**1. Introduction**

Landslides are one of the most devastating geo-hazards that cause heavy casualties and great economic losses around the world every year. China is one of the countries that are most prone to landslides due to its geological, geomorphological and climate conditions [1,2]. According to the National Geological Disaster Bulletin issued by the Ministry of Natural Resources of the People's Republic of China from 2007 to 2016, an average of 762 people are reported dead or missing per year due to landslides, debris flows and other geological disasters, resulting in direct economic losses of 4.2 billion Chinese yuan. Since precipitation is the main trigger factor for landslides, the frequency and intensity changes of precipitation, which are caused by the global warming [3], will probably lead to an increase in the potential occurrence of landslides in the future [4–6]. In addition, as China is currently experiencing economic booming and developing rapidly, the early warning of landslides is crucial in China.

Rainfall thresholds can be used to predict the possible occurrence of slope failures in an area [7]. Currently, many countries and regions have developed the early warning systems (EWSs) based on rainfall thresholds to provide geological hazard information for the public [8–10].

By using the global rainfall threshold, combined with the landslide and debris flow susceptibility in China, a real-time EWS has been established for landslides and debris flows in China [11]. The EWS was verified by using 106 events in China during 2016–2017, and the results showed that the early or delayed warning is effective for 69% of the total events, while 72% of landslide events in the rainy season (May to September) can be warned, but only 35% of landslide events in the non-rainy season can be warned in advance. Some landslide events which did not reach the critical value of rainfall triggering the landslide and debris flow cannot be predicted, which may be due to the fact that these rainfall thresholds are not applied for China [12]. Thus, it is necessary to establish rainfall thresholds suitable for China to improve the positive prediction rate for landslides.

In fact, some rainfall thresholds have been developed in China, but most of them were established with few events or just suitable for specific areas. For example, Huang et al. (2015) have calculated the rainfall thresholds with 50 landslide events in Huangshan region of Anhui Province [13]. Zhou and Tang (2013) have defined a rainfall threshold with 11 rainfall events inducing debris flows in the Wenchuan earthquake-stricken area in Sichuan Province [14]. Li et al. (2017) have established a rainfall intensity–duration threshold in China, but only with 60 landslide events happening during 2005–2011 [15]. By far, few studies have used a large amount of historical data of landslides to establish rainfall thresholds in China.

Rainfall thresholds can be established by using both physical and empirical methods. Rainfall infiltrates into the slope and changes the pore water pressure, reducing the shear stress and thus may result in the slope failure [16]. Based on the physical methods, these physical processes can be analyzed and the critical value of rainfall can be calculated out [17,18]. However, such methods require an extensive data collection of geological information that requires high spatial resolution, so it's not feasible for threshold calculation at large regional or global scales [19]. Based on the rainfall information of historical landslide events, the empirical methods can determine the low boundary with statistical methods for the rainfall conditions that result in slope failures. Several statistical methods have been applied to obtain the rainfall thresholds, such as Bayesian statistics [20], logistical regression [21], confidence level [22] and quantile regression [23,24]. Thus, the empirical methods are more suitable for larger areas.

Guzzetti et al. (2008) have classified the rainfall thresholds into four categories: (i) rainfall intensity–duration (I–D) thresholds, (ii) thresholds based on the total rainfall of the events, (iii) rainfall event–duration (E–D) thresholds and (iv) rainfall event–intensity (E–I) thresholds [25]. The I–D and E–D thresholds are the most commonly, world-widely used among them. The E–D thresholds are equivalent to the I–D thresholds, while the E–D thresholds can avoid unnecessary conversions. Therefore, the E–D thresholds are selected in this paper.

This paper aims to calculate the E–D and normalized (the accumulated rainfall is normalized by mean annual precipitation, MAP) (EMAP–D) thresholds in the whole of China, and to validate the thresholds for early warning of landslides. A landslide inventory with more landslide events covering the whole region of China will be used. The rest of this paper is organized as follows. Section 2 describes the landslides inventory, rainfall dataset and quantile regression method to calculate the rainfall thresholds, and the thresholds established in this study are shown in Section 3. In Section 4, the thresholds defined in this paper are compared with the existed thresholds and validated with those from the previous analysis [11]. Finally, the main conclusions are summarized in Section 5.
