*3.1. Daily Mean Precipitation*

Figure 1b–d displays the spatial distributions of two-year daily mean precipitation for GSMaP and CGDPA precipitation products. Generally speaking, the spatial distributions of GSMaP and CGDPA precipitations were similar, showing a downward gradient from the southeast China to the northwest China. However, a pronounced difference in the precipitation amount was found between the gauge observations and GSMaP satellite precipitation products. For example, compared with the CGDPA, the GSMaP\_NRT tended to underestimate the gauge observations in the southeast and overestimated them in the northwest. Impressively, the GSMaP\_NRT significantly overestimated the gauge precipitation in the Sichuan province due to the indirect retrieval of satellite precipitation. After the parameterized gauge calibration, the errors in GSMaP\_NRT were effectively suppressed, and the GSMaP\_Gauge\_NRT had a more reliable performance than the GSMaP\_NRT in capturing the spatial patterns of precipitation over China. Therefore, the GSMaP\_Gauge\_NRT combined historical gauge information to reduce biases, making it consistent with the ground measurements. This suggests that the parameterized adjustment procedures can effectively enhance the quality of the original GSMaP\_NRT satellite precipitation estimates.

#### *3.2. Comparison and Validation of GSMaP\_NRT and GSMaP\_Gauge\_NRT Products*

Figure 2 shows the spatial maps of CC, RMSE, and POD, which were computed from two GSMaP products against gauge observations over the Mainland China at the 0.25◦ × 0.25◦ resolution grid. Generally speaking, the spatial distributions of GSMaP\_Gauge\_NRT exhibited a great improvement compared to that of the GSMaP\_NRT, with higher CC, lower RMSE, and slightly better POD values. Over Mainland China, the CC increased from 0.58 with GSMaP\_NRT to 0.67 with GSMaP\_Gauge\_NRT and the RMSE dropped from 9.11 mm to 7.07, besides a small change of POD between GSMaP\_NRT (0.69) and GSMaP\_Gauge\_NRT (0.70). With respect to the spatial performance of error metrics, both GSMaP\_NRT and GSMaP\_Gauge\_NRT products exhibited similar features. That is, worse values of CC and POD occurred in the northwest and improved toward the southeast, while higher RMSE existed in the southeast. This phenomenon was reasonable because the RMSE value increased with increasing precipitation amounts, and southeastern China has more precipitation than other areas in China. Distributions of CC, RMSE, and POD indicate that GSMaP\_Gauge\_NRT performs better than GSMaP\_NRT over Mainland China, suggesting that the calibration in near-real-time can effectively reduce the error and improve detectability of GSMaP\_NRT.

We further inquired about the temporal behavior and the seasonal statistics of the GSMaP\_NRT and GSMaP\_Gauge\_NRT products over Mainland China. In order to ensure a more accurate comparison, only those grids that contained at least one gauge were taken to compute the statistical indices. Figure 3 depicts the monthly precipitation and monthly variations of statistical metrics from gauges and GSMaP precipitation products by calculating at daily scale. Table 1 summarizes the seasonal statistics including spring (March–May), summer (June–August), autumn (September–November), and winter (December–February). It can be see that both GSMaP\_NRT and GSMaP\_Gauge\_NRT products can generally capture the intra-annual and seasonal variation patterns of precipitation over China, with the rainy summer and dry winter (Figure 3a). The GSMaP\_NRT showed much more precipitation than gauge observations in most months, and the GSMaP\_Gauge\_NRT reduced this overestimation, which was more consistent with gauge observations. The time series of statistical indices clearly show that the GSMaP\_Gauge\_NRT outperforms GSMaP\_NRT with higher correlation, lower error, and better detection (Figure 3b–d). This further confirms that the calibration in the GSMaP\_Gauge\_NRT can substantially improve the quality of the original GSMaP\_NRT precipitation product. However, it is worth noting that the BIAS was increased in some months. This issue may be due to the fact that the overestimation and underestimation at different regions could cancel each other out when calculating the BIAS value. Focusing on the curves of CC and RMSE, we can conclude that the GSMaP\_Gauge\_NRT had better agreement with gauge observations than the GSMaP\_NRT products over China. In addition, we note that the performance of satellite precipitation productions showed obvious seasonally dependent variations, with better statistical indices in summer and worse in winter (Table 1). Taking CC as an example, the value of CC decreased from 0.62 in summer to 0.41 in winter for GSMaP\_NRT and from 0.67 to 0.58 for GSMaP\_Gauge\_NRT. During the winter months, the snow brought by the westerly winds was the main form of precipitation in the north and west regions of China. However, the complex radiative properties of ice particles and snowflakes restricted the retrieval capability of microwave radiation. On the one hand, the low-frequency channels of PMW sensors contain limited snow detection information, but account for most of PMW channels and are traditionally used to retrieve rain drops. On the other hand, more snow-related high-frequency channels will be seriously interfered in the snow-covered background surfaces, which often present a similar passive microwave signature as the falling snow [51]. Therefore, measuring solid precipitation and snow is a challenging task for the satellite precipitation retrievals, which was pointed by many previous studies [15,23,25,52]. The GSMaP\_Gauge\_NRT product can combine the gauge information to reduce the precipitation error of GSMaP\_NRT in winter, showing more consistency with gauge observations.

**Figure 3.** (**a**) Time series of mean monthly precipitation and monthly variations of statistical indices over grid boxes with at least one gauge in Mainland China: (**b**) Correlation coefficient (CC), (**c**) root mean square error (RMSE), (**d**) relative bias (BIAS), and (**e**) probablility of detection (POD).


**Table 1.** Seasonal statistics of GSMaP\_NRT and GSMaP\_Gauge\_NRT against ground observations from selected 0.25◦ grid boxes over the Mainland China.

Considering the diverse climate of China, it was rational to subdivide national-scale evaluation into regional analyses. Figure 4 shows the scatterplots of daily GSMaP\_NRT and GSMaP\_Gauge\_NRT against gauge observations for the selected grids over different climate regions. Clearly, over the four climate regions, all the scatterplots show that the scatter points of GSMaP\_Gauge\_NRT were clustered closer to the 1:1 line than those of GSMaP\_NRT estimates, meaning that the GSMaP\_Gauge\_NRT was more in agreement with gauge observations. The GSMaP\_NRT estimate significantly overestimated the gauge precipitation with BIAS range from 0.82% to 149.08% (see left column in Figure 4). After the calibration, these biases were effectively minimized in the GSMaP\_Gauge\_NRT product. Correspondingly, the CC values increased from GSMaP\_NRT to GSMaP\_Gauge\_NRT, and the RMSE values showed an apparent downward trend. However, in terms of the contingency table statistics, the improvements were not obvious. This suggests that the calibration can effectively reduce the bias but is not good at improving the skill of detecting rainy events. Additionally, Figure 4 illustrates that the two near-real-time GSMaP products had different performances at four climate regimes, with better agreement from gauge observations over the humid region (Figure 4a,b) and an unsatisfactory performance over the arid region (Figure 4g,h). This was likely caused by the different retrieval skills of rainfall types over different climate regimes. The arid region, covered with desert and high mountains, was dominated by short-lived convective precipitation and orographic precipitation. However, the satellite-based precipitation retrieval had difficulty coping with these two precipitation conditions. Moreover, the light rainfall and winter snow in arid region further imposed another challenge to the satellite-based precipitation estimates. Consequently, the satellite precipitation products usually had an unsatisfactory performance over the arid region in China, and this is consistent with the result of Yong et al. [39] and Chen et al. [46]. Compared to GSMaP\_NRT, the GSMaP\_Gauge\_NRT obviously improved the data accuracy, but it had low CC (0.36) and BIAS (35.08%) values over the arid region. Such results indicate that the current near-real-time gage calibration algorithm of GSMaP still have significant room for further improving the data quality over the arid region.

Theoretically, a value of satellite precipitation can be divided into four categories based on its ability to identify rain occurrences. A hit event means that both the satellite estimate and gauge reference detected rain, while miss precipitation suggests that a rain event was reported by gauge observation but not detected by satellite. In contrast, false precipitation means that precipitation was detected by satellite but not observed by gauge, and the rest part of precipitation means that that both satellite and gauge showed no rain. Based on the different precipitation events, we further computed the intensity distributions of daily precipitation amount to look into the error characteristics of GSMaP products (Figure 5). The rainfall intensity was binned with logarithmic scale across the range of 1–256 mm/day, and the daily averaged precipitation accumulation of each bin was calculated on the y-axis. The intensity distribution, which has different error components, can reveal detailed information on the error features at the event scale. As shown in Figure 5, the intensity distributions of total precipitation were generally similar to those of hit precipitation, suggesting that hit event was the dominate component of total precipitation. However, over the arid region, the false precipitation also accounted for a considerable proportion of the total precipitation. Considering that the false precipitation will amplify the total precipitation amount of satellite, the obvious overestimation of two near-real-time GSMaP products over the arid region was partly caused by the false rainy events. On the other hand, compared to GSMaP\_NRT, the intensity distributions of GSMaP\_Gauge\_NRT were closer to the gauge observations (first column in Figure 5), which indicates that the GSMaP\_Gauge\_NRT has better performance than the GSMaP\_NRT. Additionally, it can be seen that the calibration of GSMaP\_Gauge\_NRT mainly changed the intensity distributions at a moderate–high rain rate. For example, the GSMaP\_NRT had more precipitation amounts than gauge observations over the semi-humid region (Figure 5e). After the calibration, the curve of GSMaP\_Gauge\_NRT descended and was more consist with the gauge. In the third column of Figure 5, it was found that the GSMaP\_NRT and GSMaP\_Gauge\_NRT showed basically identical intensity distributions of miss precipitation over all four climate regions. This implies that the GSMaP\_Gauge\_NRT failed to correct precipitation events undetected by the satellites in the calibration process. Therefore, the future correction efforts of incorporating different precipitation component is recommended to improve the precision of satellite precipitation.

**Figure 4.** Scatterplots of the daily precipitation for GSMaP\_NRT (left) and GSMaP\_Gauge\_NRT (right) versus gauge observations at selected grid boxes over four climate regions: (**a**,**b**) Humid region; (**c**,**d**) semi-humid region; (**e**,**f**) semi-arid region; (**g**,**h**) arid region.

**Figure 5.** Daily intensity distribution of the total, hit, miss, and false precipitation over four climate regions: (**a**–**d**) Humid region; (**e**–**h**) semi-humid region; (**i**–**l**) semi-arid region; and (**m**–**p**) arid region. The total observed precipitation (black line) is also shown in the first two columns.
