2.2.2. GSMaP Near-Real-Time Precipitation Products

As one of Japanese GPM projects, GSMaP was implemented under the Japanese Precipitation Measuring Mission (PMM) science team with the target of providing a global precipitation map with high precision and high resolution [34]. The GSMaP algorithm uses various of PMW/IR sensors to produce the "best" precipitation estimates through several steps. First, several PMW radiometers carried by different satellites, such as the GPM microwave imager (GMI), TRMM microwave imager (TMI), special sensor microwave imager/sounder (SSMIS), advanced microwave scanning radiometer 2 (AMSR2), advanced microwave sounding unit-A (AMSU-A), and microwave humidity sounder (MHS), are used to retrieve quantitative precipitation estimates [33,46]. Then, it uses the cloud motion vector derived from successive geo-IR images to propagate the precipitation area for filling the gaps between PMW-based estimates, which is similar to CMORPH. In addition, a Kalman filter model is applied to modify precipitation rates after propagation. Finally, the forward and backward propagated precipitation estimates are weighted and combined to produce the standard GSMaP\_MVK product. At the beginning of design, the GSMaP algorithm developers did not consider near-real-time operation and data availability. To meet this demand, a near-real-time product of the GSMaP (GSMaP\_NRT) with resolutions of 0.1◦ and 1 h was developed. Different from GSMaP\_MVK, the GSMaP\_NRT only employs forward cloud movement to hold operability in near-real-time. The emergence of GSMaP\_NRT product attracts a lot of data users, owing to its short latency (~3 h after observation). To reduce the bias in the GSMaP\_NRT product, a new algorithm introducing gauge information to GSMaP\_NRT (i.e., GSMaP\_Gauge\_NRT) is currently under development. In the GSMaP\_Gauge\_NRT product, a precipitation error parameters model was created based on the historical database of GSMaP\_Gauge. Then, these parameters were used to adjust the GSMaP\_NRT estimation in near-real-time to improve the precision of GSMaP\_NRT. Considering that the GSMaP\_Gauge\_NRT product does not use the gauge measurement directly, this makes the GSMaP\_Gauge\_NRT independent of the ground gauge observations. Recently, the latest GSMaP algorithm upgraded to version 7, and its near-real-time products were made available after 17 January 2017. In this study, the GSMaP\_Gauge\_NRT product and the uncalibrated GSMaP\_NRT product were investigated over a complete two-year period (from September 2017 to August 2019). Both satellite precipitations were aggregated into daily amounts, with a 0.25◦ × 0.25◦ resolution corresponding to the gauge data.

#### *2.3. Methods*

In this study, we conducted the assessment and comparison of GSMaP precipitation based on continuous statistical metrics and contingency table metrics. The continuous metrics included correlation coefficient (CC), mean error (ME), root mean squared error (RMSE), and relative bias (BIAS), which are widely used to quantitatively represent the degree of agreement and the error between satellite precipitation and gauge observations. These continuous metrics were calculated by the following equation:

$$\text{CC} = \frac{\sum\_{i=1}^{n} \left( G\_i - \overline{G} \right) \left( S\_i - \overline{S} \right)}{\sqrt{\sum\_{i=1}^{n} \left( G\_i - \overline{G} \right)^2} \cdot \sqrt{\sum\_{i=1}^{n} \left( S\_i - \overline{S} \right)^2}} \tag{1}$$

$$\text{ME} = \frac{1}{n} \sum\_{i=1}^{n} (\mathbf{S}\_i - \mathbf{G}\_i) \tag{2}$$

$$\text{RMSE} = \sqrt{\frac{1}{n} \sum\_{i=1}^{n} \left(\mathbf{S}\_i - \mathbf{G}\_i\right)^2} \tag{3}$$

$$\text{BIAS} = \frac{\sum\_{i=1}^{n} (S\_i - G\_i)}{\sum\_{i=1}^{n} G\_i} \times 100\% \tag{4}$$

where *Si* and *Gi* are the precipitation values from satellite estimation and gauge data, respectively; correspondingly, *S* and *G* are their mean precipitation, and *n* is the number of samples.

In addition, three contingency table metrics were adopted to evaluate the capability of satellite precipitation in the detection of precipitation events. These categorical metrics were the probability of detection (POD), false alarm ratio (FAR), and critical success index (CSI). POD is usually used to represent the fraction of precipitation events that correctly detected by satellite among all the actual precipitation events. FAR denotes the ratio of false alarm by satellite among the total satellite detected events. The CSI, combining the correct hit, false alarm, and missed event, is a more comprehensive score. The formulas of these contingency table metrics are listed below:

$$\text{POD} = \frac{H}{H+M} \tag{5}$$

$$\text{FAR} = \frac{F}{H+F} \tag{6}$$

$$\text{CSI} = \frac{H}{H + M + F} \tag{7}$$

where *H*, *M*, and *F* are the numbers of different precipitation events: Hit (both satellite estimates and gauge observations detect rain), miss (observed rain that is not detected by satellite), false (rain detected but not observed). Here, a commonly used threshold of 1.0 mm/day was set to define the rain/no rain event, as suggested by many previous studies [47–50].

For more detailed description of above continuous statistical metrics and contingency table metrics, readers can refer to Yong et al. [25] and Lu et al. [30]. We need to point out that all metrics were calculated in the 0.25◦ × 0.25◦ grid boxes with at least one gauge in order to ensure more convincing results (gauge distribution shows in Figure 1a). However, we also calculated metric value in every grid box to enable a visualization when presenting continuous spatial distribution (Figure 2).

**Figure 2.** Spatial distribution of statistical indices derived from the GSMaP\_NRT (left column) and GSMaP\_Gauge\_NRT (right column) daily precipitation at 0.25◦ × 0.25◦ resolution over the Mainland China: (**a**,**b**) Correlation coefficient (CC), (**c**,**d**) root mean square error (RMSE), and (**e**,**f**) probablility of detection (POD).

#### **3. Results**
