Precipitation plays a critical role in the global hydrological cycle and energy exchange of the atmosphere [1
]. The uneven distribution of precipitation over large areas may directly cause the hydrologic system to be unstable [5
] and easily lead to urban flooding, mountain mudslides, and flash floods [8
]. Thus, observing the precipitation with high accuracy on a global scale is helpful in flood early warning, water resources management, and weather and geological disasters monitoring [11
]. However, many more challenges remain in obtaining accurate precipitation over large areas. Currently, rain gauges, ground-based Doppler weather radars, and spaceborne active and passive remote sensors are the mainstream measurement of precipitation [17
]. Although rain gauges can provide the most accurate precipitation at certain locations, its low-density distribution limits the high spatial resolution precipitation observation, especially in the complex terrain regions [18
]. The ground-based Doppler weather radars can provide real-time, accurate, and refined internal structure of precipitation [20
], while its actual detection performance is often affected by factors such as the beam blockage, the electromagnetic wave attenuation, and other measurement errors [21
]. Some of these influencing factors can be resolved by using networked radars, but it is unsuitable for application in the mountainous and oceanic regions [24
]. Spaceborne remote sensors may be the most reliable and stable way to measure the global precipitation due to their extensive spatial coverage and continuous monitoring [3
The spaceborne remote sensors include both passive and active types for their different working principles. The passive spaceborne remote sensors mostly use visible light, infrared, or microwaves, which are difficult to obtain the accurate internal structure of precipitation [28
]. The most effective means to obtain the details of precipitation in large areas rely on the active spaceborne microwave precipitation radar [13
]. The first Ku-band active microwave precipitation radar and microwave imager (TMI, 10–85.5 GHz), which were boarded on the Tropical Rainfall Measuring Mission (TRMM) satellites, were successfully launched by the National Aeronautics and Space Administration (NASA) and National Space Development Agency (NASDA) in 1997 [6
]. Since then, various satellite precipitation products with high spatiotemporal resolution have been made available, such as NASA’s TRMM Multi-satellite Precipitation Analysis (TMPA), NOAA’s Climate Prediction Center MORPHing technique (CMORPH) [32
], the Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks (PERSIANN), and Global Satellite Mapping of Precipitation (GSMaP) [33
]. These products can greatly improve the accuracy of heavy precipitation early warning, and promote the study of the relationship between global precipitation and climatic change [34
However, due to its limitations in detection performance [37
], the TRMM cannot detect the solid precipitation and light precipitation less than 0.5 mm/h, both of which bear important implications for accurate evaluation of the global energy and water resources distribution [38
]. To solve this issue, the Global Precipitation Measurement (GPM) mission, an international constellation of satellites carrying more advanced spaceborne meteorological observation instruments, was developed by NASA and the Japan Aerospace Exploration Agency (JAXA) and launched in 2014. The GPM core observatory carries, in particular, a dual-frequency precipitation radar (Ku-band and Ka-band, DPR) and a conical-scanning multichannel GPM Microwave Imager (GMI, 10–183 GHz). In addition, the GPM mission provides the Integrated Multi-satellitE Retrievals for GPM (IMERG), which include near-real-time (NRT) “Early” (IMERG-E) and “Late” (IMERG-L), and post-real-time (PRT) “Final” (IMERG-F) products. Compared to the TRMM precipitation products, the GPM products have larger spatial coverage and higher spatiotemporal resolution (0.1° × 0.5 h vs. 0.25° × 3 h) [39
At present, a number of comparison studies focus on the accuracy and reliability of the IMERG-F product and two TMPA products (3B42 and 3B43) [3
]. For instance, Xu et al. (2019) [40
] found that IMERG-F has a better precipitation detection capability than 3B42 at daily resolution in the Huang-Huai-Hai Basin Plain, China (probability of detection are 0.83 and 0.67, respectively), especially when detecting trace and solid precipitation. Tang et al. (2016) [3
] evaluated IMERG-F and 3B42 products at three hourly and daily resolutions over mainland China, and found that IMERG-F shows higher correlation coefficient (CC) than 3B42 (0.53 vs. 0.42 for three hourly products, 0.71 vs. 0.68 for daily products). Lu et al. (2018) [27
] compared the monthly IMERG-F and 3B43 products over Xinjiang, China, and found that their relative bias (RB) are 7.76 and 10.24%, and CC are 0.68 and 0.62, respectively. Thus, these studies demonstrate that the IMERG-F product shows better precipitation detection capabilities than the TRMM precipitation products.
As for the comparative evaluation of the IMERG-E, IMERG-L, and IMERG-F products, there are only a few studies [43
]. For instance, Wang et al. (2017) [43
] evaluated the IMERG daily precipitation products over the Beijiang River, China. Sungmin et al. (2017) [44
] compared the IMERG version 3 Early, Late, and Final precipitation products in southeastern Austria. Tan et al. (2017) [45
] assessed the IMERG precipitation products in Malaysia. However, there is little research on the quantitative assessment and comparisons among the IMERG-E, IMERG-L, and IMERG-F hourly products with gauge-based products that have been proven to be very accurate, especially in the complex terrain regions. Meanwhile, the observation data used in many IMERG product evaluation studies span less than one year, thus the results are not as robust.
It is recognized that evaluation on the accuracy and reliability of the NRT IMERG hourly products is conducive to improvement of data quality. In particular, high quality precipitation data with high spatiotemporal resolution used in weather forecast and hydrological models can optimize the forecast accuracy of weather and geological disasters in complex terrain regions [48
]. Regrettably, these studies have not been demonstrated in Sichuan Basin, which is one of the most complex terrain regions in the world and an ideal model for studying the influence of topography on precipitation.
Therefore, in this study, the three IMERG hourly products (IMERG-HE, IMERG-HL, and IMERG-HF, whereas “H” signifies “hourly”) during the period of 2016 to 2018 are quantitatively assessed at high spatial resolution (0.1° × 0.1°) over Sichuan Basin. Moreover, a high quality gauge network data with stringent quality control are used along with the three IMERG products in the present study to: (1) assess and analyze the differences in the measurement accuracy of the IMERG hourly products over Sichuan Basin; (2) compare the spatial distribution characteristics among the three IMERG hourly products; and (3) evaluate the capability of reproducing the diurnal cycle structure of precipitation for the three IMERG hourly products. This study reveals the errors, spatial distribution, and diurnal cycle characteristics among the three IMERG products at grid and regional scales, and provide valuable insights for the algorithm developers to improve the IMERG product quality and for users to select high-quality IMERG products in many relevant applications.
The remaining sections of the paper are structured as follows: Section 2
introduces the study area, precipitation datasets, the data quality control methods, data preprocessing, and statistical metrics. Section 3
evaluates the quality of all three IMERG products at grid scale. Section 4
presents comprehensive evaluations of all three IMERG products with multi-time scales at regional scale, and compares the accuracy and detection capability of various precipitation intensity. Section 5
focuses on the analysis of the capability in reconstructing the diurnal cycle of precipitation among the IMERG-HE, IMERG-HL, and IMERG-HF products at seasonal scales. Finally, discussion and summary are provided in Section 6
3. IMERG Grid-Scale Evaluation
The IMERG hourly products are evaluated against the CMPA data via the continuous and contingency statistical metrics at each grid cell. All of the IMERG hourly products and CMPA data used for this comparison are grid data with 1505 cells (0.1° × 0.1°) over the Sichuan Basin. For the grid-scale evaluation, the time series of each grid cell from January 2016 to December 2018 is analyzed by calculating the statistical metrics of CC, RB, RMSE, POD, FAR, and CSI, as well as their distributions over the Sichuan Basin.
The spatial distributions of the statistical metrics for the IMERG-HE, IMERG-HL, and IMERG-HF products at grid scale over the Sichuan Basin are shown in Figure 2
. In general, all three IMERG products exhibit similar spatial patterns of continuous statistical metrics, while the IMERG-HL and IMERG-HF products present the higher accuracy and lower error, especially in the eastern Sichuan Basin (CC > 0.55 and RMSE < 0.7 mm), which means that the IMERG-HL and IMERG-HF products have made a significant improvement in the agreement with the CMPA data. However, CCs of the IMERG-HL and IMERG-HF products are lower in the northwestern Sichuan Basin (CCs < 0.45). In terms of RB, the NRT IMERG hourly products generally tend to slightly overestimate or underestimate precipitation in most interior and marginal areas, but severely overestimate precipitation in the northwestern and southwestern Sichuan Basin. Nevertheless, the PRT IMERG-HF product significantly expands the area of overestimation, exacerbates the extent of overestimation, but reduces the extent of underestimation. In brief, the IMERG-HF product shows a slight overestimation in most areas, but shows a severe overestimation in the northwest and southwest, as well as an underestimation in the marginal regions of Sichuan Basin. Compared to the NRT IMERG hourly products, the RMSE of the IMERG-HF product is generally lower in most areas, which implies that the precipitation estimation has been effectively improved, while this improvement is limited in the western and southwestern Sichuan Basin. For the capability of the precipitation detection, the IMERG-HF product apparently presents higher POD (>0.5) than the IMERG-HE and IMERG-HL products in most areas of the Sichuan Basin as a whole, especially in the western, northwestern, and eastern regions (POD > 0.6). However, the FAR spatial pattern of the IMERG-HF product shows no significant difference compared to the IMERG-HL product, and both of them present high FARs (>0.7) in the northwestern regions. As for CSI, comparing with the IMREG-HE and IMERG-HL products, the IMERG-HF product shows better precipitation detection capability, and has a higher CSI (>0.35) in the eastern and northeastern regions. This phenomenon indicates that the IMERG-HF product improves the accuracy of precipitation, but severely overestimates precipitation and exaggerates the FAR of light precipitation events, which commonly occur in the northwestern Sichuan Basin (also see Figure 1
The box plots of the metrics for the IMERG-HE, IMERG-HL, and IMERG-HF products are shown in Figure 3
, which represents dispersion characteristics of the statistical metrics. After being arranged in the ascending order, the statistical data is divided by quartiles into four equal parts, which include the first quartile (Q1, 25%), the second quartile (Q2, 50%), and the third quartile (Q3, 75%). More specifically, the bottom and top edges of the box are Q1 and Q3, and the band inside the box is Q2. Furthermore, the ends of the upper and lower whiskers indicate the maximum and minimum values, respectively. The distributions of the metrics for the three IMERG products show strong symmetry, but the ends of the whiskers for CC, RB, RMSE, and FAR are far from Q2, indicating that these metrics are not concentrated. In terms of CC and RMSE, the IMERG-HF product (Q2 of CC and RMSE are 0.54 and 0.77 mm, respectively) presents comparable performance with the IMERG-HL product (Q2 of CC and RMSE are 0.53 and 0.78 mm, respectively). By comparison, the IMERG-HE product presents a little poorer performance (Q2 of CC is 0.46 but higher Q2 of RMSE at 0.86 mm) than the IMERG-HL product. As for RB, all three IMERG products overestimate the precipitation in most areas (shown in Figure 2
d–f), with Q2 of 6.18%, 7.99%, and 21.16%, respectively, which indicates that the degree of overestimation of the IMERG-HF product is greater than that of the IMERG-HE and IMERG-HL products. By analyzing the metrics above, the IMERG-HL product shows the best agreement with the CMPA data than the IMERG-HE and IMERG-HF products at the grid scale evaluation. For the capability of the precipitation detection, the IMERG-HF product apparently presents higher POD and CSI (Q2 are 0.52 and 0.32, respectively) than the IMERG-HL (Q2 are 0.46 and 0.30, respectively) and IMERG-HE products (Q2 are 0.40 and 0.26, respectively); but Q2 of the FAR (0.53) is slightly higher than the IMERG-HL product (0.52) and lower than the IMERG-HE product (0.54). This implies that after correcting the deviation using the GPCC data, the light precipitation detection capability of the IMERG-HF product is improved obviously, but the FAR of the IMERG-HF product has not been improved at all.
4. IMERG Region-Scale Evaluation
4.1. Annual Evaluation
The three IMERG hourly products are compared and evaluated at a regional scale in order to facilitate the quantitative evaluation of the precipitation estimation deviation during the three-year period. The statistical metrics of these products are calculated and shown in Table 2
The performance of the three IMERG hourly products at regional scale is very similar to their performance at grid scale in almost all of the statistical metrics, except for the RB. The RBs of the three IMERG hourly products at regional scale (3.84%, 4.97%, and 20.18% for HE, HL, and HF products, respectively) are slightly decreased compared to the grid scale evaluation (6.18%, 7.99%, and 21.16%, correspondingly and respectively), which indicates that the IMERG hourly products perform slightly better at regional scale than at grid scale. More specifically, as for the CC and RMSE, the IMERG-HL product presents comparable performance (CC of 0.51 and RMSE of 0.80 mm) as for the IMERG-HF products (CC of 0.53 and RMSE of 0.79 mm), and both of the products show better performance than the IMERG-HE product. However, the RB of the IMERG-HL product (4.97%) is significantly lower than that of the IMERG-HF product (20.18%). Moreover, the RBs of the three IMERG hourly products are positive numbers, which indicate that these products all overestimate precipitation at regional scale over Sichuan Basin during the period of 2016 to 2018, and the most serious overestimation comes from the IMERG-HF product. Moreover, the IMERG-HF product has the highest POD and CSI (0.51 and 0.32, respectively) compared to the other IMERG products, and has the same FAR (0.53) as the IMERG-HL product. Overall, the IMEG-HL product performs better in accurately estimating precipitation, while the IMERG-HF product shows a better precipitation detection capability in regional evaluation. Note that the gaps of all the statistical metrics for the IMERG-HL and IMERG-HF products are narrow in regional scale evaluation, except for the RB. This implies that the performance of the IMERG-HF product has not dramatically improved, thus, the precipitation estimation method used by the IMERG-HF product should be further improved.
4.2. Seasonal Evaluation
The statistical metrics from the evaluation of seasonal precipitation from the three IMERG hourly products are shown in Figure 4
. In general, the seasonal variation of the statistical metric profiles of the three IMERG hourly products are remarkably similar, and the gaps among the three IMERG hourly products for all the metric profiles are very close except for the RB. More specifically, the IMERG-HL and IMERG-HF products have a similar precipitation detection capability and present better performance than the IMERG-HE product in all seasons except for the RB. As for the RB, the IMERG-HF product exhibits more severe overestimation (21.49% in spring, 22.49% in summer, and 17.77% in autumn) than the IMERG-HE and IMERG-HL products, but only a slight overestimation in winter (3.87%). In contrast, the IMERG-HE and IMERG-HL products severely underestimate the precipitation in winter (−28.25% and −27.33%, respectively). By synthesizing all the statistical metrics in every season, compared to the other IMERG hourly products, the IMERG-HL product shows the best performance in precipitation observation in spring, summer, and autumn. As for winter precipitation, the IMERG-HF product shows the best precipitation detection capability and accuracy.
From the above analysis, the quality of the IMERG-HF product has not been improved much by incorporating the monthly observation of GPCC; in fact, it is even worse than the IMERG-HE and IMERG-HL products in terms of RB for spring, summer, and autumn precipitation. Several factors may explain this phenomenon: (1) more convective precipitation events occurred in summer, which may affect the detection capability of the microwave and infrared sensors, and consequently, reduce the detection accuracy; (2) the scale of convective precipitation may be smaller than the resolution of GPM, which may also result in the overestimation of the IMERG products in summer [64
]; (3) few precipitation events or light precipitation occur in winter over Sichuan Basin, and the large size of cloud drop increased by higher concentration of aerosols may intercept the ability of precipitation estimation by IMERG products [65
]; and (4) monthly observations of GPCC may be more accurate than climatological gauge data only in winter over Sichuan Basin, which may result in the IMERG-HF product to performs better than the NRT IMERG hourly products in winter.
4.3. Monthly Evaluation
The monthly quantitative evaluation of the three IMERG hourly products at regional scale during January 2016 to December 2018 is also conducted, and the results are shown in Figure 5
. It is evident that all three IMERG products have the weakest correlation with the CMPA data in January and December during 2016 to 2018. As for the RB, the IMERG-HF product overestimates the precipitation in almost all months, but only underestimates the precipitation in December. However, the IMERG-HE and IMERG-HL products tend to underestimate the precipitation in winter months, and overestimates the precipitation in other months. In addition, the RMSE profiles of the three IMERG hourly products have the obvious characteristics of seasonal variation, with the peak in June or July and the valley in January or December. As for the POD and FAR, all three IMERG hourly products with high PODs and low FARs can generally capture the precipitation during March to October, during which both of the IMERG-HL and IMERG-HF products show the similar precipitation detection capability and have better performance than the IMERG-HE product. However, the IMERG-HF product presents a little better precipitation detection capability than the other IMERG products in the winter months. This indicates that the IMERG-HL and IMERG-HF products with more calibrations are indeed able to improve the precipitation detection capability over Sichuan Basin, but are still too sensitive (with high FAR) in winter, which implies that the IMERG-HF does not exhibit the obvious improvements in terms of the FAR compared with other IMERG products over Sichuan Basin.
Considering what has been presented above, the IMERG-HF product exhibits a better performance in January and December, but for the rest of the months, the IMERG-HL product shows better agreement with the CMPA data and has a better precipitation detection capability than the other IMERG hourly products.
4.4. Precipitation Dectection Capability
The accuracy and detection capability of the IMERG hourly products is closely related to the intensity of precipitation. Therefore, evaluation of the IMERG hourly products for different precipitation intensities may provide favorable guidance for data selection in early warning of drought and extreme precipitation events. According to the distribution characteristics of annual precipitation over Sichuan Basin and the precipitation threshold (0.1 mm/h) set forth earlier, we divided the intensity of precipitation into nine intervals, and calculated the cumulative precipitation and frequency of the precipitation events at all nine intervals from the IMERG grid-scale products and the CMPA data during 2016 to 2018. The statistical results are shown in Figure 6
The distribution of the cumulative precipitation shows that the IMERG-HF product severely overestimates the precipitation compared to the CMPA data in the intensity range from 1.0 to 10 mm/h, while the IMERG-HE and IMERG-HL products only slightly overestimates the precipitation in the intensity ranges from 1.0 to 5.0 mm/h and greater than 10 mm/h, and underestimates the precipitation in the rest of the precipitation intensity ranges.
The distribution of precipitation frequency indicates that as for the light precipitation less than 1.0 mm/h, the IMERG-HF product is closer to the CMPA data than the IMERG-HE and IMERG-HL products, which suggest that both of the IMERG-HE and IMERG-HL products show the insufficient detection capability for the light precipitation, but the missed detection rate of the IMERG-HE product is significantly greater than that of the IMERG-HL product. More specifically, for the precipitation from 0.1 to 0.3 mm/h, the frequency values of the CMPA data and the three IMERG products are 3.38%, 2.48%, 3.01%, and 3.62%, respectively. These results indicate that the IMERG-HF product slightly exaggerate the detection ability of light precipitation. As the precipitation intensity increases (>1.0 mm/h), the IMERG-HE and IMERG-HL products show more consistent detection capability with the CMPA data, while the IMERG-HF product still has a certain degree of FAR.
In general, the accumulative deviations between the IMERG hourly products and the CMPA data are predominantly contributed by the precipitation with intensity from 1.0 to 10 mm/h. In this range, the overestimation of the IMERG-HE and IMERG-HL products are comparable, but less than that of the IMERG-HF product. In addition, the IMERG-HF product can capture more light precipitation events than the IMERG-HE and IMERG-HL products for the precipitation less than 1.0 mm/h, while the IMERG-HE and IMERG-HL products show a better detection capability for the precipitation more than 1.0 mm/h.
5. Evaluation on Reconstruction Ability of the Diurnal Precipitation Cycle
The diurnal cycle of precipitation is an important element in the evolution of weather systems, and it is greatly affected by the season and terrain [66
]. For the Sichuan Basin, obtaining an accurate diurnal cycle of precipitation will help further study the influencing mechanisms of terrain on atmospheric dynamics and thermodynamics. Therefore, for the three IMERG hourly products, their reconstruction ability of the diurnal precipitation cycle during January 2016 to December 2018 (1096 days in total) over Sichuan Basin was also evaluated.
Because the structure of the diurnal precipitation cycle is strongly influenced by the seasons, we analyzed the IMERG hourly products and CMPA data in each season. The specific processes are as follows: (1) convert from UTC of the IMERG products and CMPA data to LST at the center latitude for Sichuan Basin; (2) adjust the time to get the correct observations from 01:00 to 24:00 LST in one day, because the CMPA data at 00:00 LST represents the accumulated precipitation in the previous hour; (3) gather the accumulation time of effective data and calculate the mean precipitation intensity. The diurnal cycle and the continuous statistical metrics of the IMERG-HE, IMERG-HL, and IMERG-HF products for each season during 2016 to 2018 are shown in Figure 7
and Table 3
As for the variation characteristics of the diurnal precipitation cycle, the mean precipitation intensity of the CMPA data and the three IMERG hourly products show the same trend with time. However, mean precipitation intensity in spring, summer, and autumn is significantly higher than that in winter. In addition, the IMERG-HF product overestimates precipitation in all the seasons, which also results in a large RB (20.18%) in regional evaluation (Table 2
). For the time of peak precipitation in the diurnal cycle, the times of the peak precipitation observed by the CMPA data in the four seasons are at 03:00, 04:00, 06:00, and 02:00 LST, respectively. From Figure 7
, it is evident that the three IMERG hourly products can accurately capture the peak precipitation in spring and autumn. However, for the summer and winter peak precipitation, only the IMERG-HF product could accurately reproduce the characteristics of the peak precipitation, while the IMERG-HE and IMERG-HL products show a slight time lag and advance of diurnal peaks, respectively. More specifically, the IMERG-HE and IMERG-HL products show a time lag of 2 h and 1 h, respectively, for the peak precipitation in summer, while both of them advance the peak precipitation by 2 h in winter.
In addition, CC, RB, and RMSE are calculated and shown in Table 3
. In spring, the CCs of the IMERG-HE, IMERG-HL, and IMERG-HF products are very close (0.95, 0.96, and 0.94, respectively), but the RB and RMSE of the IMERG-HE product are the smallest among the three IMERG hourly products. Additionally, although the CC of the IMERG-HL product is slightly less than that of the IMERG-HF product in summer and autumn, the RB and RMSE are much better than that of the IMERG-HF product. However, for the diurnal precipitation cycle in winter, CC, RB, and RMSE of the IMERG-HF product are better than those of the IMERG-HE and IMERG-HL products.
In general, by synthesizing all the statistical metrics and the ability of accurately capturing peak precipitation, the IMERG-HE and IMERG-HF products show the better capability of reproducing the diurnal cycle in spring and winter, respectively. Additionally, for daily peak precipitation in summer, despite the fact that the capture time of the IMERG-HL product is 1 h later than that of the CMPA data, its comprehensive performance is better than the other IMERG hourly products. With regard to autumn precipitation, the IMERG-HL product performs better than the other IMERG products. Nonetheless, there is still a relatively large deviation in the reconstruction of the diurnal cycle between the IMERG hourly products and the CMPA data, which means that the IMERG hourly products have a long way to go for the optimization of precipitation estimation algorithm. Specifically, the improvement of precipitation estimation algorithm should focus on reducing the overestimation of the daily precipitation in spring, summer, and autumn, and improving the correlation coefficient in winter over Sichuan Basin.
In order to more carefully evaluate the ability of the IMERG hourly products in reproducing the diurnal cycle, we divided the time range from 01:00 to 24:00 into three periods with an interval of 8 h, i.e., 01:00–08:00, 09:00–16:00, and 17:00–24:00. After that, for all precipitation events (intensity > 0.1 mm/h), the mean intensity and frequency of the CMPA data and the three IMERG hourly products are calculated for the four seasons, and their statistical results are shown in Figure 8
As for the precipitation frequency, in the period of 01:00–08:00 LST in all four seasons, the precipitation frequency observed by the CMPA data is more than 40%, which is much higher than that of other time periods. This phenomenon fully reflected the characteristics of “nocturnal rain” in the Sichuan Basin. Comparing the precipitation frequency of the diurnal cycle observed by the IMERG hourly products to the CMPA data, it can be seen that the two data are very consistent in summer, but some differences exist in other seasons. More specifically, the difference between them mainly occurs at 09:00–16:00 in spring, 01:00–08:00 in autumn, and 09:00–24:00 in winter. In addition, for the precipitation frequency of the diurnal cycle, the IMERG-HE product is apparently closer to the CMPA data than the IMERG-HL and the IMERG-HF products in spring and autumn, but the IMERG-HL product and the IMERG-HF product perform slightly better in summer and winter, respectively.
As for the mean precipitation intensity within the diurnal cycle, compared to the IMERG-HE and IMERG-HL products, the IMERG-HF product overestimates the precipitation more severely throughout the day in spring, summer, and autumn. Specifically, all the IMERG hourly products overestimate the precipitation from 01:00 to 08:00 and 17:00 to 24:00 LST in spring, but accurately estimate the precipitation from 09:00 to 16:00 LST in spring. In summer, all the IMERG hourly products also overestimate the precipitation from 01:00 to 08:00 LST, but slightly overestimate the precipitation at other time periods. Moreover, for all time periods of a day in autumn, the IMERG-HF product overestimates the precipitation, while the IMERG-HE and IMERG-HL products slightly underestimate the precipitation. In winter, fewer precipitation events occurred in Sichuan Basin, thus, for the CMPA data and IMERG hourly products, the mean precipitation intensity of the diurnal cycle is relatively small throughout the day, resulting in little difference between them. Overall, all the IMERG hourly products overestimate the early morning (01:00–08:00 LST) precipitation, which may be linked to the increased cloud drop size caused by a higher concentration of aerosols over Sichuan Basin [65
In conclusion, the statistical results of precipitation intensity and frequency within a diurnal cycle at each time period show that the IMERG-HE product reflect the variation characteristics of diurnal precipitation more accurately in spring, while the IMERG-HL product performs better than the other IMERG hourly products in summer and autumn. In addition, the IMERG-HF product has a better capability of diurnal precipitation reconstruction in winter.
6. Summary and Conclusions
This study is the first to evaluate the precipitation estimates from the IMERG-HE, IMERG-HL, and IMERG-HF products using the CMPA data over Sichuan Basin across three years (from January 2016 to December 2018). The evaluation results provide useful references for other similar regions, especially for complex terrain regions. Moreover, the evaluation work can also provide valuable insights, not only for the algorithm developers to improve the retrieval processes of the IMERG hourly products, and to achieve a better data quality, especially for the NRT IMERG hourly products, but also for the meteorology related users to select high quality IMERG hourly products in many relevant applications, such as flood early warning, weather disaster forecasting, or hydrological model development. The main conclusions are summarized as follows:
(1) For the grid-scale evaluation of the IMERG hourly products, the IMERG-HL product shows comparable performance (high CC and low RMSE at 0.53 and 0.78 mm, respectively) to the PRT IMERG-HF product (CC and RMSE are 0.54 and 0.77 mm, respectively), but the IMERG-HL product reduces RB from 21.16% to 7.99%. By comparison, both of them have a much better performance than the IMERG-HE product (CC and RMSE are 0.46 and 0.86 mm, respectively). In terms of spatial distribution, both of the IMERG-HL and IMERG-HF products have high CC (>0.55) and low FAR (<0.4) in the eastern Sichuan Basin, where heavy precipitation events occur frequently. However, they show low CC (<0.45) and high FAR (>0.65) in the northwestern Sichuan Basin, where light precipitation events occur frequently. For the capability of detecting the precipitation events, the IMERG-HF product demonstrates better performance with highest POD (0.52) and CSI (0.32). However, the IMERG-HF product presents slightly poorer performance (FAR is 0.53) than the IMERG-HL product (FAR is 0.52). Thus, after correcting the deviation using the GPCC data, the precipitation detection capability of the IMERG-HF product is improved obviously, but the FAR has not been improved at all, implying that precipitation retrieval algorithms used in the IMERG-HF product should be revised.
(2) Compared to the grid-scale evaluation, the performance of the IMERG hourly products on regional evaluation is improved a little in regard to almost all statistical metrics, except for the RB. However, the IMERG-HF product still overestimates the precipitation more seriously (RB is 20.18%) than the IMERG-HE and IMERG-HL products (RBs are 3.84% and 4.97%, respectively). In addition, the IMERG-HL product gets a better ability to accurately estimate precipitation than the other IMERG hourly products. Nonetheless, the IMERG-HF product, with the highest POD and lowest FAR, shows a better capability than the other IMERG hourly products in detecting precipitation events over larger areas. For the seasonal evaluation, the performance of the IMERG-HF product is close to the IMERG-HL product with similar CCs, RMSEs, PODs, FARs, and CSIs for all four seasons. Additionally, in terms of RB, the IMERG-HL product shows better performance than the IMERG-HF product in spring, summer, and autumn. Conversely, in winter, the IMERG-HF product only slightly overestimates the precipitation (RB is 3.87%), but the IMERG-HE and IMERG-HL products severely underestimates the precipitation (RB are −28.25% and −27.33%, respectively). The largest RB of the NRT IMERG hourly products suggests that the IMERG algorithm developers should take immediate actions to correct data when the NRT products are used in the study of hydrological forecasting and drought monitoring. For the monthly evaluation, the CC, RMSE, POD, FAR, and CSI profiles of the three IMERG hourly products have the obvious characteristics of seasonal variations, with the peak in June or July and the valley in January or December. In addition, for the rest of the months, the IMERG-HL product is slightly better than the other IMERG hourly products. Compared to the NRT IMERG hourly products, the IMERG-HF product shows better capability of light precipitation (<1.0 mm/h) detection. However, for precipitation intensity more than 1.0 mm/h, the precipitation detection capability of the NRT IMERG hourly products is closer to that of the CMPA data. Furthermore, in the aspect of the cumulative precipitation, the difference between the IMERG-HF product and the CMPA data is the largest than the NRT IMERG hourly products, and the difference mainly comes from moderate precipitation events (1.0–10 mm/h).
(3) Regarding the peak precipitation reconstruction ability of the diurnal cycle, the IMERG-HF product can accurately capture the precipitation in all four seasons, while the IMERG-HE and IMERG-HL products show slight time lag (2 h and 1 h, respectively) of diurnal peaks of intensity for summer precipitation, and show the same time advance (2 h) for winter precipitation. By synthesizing the analysis of the precipitation intensity and frequency, it can be found that the IMERG-HE product shows better capability in reproducing the diurnal cycle for the spring precipitation, while the IMERG-HL product performs better in summer and autumn, and the IMERG-HF product can better reflect the characteristics of the diurnal cycle in winter. In addition, we found that the time of overestimation in the three IMERG hourly products for spring, summer, and autumn precipitation mainly happens in the early morning (01:00–08:00 LST). Therefore, the capability of all the IMERG hourly products should be improved in the future, especially for the early morning precipitation.
In summary, the performance of all the IMERG hourly products is not very satisfactory in most of the above evaluations. Moreover, the IMERG-HF product is not always superior to that of the NRT IMERG hourly products in some evaluations. For instance, its RB and FAR is significantly higher than the NRT IMERG hourly products in grid-scale and region-scale evaluations, which indicates that the monthly gauge correction aggravates the overestimation of the IMERG-HF product over Sichuan Basin. Therefore, the quality of the IMERG hourly products needs to be improved further to get more accurate precipitation estimates, especially for light precipitation. Our future studies on the evaluation of the IMERG hourly products will focus on exploring the source of error and uncertainty over complex mountainous regions, in order to better understand the variations of the precipitation estimation error with different data sources or thresholds and to prompt the development and application of the IMERG algorithms and products.