*2.2. Data Sources*

Daily precipitation data, covering 1960 to 2012, from the 13 standard meteorological stations (Table 1) in the WRB were obtained from the China Meteorological Administration (http://cdc.cma.gov.cn, accessed on 22 August 2015). These meteorological stations are maintained according to the standard methods of the National Meteorological Administration of China. For the same period, daily streamflow data from the Huaxian gauge (Figure 1) was obtained from the Hydrological Yearbooks of China (http://loess.geodata.cn, accessed on 10 May 2016). All meteorological and hydrological data used in this study have been submitted to quality control by government agencies before release.


**Table 1.** Meteorological stations used in this study.

#### **3. Methods**

#### *3.1. Baseflow Separation Algorithm*

To improve the accuracy of baseflow estimates in this study, revised and validated baseflow separation was implemented [40]. Baseflow has a lag time concerning the last precipitation event [41]. Generally, the baseflow recession is linked with the surface and sub-surface flow characteristics and follows an exponential decay curve [42]:

$$Q\_b = Q\_0 \alpha^t \tag{1}$$

where *Qb* is the baseflow at time *t*, and *α* is the recession constant determined by recession analysis. The baseflow can be calculated using the baseflow separation method.

Baseflow separation is a fundamental issue that has been comprehensively documented [8,43,44]. Several algorithms have been proposed to separate baseflow from total observed streamflow [45–47] and can be classified as trace-based, water balance, and graph approaches according to general applications. Digital filters are the most widely used tools for small-data input and is reducible (e.g., only daily streamflow records and more objective) [40]. The Lyne–Hollick method was used here, expressed as [48]:

$$Q\_{q\_i(i)} = \mathfrak{a}Q\_{q\_i(i-1)} + \frac{1+\mathfrak{a}}{2}(Q\_i - Q\_{i-1}) \tag{2}$$

where *Q* is total streamflow (m3/d), *Qq* is quick flow (mm/d), *i* is the time step (day), and *α* is the filter parameter (recession constant, in 1/day). Baseflow (*Qb*, m3/d) can subsequently be calculated as *Qi* minus *Qq*. The baseflow index (BFI, calculated as *total Qb/total Q*), is a standard indicator of the baseflow contribution to total streamflow. Herein, the calibrated Lyne–Hollick method was employed to separate the long-term baseflow. This approach has been validated by Zhang, et al. [40].

The recession constant can be obtained using the recession analysis developed by Brutsaert, et al. [43]. This recession approach efficiently reduces uncertainties when estimating the initial points in the recession limb. Details of recession analysis are given in Cheng, et al. [44].
