of IMERG. IMERG performs better than other products during winter in TP, which is unexpected, calling for further investigation. Although SM2RAIN is among the worst products in the whole China, TP, and NE areas, its accuracy is relatively higher in XJ because soil moisture is more sensitive to precipitation in arid regions and seldom saturated due to the scarcity of extreme precipitation. Reanalysis products including ERA5, ERA‐Interim, and MERRA2 perform better during winter, except in TP where large standard deviation and RMSD degrade their overall performance, although they acquire the highest CC. In NE, reanalysis products can also capture precipitation well, except during summer when precipitation is notably larger than other seasons. This is because reanalysis products are better at estimating stratiform precipitation than convective precipitation.
As shown in performance diagrams (Figure 2.6), GSMaP is the best in detecting precipitation occurrence in all seasons and subregions. Points representing the three reanalysis products are always clustered together and only second to GSMaP in most cases, indicating that their occurrence detection mechanism is excellent but probably consistent. IMERG performs reasonably well in spring, summer, and autumn but notably worse in winter, as also shown in Taylor diagrams. Points of IMERG_cal and IMERG_uncal are always overlapped, meaning that the monthly‐scale gauge adjustment has little impact on occurrence detection on the daily scale.
PCDR and CHIRPS are better than other satellite products at detecting precipitation during winter in TP, XJ, and NE, showing the potential of infrared data in the cold season. It is found that CMORPH is the worst during winter, particularly in XJ and NE, where it cannot capture precipitation events correctly according to the extremely small POD values and large FAR values. CMORPH relies on passive microwave data to acquire seamless precipitation estimates, and infrared data are only used to calculated cloud motion vectors. In contrast, microwave‐infrared combined products such as IMERG utilize infrared data to fill the gap between passive microwave sensors and replace passive microwave estimates over snowy and icy surface types. The limitation of passive microwave data leads to the degraded quality of CMORPH during winter.
2.4.3. Snowfall Pattern, Evaluation, and Trend
The snowfall distributions of IMERG, ERA5, ERA‐Interim, MERRA2, and CGDPA are shown in Figure 2.7. CGDPA does not provide direct snowfall estimates. Therefore, we first calculate rainfall‐snowfall ratios from three reanalysis products and IMERG, and then use the ratios to estimate snowfall from CGDPA. According to Figure 2.7, snowfall is mainly distributed over mountainous regions in China, including the whole TP, the Tianshan Mountains in XJ, and the Altai Mountains, which cover a small area in the northern corner of XJ. For the TP, the Qiangtang Basin (30°–35°N and 80°–90°E) shows little snowfall due to its dry climate. The high latitude regions are widely covered by snowfall, whereas the total amount is not large. It is worth mentioning that the Brahmaputra Grand Canyon (about 30°N, 95°E) has a very large amount of snowfall benefiting from the high precipitation caused by the southwest monsoon.
Meanwhile, these products also show substantial differences. The snowfall amounts produced by reanalysis products are much larger than those from CGDPA and IMERG. Among the three reanalysis products, ERA‐Interim has a coarse resolution and thus cannot represent the spatial details of snowfall. Moreover, ERA‐Interim fails to capture snowfall in southern regions. MERRA2 generates much stronger snowfall along the Himalayan Mountains than other products. CGDPA shows similar patterns with reanalysis products because it depends on reanalysis data to realize precipitation phase discrimination, whereas the snowfall intensity of CGDPA is smaller than that of reanalysis products, particularly over the northern and western TP. When station‐based snowfall measurement information is absent, the reanalysis‐based method (Figure 2.7a, b, c) and the temperature‐based method (using a temperature threshold to distinguish rain and snow, i.e., Figure 2.7d) are used to obtain snowfall estimates. As shown in Figure 2.7, both methods indicate that CGDPA shows lower snowfall amounts than reanalysis products. Although it is known that stations may have large errors in snowfall measurement, it is hard to conclude whether CGDPA underestimates or reanalysis products overestimate snowfall without further studies. IMERG generates the weakest snowfall across China.
Due to the lack of snowfall observations, we use MTC to evaluate the quality of CGDPA, ERA5, MERRA2, and IMERG snowfall estimates (Figure 2.8). ERA‐Interim is not included because it is similar to ERA5, as shown in Figure 2.7. Only samples meeting two criteria are selected: (1) all products detect the occurrence of a snowfall event (i.e., snowfall probability >50%), and (2) a grid must have at least 200 effective snowfall events. The criteria are designed to (1) reduce the effect of different rainfall‐snowfall classification schemes and (2) exclude zero snowfall samples, which are dominant in numbers and will greatly affect evaluation results. Two triplets are designed. The first includes CGDPA based on MERRA2 snowfall information, IMERG, and ERA5. The second includes CGDPA based on ERA5 snowfall information, IMERG, and MERRA2. Both triplets can satisfy the independence requirement of MTC. The two triplets produce similar CC and RMSE values for CGDPA and IMERG, which can partly verify the effectiveness of MTC. The metric maps for CGDPA and IMERG in Figure 2.8 are provided by the first triplet.
Figure 2.7 Spatial distributions of snowfall from 2000 to 2018 for (a–d) CGDPA, (e) ERA5, (f) ERA‐Interim, (g) MERRA2, and (h) IMERG. CGDPA1‐4 employs rainfall‐snowfall classification data provided by (e–h), respectively.
Source: Based on Tang et al. (2020), Figure 12, p 13 / Elsevier.
In most regions of China, including XJ and NE, ERA5 performs better than other products with the highest CC, while MERRA2 shows the highest CC in the TP. It is noted that there is a hollow eastern to the TP, which is the relatively warm Sichuan Basin with limited snowfall. CGDPA exhibits relatively higher CC in east and south China than in north and west China. Overall, CGDPA is much worse than ERA5 and MERRA2 in snowfall estimation concerning CC, whereas it is hard to determine which one is the best based on RMSE. By comparing Figures 2.7 and 2.8, it is found that ERA5 and MERRA2 exhibit large RMSE in regions with limited snowfall, probably due to the overestimation of the frequency of light precipitation. In contrast, CGDPA exhibits large RMSE in regions with abundant snowfall caused by its underestimation of snowfall amount, indicating that the current design of rain gauges cannot satisfy snowfall measurements. Comparing with CGDPA and reanalysis products, IMERG shows smaller CC over China and much larger RMSE over TP, XJ, and NE due to its underestimation.
Figure 2.8 CC and RMSE of CGDPA, ERA5, MERRA2, and IMERG according to the triple collocation analysis using data from 2000 to 2018. Only grid cells with more than 200 effective snowfall samples are included. The first triplet includes CGDPA, IMERG, and ERA5, and the second triplet
