shows the Log-transformed annual precipitation estimates and
Figure 3 shows the original linear values. Please note that, with respect to
Figure 2, we have changed the independent variable point of view, and thus, in
Figure 3, the gauge amounts are reported on the abscissa, while GPM (augmented by 2 dB), RADAR, and CombiPrecip are shown on the ordinate. We have opted for GAUGE values on the abscissa to check the reliability and representativeness of the values corresponding to the large gauge amounts collected every 10 min that were acquired during the first 2 years of activity of the GPM Core Observatory satellite. As stated in
Section 2, these 528 samples are synchronous with the GPM overpasses over the Swiss territory and represent a “random” subsample of 0.5% of all available samples. Using all samples, the rainiest site (ROE, 1896 m altitude) shows ~2500 mm/a, while the second and third rainiest sites (SAE, 2502 m; GSB, 2472 m) show ~2350 mm/a and ~2150 mm/a, respectively. As can be easily seen in
Figure 1 and
Figure 2, using the GPM-synchronous 528 samples, the rainiest site is found to be GSB (~3220 mm/a) followed by Saentis (SAE, ~2770 mm/a).
Figure 2 also shows that while the RADAR and CPC underestimation at the SAE site are still “acceptable”, at the GSB site they are not. All three products suggest that the rain gauge is not reliable at that site. This fact is well known at MeteoSwiss: for instance, the correlation coefficient between 6589 wet–wet RADAR–GAUGE hourly amounts during the period 2005–2009 is as low as 0.133. Additional investigations have suggested that this rain gauge site is far from being ideal. At the end of 2016, a new rain gauge was installed in a better position (~1 km away from the old site). Preliminary comparisons between the radar and the two gauge estimates seem to confirm that the site of the new gauge is much better. Thanks to GPM (see
Figure 2), MeteoSwiss has an additional confirmation of the problem at GSB. It is also worth noting that the correlation and scatter values presented in
Section 3 are not sensitive to the GSB data pair; rather, they are reasonably resistant. In fact, by eliminating GSB, the correlation coefficient between GAUGE and GPM increases from 0.403 to 0.411 (Log values) and from 0.290 to 0.312, and the scatter decreases from 2.84 dB to 2.78 dB using GAUGE as a reference. As expected, the elimination of the GSB site does not influence the agreement of GPM with RADAR or with CombiPrecip.
The performance of the various precipitation products can also be evaluated using additional scores, like the Root Mean Square Error (RMSE), the Heidke skill score (HSS), and the Nash-Sutcliffe (NS) efficiency.
Table 3 shows the values of the root mean square difference (“error”) between GPM and the three MeteoSwiss precipitation products. The results are identical to those obtained with the correlation coefficient (
Table 1) and the scatter (
Table 2): the best agreement is observed with CombiPrecip, followed by the ground-based radar, and finally the gauge network. Note that the RMSE can be decomposed in two orthogonal contributions: the square of the mean difference and the variance of the differences. Original GPM values give a significant mean error component because of the large underestimation. Therefore, in the second row of
Table 3 we also show the values after a posterior adjustment of +2 dB, which is the same factor used in
Figure 1 and
Figure 2. However, due to the relatively poor correlations (see
Table 1, line 2) between GPM and the three products, this causes not only a reduction of the squared mean difference, but also an increase of the variance of the difference. Nevertheless, a compensation of 2 dB still gives a better RMSE than the one obtained with the original GPM estimates, as can be seen by comparing the values in the two lines of
Table 3 and
Table 4. An adjustment of 3 dB brings the square of the mean difference to zero, but causes an enormous increase of the variance of the difference, so that the RMSE becomes even larger than the one obtained with the original (underestimated) GPM estimates.
Similar results can be seen in
Table 4, with a slight improvement when using the new version of the GPM-DPR level-2 data, namely, V05A.
Table 5 shows the values of the Heidke Skill Score in the case of a polychoric metrics; in particular, we have opted for a 4 × 4 contingency table, with the GPM estimates augmented by 2 dB along the columns, and (alternatively) the three MeteoSwiss products along the rows. The three thresholds used to define the four categories are: 650, 850, and 1250 mm/a. These are exactly the same values used in
Figures S2 (GPM estimates), S3 (RADAR estimates), and S4 (CombiPrecip estimates). The HSS values confirm the results obtained using correlation, scatter, MADM, and RMSE: the best agreement is observed between GPM and CombiPrecip, followed by RADAR and GAUGE. In all cases, the GPM version V05A performs better than the V04A.
The definition of the NS efficiency is based on a regression through the origin, instead of an ordinary least-squares regression (like the one used to derive the correlation coefficients shown in
Table 1). The ratio between the RMSE and the standard deviation of the reference is the crucial quantity in the definition of the NS efficiency, which is one minus such a ratio. A performance evaluation based on the NS efficiency confirms what was evidenced by the other scores: the best agreement is between GPM and CombiPrecip. Again, results are slightly better when data from the GPM version V05A are used instead of those from the V04A version. In addition, all the five scores used provide the same ranking picture: the estimates of accumulated precipitation from space during the first 2 years of GPM life in 221 Swiss sites agree better with those of the CombiPrecip product, followed by the those of RADAR and then, as expected, those of GAUGE. The extensive statistical analysis presented so far also shows that the GPM version V05A agrees better than the V04A version with the three MeteoSwiss precipitation products.
So far the quantitative analysis was based on 221 total amounts and 4 precipitation products. Three products, however, carry continuous spatial information regarding the precipitation fields in all the 3456 pixels (25 km
2) of the study area. The corresponding images for Log-transformed and linear total precipitation amounts are shown in
Figures S3 and S4, respectively: 528 overpasses are already enough to highlight several climatological features of the area surrounding the northern part of the Western Alps (from Monte Bianco to Ortler and Piz Murnin) and the Jura mountains. By comparison with
Figure S1, it can clearly be seen that orography strongly affects precipitation in the study area. The dominant weather pattern across much of Switzerland is characterized by a westerly-northwesterly wind bringing precipitation that falls primarily on the northwestern side of the Alps. Another dominant pattern on the southern side of the Alps (Piemonte, Canton Ticino, Lombardia) is also clearly visible, although it appears more “convective” (there are several high precipitation patterns composed of just few pixels, corresponding to areas that are smaller than 100 km
2). This southern part of the Alps tends to get its precipitation from a different weather pattern (characterized by a southerly wind from the Mediterranean), resulting in high amounts of precipitation on the western or central Prealps or Alps, depending on the direction (and strength) of the winds in the low (typically south–easterly) and upper (typically south–westerly) levels [
46]. In the southwestern Alps and in the surrounding flatlands and hills, autumn is the season in which the heaviest daily rainfalls occur [
47]. This fact has long been known: the description of these “late-summer” Mediterranean storms can be found even in "Naturalis Historia" by the Roman author Plinius. There is then another area of higher precipitation in the smaller Jura Mountains in the northwest of Switzerland, which form the first significant mountain range encountered by the westerly fronts, hence generating a strong orographic enhancement of precipitation.
It is interesting to note (
Table 6) the values of the Coefficient of Variation (CoV, relative dispersion) that characterizes the probability distribution functions of the linear precipitation amounts (3456 pixels) over the study area (86,400 km
2). It is observed that GPM presents the largest value, followed by RADAR, while CombiPrecip presents the smallest value. This is also the case when dividing the study area in two (roughly equal) parts, inside and outside Switzerland (lines 2 and 3), and it is caused by an increased underestimation tendency of GPM with increasing orographic complexity. Consequently, the difference in the relative dispersion between GPM and RADAR (or CombiPrecip) is larger inside Switzerland, where there are more mountains. To further explore the effect of an increasing orographic complexity, we divided the data set in two parts, i.e., complex and flat terrain, using the same strategy and threshold as in [
25] (Section 2.d, p. 1253). It can be observed that the CoV characterizing 135 sites in complex terrain is larger than the CoV in non-complex terrain. The stratification upon altitude also shows the quite different level of underestimation between GPM and RADAR or CombiPrecip. Using 135 sites (pixels) in complex terrain, in fact, the Mean Field Bias is −3.8 dB with respect to the RADAR, and −3.5 dB with respect to CombiPrecip. Using 86 sites in non-complex terrain, the underestimation is −0.9 dB with respect to both products. This confirms findings by Speirs et al. [
25] using RADAR as reference: both detection and estimation performance of GPM over Switzerland are significantly better in flatter than in complex terrain.
As a result of 2 years of GPM observations, a new, completely independent QPE product is now available for MeteoSwiss. In this note, GPM total amounts have been compared in detail with those of the other MeteoSwiss QPE products. Unfortunately, the GPM estimates are less reliable just in those high altitude areas where MeteoSwiss estimates are also more uncertain or absent (GAUGE). Consider, for instance, that there are only three MeteoSwiss gauges (~1.4% of the total) at an altitude over 2500 m (the altitude of the above mentiones SAE and GSB sites), while more than 10% of the Swiss territory is at an altitude over 2500 m. There is a clear need to improve the knowledge of the complex precipitation fields at such high altitudes. To conclude, a key point is whether 528 GPM overpasses (a few hundred of them with some wet pixels in Switzerland) are sufficient to derive an accumulation product able to discriminate which MeteoSwiss precipitation product shows the more realistic spatial variability of precipitation. Thanks to the experience gained in mountainous regions in the southeastern Mediterranean [
43,
44] and western Alps [
25], we are confident that this is the case. However, it will be necessary to extend the analysis for a few more years. To this end, it is planned, in fact, to analyze data from 2016, hence fully exploiting the observations from the recently installed ground-based radar in the Eastern part of the country. Further extension to 2017 and subsequent years will certainly make the data set more robust and representative, especially in those areas where the contribution to the generation of precipitation was caused more by convective instability than by baroclinic instability. Consider, for example, that eight years of spaceborne radar were used in cases where the TRMM-PR level 2 data alone were used to characterize precipitating convective systems in the southeastern Asia [
48], as well as their regional, seasonal, and diurnal variations [
49].
Just as the TRMM Precipitation Radar was used in the last 20 years, the GPM-DPR is primarily intended to be used as the calibration basis for radiometer-derived global precipitation products [
1]. The GPM core satellite, in fact, carries on board not only the active DPR, but also the passive microwave radiometer called GMI [
50]. As such, it is important that the DPR provides a reliable reference in as many circumstances as possible. The work presented here is focused on the performance of the DPR for quantitative precipitation estimations in complex terrain (the western Alps). DPR performance is found to be remarkably better in flatter terrain (the Swiss plateau) than in complex terrain. In turn, this may affect passive radiometer estimates, which are known to be more problematic over complex terrain than in flat areas (e.g., [
51]). Furthermore, it has been recently shown [
52] that radiometers estimates over mountainous terrain could be improved through a simple parametrization that accounts for orographic enhancement. Hence, our future analysis will be carried out also as a function of the complexity of orography (complex vs. flatter terrain).
5. Conclusions
Switzerland is a unique and ideal ground validation site for comparing spaceborne weather radar level 2 precipitation estimates (GPM,
Section 2.1) with other different precipitation products, as the mountainous and relatively small (~41,000 km
2) country is equipped with a network of telemetered, automatic rain gauges (GAUGE,
Section 2.3), and with a state-of-the-art dual-polarization network of weather radars (RADAR,
Section 2.2). A third and innovative MeteoSwiss precipitation product is available (CombiPrecip,
Section 2.4), which is a real-time geostatistical merging of RADAR and GAUGE. To the best of the authors’ knowledge, there is no existing published work examining the GPM-DPR precipitation product and three other different precipitation products in mountainous terrain. On average, GPM underestimates precipitation amounts by −3.0 dB with respect to RADAR, and by −3.4 dB with respect to GAUGE.
Ground-based and spaceborne radars provide a complementary view of the complex and highly variable precipitation fields: the former measures precipitations from a lateral direction, while the latter looks at the precipitation fields from the top. The radar backscattering volume increases with the square of the range. Consequently, the changes in resolution of the GPM radar are almost negligible, since its measurements originate from similar distances of 400–420 km over the whole swath. On the contrary, the sampling volume of the ground-based radar changes significantly, since it has to measure precipitation from close to large distances. Hence, the GPM can be used as a reference to evaluate the proper spatial variability of both the RADAR and the CombiPrecip products, since the former is used as an external drift to derive the latter in the geostatistical merging based on GAUGE kriging. This explains why, in this note, not only an evaluation is made of the GPM level 2 estimated surface precipitation product against syncrhonous estimates from the MeteoSwiss RADAR, GAUGE ,and Combiprecip products, but GPM is also used as a reference to evaluate the dispersion of the multiplicative difference between each MeteoSwiss products and an unbiased version of the GPM-derived field.
The evaluation of the GPM level 2 precipitation product in the Swiss complex terrain is of interest for those who plan to use GPM in other mountainous areas. The GPM core satellite is, in fact, adding significant new coverage to mountainous areas, especially in poorly instrumented parts of the world and at latitudes not previously covered by TRMM. This study shows that, in terms of average precipitation, a clear underestimation by GPM (with respect to the dense ground-based radar or the gauge networks) can be expected also in other mountainous areas of the world. Regarding the dispersion of the GPM error around the mean, even by integrating 2 years of overpasses, the overall scatter associated with the high-resolution DPR cells (20 km2) is of the order of 2.6 dB. This figure represents an overall post facto assessment, and it is simply a quantitative assessment of the scatter of the differences between the DPR and the other three ground-based references, namely, the rain gauge network (2.8 dB), the weather radar network (2.6 dB), and a geostatistical combination (CombiPrecip) of these two ground-based networks (2.4 dB).
Regarding the use of GPM for evaluating the performance of the three MeteoSwiss products, the smallest dispersion is observed between GPM and CombiPrecip (scatter ~1.8 dB), followed by GPM and RADAR (scatter ~2.2 dB), whereas a larger disagreement was found between GPM and GAUGE (scatter ~2.4 dB). GPM confirms that, for precipitation mapping in the Alpine region, the best results are obtained by combining ground-based radar and rain-gauge measurements using a geostatistical approach.