Based on Tang et al. (2020), Figure 13, p 14 / Elsevier.
The trend of snowfall based on different products during the study period is shown in Figure 2.9. The trend according to reanalysis products can pass the significance test at the 95% significant level. According to ECMWF products (ERA5 and ERA‐Interim), the snowfall amount decreases at the rate of –2.90 and –2.29 mm/decade. This value is only –1.27 mm/decade for CGDPA. In contrast, MERRA2 and IMERG show that the snowfall amount increased from 2001 to 2018, with the increasing rate being 6.53 and 1.13 mm/decade, respectively. It is noted that the TP shows a very significant trend of snowfall according to all products except CGDPA, whereas MERRA2 and IMERG show an inverse trend compared to CGDPA, ERA5, and ERA‐Interim. It is amazing to find that the increasing rate of MERRA2 is as large as 55.63 mm/decade. Some studies attempted to explore the trend of snowfall in TP using various data sources. However, their conclusions are also inconsistent. In summary, IMERG needs to improve its snowfall estimation, and all products may need to increase the credibility of snowfall trends.
Figure 2.9 Snowfall trends of CGDPA, ERA‐Interim, MERRA2, and IMERG from 2001 to 2018 over China. Solid lines indicate that the trend is significant with a 95% level of confidence. The numbers on the top of the figure are the rate of trend (mm/decade).
Source: Based on Tang et al. (2020).
2.4.4. Applicability of IMERG in Flash Flood Warning
In this section, three types of rainfall products are used, including the IMERG Early run (IMERG‐E), IMERG Final run (IMERG‐F), and China Meteorological Administration hourly data (CMA). The three products are accumulated to 1 h, 3 h, 6 h, and 24 h, respectively, and then the RTI method is employed to calculate the corresponding effective accumulated rainfall. At the same time, combining the frequency of flash floods and the atlas of rainstorm statistical parameters, the multiperiod critical precipitation of 1 h, 3 h, 6 h, and 24 h (hereinafter referred to as CR1, CR3, CR6, CR24) is obtained. Then, we construct a G(x) warning model for each effective cumulative rainfall (Rt) and critical rainfall (CRt). Since this model does not consider many potential influencing factors, such as vegetation, underlying surface, etc., flash flood risk distribution maps should also be considered when issuing flash flood warnings. Integrating the flash flood risk map obtained by Ma et al. (2019), Figures 2.10 and 2.11 show the flood warning results. For specific products and finer time scales, flooding events are more detectable.
For a specific product, the detectability of flash flood events is better for finer temporal scale. The hit rate of CR6 and CR24 is lower, less than 60%. Most of the flash flood events captured by these three precipitation products are located in western Yunnan, while in the relatively flat southwest and central Yunnan, there are fewer flash flood outbreaks. In general, if a flash flood event cannot be captured in a short period, it will not be captured at a lower time scale. Compared with IMERG‐F, IMERG‐E is significantly worse at capturing flood events; IMERG‐E’s hit rate is lower than 50% on all time scales. Conversely, IMERG‐F showed considerable accuracy in capturing CR1 and CR3 of the CMA flood, with a difference of 1%. For instance, the hit rate of IMERG‐F for CR1 is about 80%. But the hit rate decreases significantly as the time scale decreases from 1h to 24h.
Figure 2.10 The performance of the flash floods captured by the three products (IMERG‐F, IMERG‐E, CMA) for different times (1 h, 3 h, 6 h, 24 h). Red indicates that the warning issued has captured the flash flood event; blue indicates that the issued warning has not captured the flash flood event.
Source: Based on Ma et al. (2020).
Figure 2.11 Percentage of flash floods caught by CMA, IMERG‐E, and IMERG‐F, based on Ma et al. (2020). Note: CRt represents the critical rainfall at time t.
Source: Based on Ma et al. (2020).
According to the accuracy in capturing flood events, CMA is better than IMERG‐F, followed by IMERG‐E. However, another important issue is the latency of those products. CMA is a merged product based on CMORPH and ground station data and thus has a nonnegligible latency time. IMERG‐F has a latency time of more than one month. IMERG‐E has a latency time of several hours. Currently, many studies use near‐real‐time satellite products such as IMERG‐E to monitor flood hazards. However, operational flood forecasting must rely on nowcast precipitation from weather models. Therefore, how to make those products useful in practical application is still challenging.
2.5. SUMMARY AND CONCLUSION
This study evaluates the performance of retrospective IMERG precipitation estimates from 2000 to 2018 at hourly and daily scales and compares it with nine satellite and reanalysis precipitation estimates in China. Various metrics and evaluation methods are employed. Special attention is paid to snowfall validation using an objective error analysis method. In addition, IMERG products are applied to capture flash flood hazards in a typical region, Yunnan Province. The conclusions are as below.
IMERG performs very well on the daily scale in the whole China and three subregions concerning all accuracy metrics. It is better than other satellite and reanalysis products, except for GSMaP. At the hourly scale, IMERG is also satisfying and exhibits better performance than previous versions through indirect comparison. PCDR and CHIRPS exhibit limited performance compared to microwave‐based or microwave‐infrared combined products. However, PCDR and CHIRPS are better at estimating precipitation during winter in TP, XJ, and NE. In contrast, CMORPH almost loses the capability of detecting precipitation occurrence for the same season and regions, indicating that infrared data are more useful than passive microwave data under a cold climate or over snowy/icy surfaces. SM2RAIN is the worst among all products. SM2RAIN performs relatively better in arid regions such as XJ and Inner Mongolia than the moister regions in south and east China because soil moisture is seldom saturated in an arid climate.
Regarding the flood warning in Yunnan Province, we find that (1) IMERG‐F presents acceptable accuracy over the study area with a relatively high hourly correlation coefficient of 0.46 and relative bias of 23.33% on the grid, while IMERG‐E shows worse performance as expected; (2) by applying the RTI method, CMA and IMERG‐F exhibit similar performance in capturing flood hazards, while IMERG‐E captures fewer floods than CMA and IMERG‐F. Besides, all products show better performance at the finer temporal scale. It should be noted that the large time lag of IMERG‐F prevents its real application and makes it suitable only in historical studies. The climatological correction of near‐real‐time satellite precipitation products (e.g., IMERG‐E) is an important research direction to make them more useful in flood monitoring.
Meanwhile, IMERG also needs improvement.
