**1. Introduction**

Flash floods, triggered by heavy rainfall (i.e., short duration, high intensity), are the rapid flooding of water within minutes up to several hours in small basins (hundred square kilometers or less) [1]. It is one of the most devastating floods in the world, which can cause grea<sup>t</sup> economic losses and casualties. From 1 October 2007 to 1 October 2015, the National Weather Service (NWS) of the U.S. reported that

278 people died from flash floods, 10% of which resulted in an average property loss of over \$100,000 [2]. China is also su ffering from severe flash floods, where 984 people have died from flash floods every year on average since 1950 [3]. With the increase in global precipitation and rapid economic development, the frequency and impacts of flash floods will be further exacerbated [4]. The most crucial approach to adapt to flash floods is accurate warnings, which can leave people with more time to respond to these emergencies. However, flash flood early warning remains challenging, especially given the conditions of a short lead time (less than 1–3 h). Accurate and continuous precipitation datasets with high spatial (i.e., 1–4 km) and temporal resolutions (i.e., 5 min to hourly) are critical to the success of flood warnings [5]. Considering climate change and land degradation processes, new tools for flood disaster monitoring and reduction are strongly required. Satellite precipitation products have wide coverage, a high spatiotemporal resolution, and easy data acquisition, and are not restricted by terrain conditions. These products are extremely important to flood warnings in mountainous areas prone to flash floods, where there is scarcely measured rainfall data, and some satellite precipitation products have been used to assess the floods over the basins. However, satellite precipitation products may underestimate extremely strong precipitation, and the accuracy of the satellite precipitation products in catching flash floods remains poorly understood [6,7].

Nowadays, numerous satellite precipitation products, such as the Tropical Rainfall Measuring Mission (TRMM) and Integrated Multi-Satellite Retrievals for Global Precipitation Measurement (GPM) Mission (IMERG), are available for providing post-real-time (PRT) estimation and near-real-time (NRT) estimation. PRT products generally undergo ground-based gauge adjustments while NRT products do not, indicating the former is usually more accurate than the latter [8]. These satellite precipitation products generally have high spatiotemporal resolutions (finer than a 0.25◦ spatial resolution and shorter than the daily temporal interval) with a wide quasi-global coverage (broader than the 50◦ N–50◦ S latitude band); these products are very useful for hydrological studies, especially in data-sparse or ungauged basins [9]. GPM is a new generation of satellite products inheriting TRMM satellite products, with more comprehensive data and higher spatiotemporal accuracy. IMERG is the Level 3 precipitation estimation algorithm of GPM, which can combine gauged rainfall data and satellite data to obtain global rainfall data. IMERG provides three types of products, including the NRT IMERG Early run (hereafter IMERG-E), Late run (hereafter IMERG-L) and the PRT IMERG Final run (hereafter IMERG-F) [10]. Previous studies have proclaimed that IMERG-F has higher accuracy, especially on land, while the IMERG-E and IMERG-L products have better timeliness, which is attractive for flood prediction and monitoring. However, due to the limitations in observation position, density, topography, etc., it is di fficult to directly analyze the actual spatial distribution of rainfall with gauged rainfall data [11]. Meanwhile, satellite products also contain uncertainties from instruments, sampling, and retrieval algorithms, and therefore should be comprehensively validated using local gauge data [12].

An appropriate flash flood warning method is another key factor to ameliorate the flash flood warning accuracy. Typical examples include Flash Flood Guidance (FFG), Soil Water Index (SWI), Rainfall Triggering Index (RTI), etc. [13]. FFG is one of the most widely used methods developed by the US River Forecasting Center in the early 1970s, and it calculates the rainfall required to produce bank-full flood conditions associated with flash floods in a given time and area. The calculation steps of the FFG are as follows: (1) gaining the current soil moisture with the hydrological model; (2) reversing the flood peak flow of the basin outlet section; and (3) obtaining the critical rainfall required when the flow reaches the early warning flow [14]. However, FFG has more data requirements since it considers almost all influencing factors. For example, the soil moisture index is an indicator that can accurately describe the trend in soil moisture in the aeration zone, which is mainly calculated from the total water depth of a three-layer tank model with fixed parameters, and FFG has only one parameter related to the infiltration time; however, it is di fficult to obtain this parameter regardless of concentration calculation or distribution calculation [15]. The RTI comprehensively considers the e ffective accumulated rainfall and rainfall intensity in the prediction of flash floods. It focuses on antecedent conditions and has been

put into practice for over 10 years in Taiwan [13]. However, this method relies too much on rainfall stations, fails to fully consider intermittent rainfall, and is greatly a ffected by rainfall field segmentation. Meanwhile, using the deduction coe fficient of "t" days in the RTI model for the antecedent rainfall calculation will result in a higher false alert rate under some rainfall patterns. Therefore, Chen et al. (2017) proposed an improved RTI to calculate the antecedent rainfall and e ffective accumulated rainfall to solve the abovementioned problems, which achieved good practical application e ffects in Taiwan's 2017 disaster warning [16].

As satellite precipitation data is widely used in large-scale watershed hydrological simulation or land surface process simulation, it is still di fficult to meet the needs of flash flood prediction in small and medium-sized watersheds. Most of the existing research focus on the corrected products, but there scarcely has been application verification for real-time products. Most of the satellite products are applied on a large scale, and little attention has been paid to their applicability in small-scale flash flood warnings. Additionally, flash floods generally occur in small and medium-sized river basins with poor economic conditions, low station network coverage density, and there is grea<sup>t</sup> di fficulty in obtaining data. Given that IMERG products have high spatiotemporal precision, it is therefore of grea<sup>t</sup> practical significance to evaluate the applicability of these products to flash flood warnings. Moreover, regional studies are also very common and popular because precipitation exhibits strong spatial variations and di fferent products show varying performance over di fferent regions. Meanwhile, there are still few studies on the inter-comparison of the IMERG products, especially in China, where the superiority of its application in flash flood warnings still needs further exploration [17]. Therefore, taking Yunnan Province in China as the study area, based on two fifth-generation IMERG products (IMERG-E, IMERG-F) and China Meteorological Administration (CMA) data, this study first evaluated these two products with particular attention paid to their systematic and random errors. Then, the empirical RTI method was utilized to evaluate the applicability of the di fferent IMERG products in flash flood warnings. This paper is further organized as follows: Section 2 describes the materials and data; Section 3 presents an assessment of IMERG-E and IMERG-F based on locally measured data and analyzes the e ffects of flash flood warnings based on satellite precipitation data; and the conclusions are summarized in Section 4.

#### **2. Materials and Data**
