2.2.1. Satellite Data

TRMM Multi-Satellite Precipitation Analysis (TMPA) (50◦ N–50◦ S) can retrieve microwave–infrared satellite precipitation estimates with gauge adjustments and can generate rational precipitation estimates at fine spatiotemporal scales (0.25◦ × 0.25◦ and 3-hourly temporal intervals) within the global scope (60◦ N–60◦ S) [21,22]. As the successor of TRMM, the GPM constellation consists of a core observation platform and 10 cooperative satellites to observe global precipitation through inter-satellite cooperation. The GPM program offers three different levels of data products, all of which are available on the NASA website (http://www.nasa.gov/mission\_pages/GPM). This study focuses on the GPM Level 3 products based on the IMERG algorithm. IMERG not only has a high spatial resolution (0.1◦ × 0.1◦, 0.5 h) and fully global coverage (60◦ S~60◦ N), but also predicts flood disasters by reducing the uncertainty associated with short-term precipitation accumulation. At present, IMERG data has been developed from the first version to the sixth version, and IMERG-V06B is the latest version of the satellite rainfall data. However, Mohammad et al. (2019) verified that IMERG V05 performs better than V06, and therefore this study has employed IMERG V05 to estimate the precipitation data [23].

IMERG has three different products: Early, Late, and Final. In the real-time phase, the IMERG data generation system generates Early products after running once and then generates Late products after running again. The main difference between them is that the Early product is generated only by the forward propagation algorithm in the cloud mobile vector propagation algorithm, while the Late product is added with the backpropagation algorithm. The time-lag product Final introduces more

sensor data sources based on the Late product. Meanwhile, the time of the extracted satellite data is matched with the station's data (08:00 a.m.) [24]. Table 1 shows the detailed information of the three IMERG products, with a study period of 2015–2018. Since IMERG-E and IMERG-L products are NRT products, they are released after 4 h and 12 h after observation, respectively; in turn, the IMERG-F is a PRT product, which has been calibrated for the deviation of the ground rainfall station, so it has a high accuracy and is usually released after two months of observation. Therefore, this study has adopted CMA data as calibration data, pays attention to the accuracy of the IMERG rainfall data to capture flash flood disasters, and combines the precipitation data estimated by IMERG-E and IMERG-F to discuss the accuracy of capturing flash flood, and thereby contributes to obtaining the static early warning thresholds.

**Table 1.** List of Integrated Multi-Satellite Retrievals for Global Precipitation Measurement (IMERG) products.


#### 2.2.2. Ground Observation Precipitation Data

The regional hourly precipitation integration products of the CMA (http://cdc.cma.gov.cn) is a reference product, which is based on the hourly precipitation observed by more than 30,000 automatic weather stations nationwide and the Climate Prediction Center morphing (CMORPH) satellite. Using the probability density function and optimal interpolation method, the precipitation fusion products with a high spatiotemporal resolution (0.1◦ × 0.1◦ and 1-h interval) have been generated in China. Among them, the optimal interpolation method is to combine the semi-hourly infrared data of the Geostationary Earth Orbit (GEO) satellite to interpolate the Passive microwave (PMW) inversion data, and then obtain the relatively fine precipitation. The gauge data involved in the CMA are subject to extreme quality control, including consistency checks, both internal and spatiotemporal. Meanwhile, the complex terrain, the low economic level in remote mountainous areas, and the sparse distribution of the ground stations have led to large errors in the measured data. The CMA data has a wide coverage and high spatiotemporal resolution; the overall error of this product is within 10%, and the error for heavy precipitation and site sparse areas is within 20%, which is more accurate than other similar products in China. Therefore, CMA products are suitable as the calibration data for product evaluation, but their scale should be consistent with the IMERG product in the calibration process [25]. Besides, when using CMA data to estimate the rainfall distribution, we should consider that the dataset may have a delayed phenomenon since it is actually a fusion of measured data and CMORPH satellite rainfall data.

The rainfall level is defined by the CMA, for example, the intraday rainfall is categorized into the rain and light-moderate rain (0–25 mm), heavy rain (25–50 mm), and rainstorm (>50 mm). Meanwhile, the minimum amount of precipitation that the gauges can measure is 0.1 mm. The critical rainfall in Yunnan Province mainly comes from the China rainstorm parameters atlas, which describes the statistical characteristics and laws of China's rainstorms. Based on the study period 2015–2018, the flash flood data were mainly from officially published data. In this study, a total of 120 flash flood events were analyzed, all of which resulted in deaths or missing of people. Additionally, some other data were also employed; for example, the topographic data came from a 1:250,000 digital elevation model (DEM) of the Yunnan Province.
