**2. Data and Methods**

## *2.1. Landslide Inventory*

The landslide information is obtained from four sources: (i) China Geological Environment Information site (http://www.cigem.cgs.gov.cn/), which provide the reports of geological disasters occurred in China; (ii) a global landslide catalog developed by Kirschbaum et al. (2015 [1]), available at https://catalog.data.gov/dataset/global-landslide-catalog-export; (iii) literatures including annual reports published by related institutions and/or government departments and local news report related to landslides, and (iv) other reports online. 771 landslide events were collected in a period covering 20 years from 1998 to 2017.

For each landslide, the corresponding information includes location, time, type of failure and triggering reason. Not all of this information is complete for every single event, i.e., some information is unknown. There are four kinds of spatial resolutions with the collected landslide events, namely, S1—province, S2—prefecture-level region, S3—county or district, and S4—town, village, country or site. As is shown in Table 1, 24 landslides (accounting for 3.1% of the total landslide events) are collected at the resolution of S1, 87 (11.3%) at S2, 286 (37.1%) at S3 and 374 (48.5%) at S4. In addition, there are also two kinds of temporal resolutions with the collected data, which are T1—the occurrence dates for 623 (80.8%) landslide events are known, and T2—the occurrence time (hour) for 148 (19.2%) landslide events are known.


**Table 1.** Information of landslide inventory in China during 1998–2017. NE: number of landslide events; S1: province; S2: prefecture-level region; S3: county/district; S4: village/town/country/site; T1: day; T2: hour; L: landslides; DF: debris flow; RF: rock fall.

In addition, information about the landslide types is also collected, which can be classified into debris-flow, rock fall and generic shallow landslide. Usually the mechanisms among them are considered to be similar, but in fact, it's not easy to distinguish them clearly, because sometimes the report uses imprecise language to describe the landslides. Thus, most of them are classified as generic shallow landslides (581, 75.36%). In fact, whether the landslide occurs for the first time or reactivates is very important in the landslide prediction, because the corresponding geological conditions would definitely change after the slope failure occurred. However, the collected landslide information does not contain this specific information, and the landslide model used here cannot simulate those processes. So, this paper mainly concentrates on the rainfall thresholds, and only the shallow landslides which occurred at the first time are considered. The above information of landslide inventory is listed in Table 1, including the number of landslide events with different spatio-temporal resolutions and landslide types in each year.

Landslide events used to reconstruct the rainfall thresholds are selected based on criteria as follows: (i) the location of the landslide has a spatial resolution of S3 or higher; (ii) the occurrence time of the landslide is known at least with a daily resolution (T1). The landslides which are not triggered by rainfall are excluded. Based on these requirements, 660 landslide events are selected to construct the rainfall thresholds.

The latitude and longitude of the landslide events can be acquired by map tool. Figure 1 shows the locations of landslide events and elevation distribution in China. The elevation data are derived from HydroSHEDS—Hydrological data and maps based on SHuttle Elevation Derivatives at multiple Scales, with a spatial resolution of 3 arc-seconds, which can be downloaded at http://hydrosheds.cr.usgs.gov [26]. Most of the landslides are distributed in transitional regions from high altitude to low altitude and hilly areas in the southern China, where heavy precipitation occurs frequently due to monsoon and typhoon activities. Landslides occurred in 1998–2015 (black triangle in Figure 1) are used to construct the rainfall thresholds, and those occurred in 2016–2017 (red triangle in Figure 1) are used for validation.

**Figure 1.** Location of landslides collected during 1998–2017 in China and the elevation (unit: m) distribution.

#### *2.2. Rainfall Data*

Three kinds of precipitation products are used to calculate the rainfall thresholds: (i) the merged precipitation product in China [27], and it is the main precipitation data used in this paper, (ii) satellite precipitation product produced by the NOAA Climate Prediction Center Morphing technique (CMORPH) [28], and (iii) CN05.1 [29] are also used in this paper. The hourly CMORPH and merged precipitation products are used to acquire the rainfall information when landslides occur, and daily CN05.1 is used to obtain the MAP for the normalization of the rainfall thresholds.

It has been pointed out that the development of satellite and ground-based radar technology has provided great support for the studies of rainfall thresholds [30]. The estimated precipitation from satellite-based data has a high spatio-temporal resolution, and the comprehensive quality evaluation in China is also good. In particular, the CMORPH performs better among the satellite precipitation products [31–33].

The CMORPH is developed by the National Oceanic and Atmospheric Administration (NOAA) Climate Prediction Center (CPC) in the United States. By using the geostationary satellite IR data to detect cloud systems and the associated motion characteristics, the cloud system advection vector is calculated, which is then used to deduct the instantaneous precipitation estimated by low orbiter satellite microwave observations. Finally, the continuous precipitation distribution is obtained. There are two versions of CMORPH: the original version (CMOPRH V0.x) and the new version (CMOPRH V1.0, used here). CMORPH V1.0 has provided near-real-time and bias-corrected products named CMORPH-RAW and CMORPH-CRT, respectively. CMORPH-RAW is satellite only precipitation and is provided in near-real-time so is suitable for real-time application. CMORPH-CRT is adjusted through matching the PDF of daily CMORPH-RAW against that for the CPC unified daily gauge analysis at each month over land, thus the CMORPH-CRT data is provided months delay. In this study, CMORPH-RAW with a temporal resolution of 30 min and a spatial resolution of −8 km, covering 60◦ S–60◦ N is used.

Based on the hourly precipitation observed by more than 30,000 automatic weather stations (AWS) in China and the satellite precipitation data retrieved from CMORPH, a merged rainfall product is developed through a two-step merging algorithm of PDF (Probability Density Function) and OI (Optimal Interpolation). The merged rainfall combines the advantages of both the AWS and satellite products, so the spatio-temporal distributions of precipitation is more accurate and reasonable. The gauge stations used in this data are far more than which used in CMORPH-CRT, and it updates more rapidly than CMORPH-CRT, so that we use merged data in China rather than CMORPH-CRT. The spatial resolution of the merged rainfall is 0.1◦ × 0.1◦, with a temporal resolution of 1 h. The dataset is available from 1 January 2008 to date, which can be downloaded online (available at http://data.cma.cn/data/detail/dataCode/SEVP\_CLI\_CHN\_MERGE\_CMP\_PRE\_HOUR\_GRID\_0.10/).

The CN05.1 dataset has been developed with observations at more than 2400 stations in China [29]. It is constructed by the "anomaly approach" [34], that is, first calculating a gridded climatology and then adding a gridded daily anomaly to the climatology to obtain the final dataset. The CN05.1 includes daily and monthly precipitation data and the period is from 1961 to 2017, with a spatial resolution of 0.25◦ × 0.25◦. Here, the monthly data in 1981–2010 is adopted to calculate the MAP for 30 years.

As is shown in Figure 2, the MAP in China ranges from 24.62 to 2315.27 mm, which is high in the southeast of the Yangtze River and gradually decreases from the southeast to northwest, as it is known that heavy rainfall occurs frequently in the southern China during the summer monsoon season. Furthermore, typhoons from the Northwest Pacific Ocean land in these areas frequently from June to October and bring abundant rainfall. The high MAP in the southeastern China could partly explain why this region with low topographic relief is prone to landslides. As the MAP varies greatly in China, and the climate affects the meteorological conditions that can result in landslides, there is a need to reduce the effects of climate diversity. Commonly, due to the differences of rainfall thresholds in different climatic regions, the MAP as a climatic index is selected to normalize the rainfall thresholds [20,24,35,36]. Therefore, the accumulated rainfall is divided by MAP in this paper to acquire the EMAP–D thresholds.

**Figure 2.** Mean annual precipitation (MAP) distribution in China during the period of 1981–2010.
