Changes of Precipitation and Extremes and the Possible Effect of Urbanization in the Beijing Metropolitan Region during 1960-2012 Based on Homogenized Observations

Received 15 November 2014; revised 16 January 2015, accepted 22 January 2015

Precipitation is an important climatic element, essential for studying water resources and the water cycle, and for quantifying climate change. However, long-term climate data series are usually interrupted by various non-natural factors, such as changes to observation locations, the environment, instrumentation, observation practices, or the algorithms used for calculating any particular climate variable (Manton et al., 2001; Tuomenvirta, 2001; Yan and Jones, 2008). The inhomogeneity caused by these factors could severely distort the true climate signal and introduce considerable bias to estimates of climate trends, variability, and extremes. As a result, various methods have been introduced for detecting and adjusting inhomogeneity in different climatic variables (Potter, 1981; Easterling and Peterson, 1995; Lund and Reeves, 2002; Li et al., 2014).

To date, several helpful temperature series homogenization techniques have been developed for many regions (Vincent et al., 2002; Trewin, 2013). However, the homogenization of precipitation series is rarely performed well, possibly because: (1) daily precipitation is a discrete variable with some non-normal distribution and stronger temporal and spatial variability; (2) the causes of inhomogeneity in precipitation series are more complex; and (3) the density of observation sites for precipitation is relatively sparse (Li et al., 2008b). Despite this, some meaningful studies have been conducted (Alexandersson, 1986; Rhoades and Salinger, 1993; Gonz\'alez-Rouco et al., 2001; Beaulieu et al., 2008, 2009; Wang et al., 2010). (Tuomenvirta, 2001) suggested that the precipitation adjustments could be 40% for 55 annual precipitation series in the North Atlantic region during 1890-1990, which were systematically biased by instrumental changes. (Begert et al., 2005) found that the adjustment factors ranged from 0.5 to 1.6 for 12 precipitation series in Switzerland from 1864 to 2000, and the introduction of automatic measuring equipment led to systematically lower measurements (approximately 5% less than the mean level). (McRoberts and Nielsen-Gammon, 2011) found that the remaining inhomogeneity from changes in gauge technology and station location may be responsible for an artificial trend of 1% to 3% per century, based on full network estimated precipitation data. (Wang et al., 2010) developed a new algorithm ("transPMFred") to detect shifts in nonzero daily precipitation series recorded at some stations across Canada and used a quantile-matching algorithm to adjust shifts in nonzero daily precipitation series.

For China, although the importance of homogeneous climate data has been considered in earlier works (Tao et al., 1991; Yan et al., 2001), most previous climate change analyses associated with precipitation have been carried out based on unadjusted data. In recent years, some authors have pointed out the inhomogeneity in precipitation series (Liu and Sun, 1995; Jiang et al., 2008; Li et al., 2008b). (Li et al., 2012) applied the standard normal homogeneity test method to monthly precipitation series at 110 stations throughout China during 1900-2009 and ultimately completed two gridded datasets (5^{°× 5°}, 2^{°× 2°}) of homogenized monthly precipitation. These homogenized datasets may be more reliable than the originals for climate change analyses. However, there is still potential for homogenization of daily precipitation series, which is important for evaluating changes of climate extremes. Daily meteorological observations have been used to assess changes in climate extremes or unusually anomalous weather fluctuations (Jones et al., 1999). As a result, a few homogenization methods, such as RHtests (Wang et al., 2010), have been developed for daily precipitation records.

We have previously developed homogenized climate datasets based on multiple analysis of series for homogenization (MASH), including daily temperature (Li and Yan, 2010) and wind speed (Li et al., 2011a) series in the Beijing metropolitan region (BMR), and daily maximum/mean/ minimum temperature series over China (Li and Yan, 2009). These results demonstrated MASH as a feasible technique for the homogenization of a large number of stations, especially in cases without complete metadata, such as in China. However, a problem arises from MASH where it is essential to adjust biases in the mean levels of the time series. To homogenize daily precipitation series, we developed an additional adjustment method to the relevant cumulative density function (CDF) dealing with climate extremes, which has been used in bias correction of model output.

In the present study, we establish a homogenized daily precipitation dataset at 15 stations in the BMR for the period 1960-2012 based on a combination of the MASH and empirical cumulative density function (ECDF) method, quantifying any secular trends in precipitation series during the past 53 years based on homogenized observations. We subsequently identify the possible influence of urbanization on the amount, frequency, and intensity of summer precipitation and extremes, and their trends. The rest of this paper is organized as follows: The daily precipitation data and methods are described in section 2. The results, including the homogenization of daily precipitation series, changes in annual and seasonal precipitation, and summer precipitation extremes are illustrated, and the possible influences of urbanization in the BMR analyzed, in section 3. Further discussion and conclusions are presented in section 4.

Observed daily precipitation records from 20 meteorological stations in the BMR were collected and processed by the Information Center of the Beijing Meteorological Bureau. The observed values were quality controlled using some conventional procedures listed by (Xu et al., 2013). Fifteen stations (Beijing-BJ, Shangdianzi-SDZ, Miyun-MY, Huairou-HR, Pinggu-PG, Changping-CP, Yanqing-YQ, Shunyi-SY, Chaoyang-CY, Tongxian-TX, Xiayunling-XYL, Daxing-DX, Fengtai-FT, Mentougou-MTG, Fangshan-FS) with few missing codes (a total of 184 days) were chosen, covering the study period of 1960-2012. Figure 1c shows the topography and geographical distribution of the 15 stations over the BMR. The metadata of relocation records at each station were also collected. All stations, apart from XYL, have relocated at least once and there were a total of 43 relocation records. There have been five relocations for BJ since 1960. The station was moved from Five Tower Temple to Daxing County on 1 January 1965, then to Zhanghua in Haidian County on 1 January 1969, to Daxing County on 1 July 1970, to Beiwa Road in Haidian District, a location closer to the city center on 1 January 1981, and back to the same suburban location in Daxing Country on 1 July 1997, with 38.8 km, 44 km, 24 km, 22 km, and 22 km of horizontal moving distance, respectively.

Figure 1.  (a, b) A square area (34-44锟斤拷N, 110-120锟斤拷E, blue shading) in China is marked for a better orientation to show the location of the Beijing metropolitan area (BMR, green shading) and Beijng station (BJ, red dot). (c) Topography (unit: m) and geographical distribution of the 15 stations over the BMR used in the study: Beijing-BJ; Shangdianzi-SDZ; Miyun-MY; Huairou-HR; Pinggu-PG; Changping-CP; Yanqing-YQ; Shunyi-SY; Chaoyang-CY; Tongxian-TX; Xiayunling-XYL; Daxing-DX; Fengtai-FT; Mentougou- MTG; Fangshan-FS.

2.2.1. Multiple Analysis of Series for Homogenization method

The latest version (v3.03) of the MASH package (Szentimrey, 1999) was used to detect and adjust possible break points due to relocations and other non-natural factors and homogenize the daily precipitation series in the BMR. Here we outline some of the basic ideas of MASH: (1) it is a relative homogeneity test procedure and does not assume that the reference series are homogeneous; (2) it includes a step-by-step iteration procedure where the role of series (candidate, reference) changes step by step in the course of the procedure; (3) additive (e.g., for temperature) or multiplicative (e.g., for precipitation) models can be used depending on the climatic elements; (4) it includes quality control and missing data completion procedures for monthly, seasonal, annual and daily data; (5) in the case of having monthly series for all the 12 months, the monthly, seasonal, and annual series can be homogenized; (6) the daily inhomogeneity can be derived from the monthly inhomogeneity; and (7) metadata (possible dates of break points) can be used automatically.

In the present study, the multiplicative model was applied to daily precipitation data underlying a quasi-lognormal distribution. A strict significance level for the test statistic (Monte Carlo method) was given at 0.01 to maintain the information in the original dataset as much as possible. The reference system (nine reference stations) of each candidate station was created based on their distances and correlation coefficients to the candidate station. The metadata of documented relocation records were used.

2.2.2. Empirical Cumulative Density Function method

The distribution-based homogenization method ECDF was introduced to the study. The idea of this method originates from bias correction techniques used in output adjustment of global or regional climate models for hydrological application or climate impact assessment (Piani et al., 2010; Yang et al., 2010; Dosio and Paruolo, 2011). The basic concept of the method assumes that the series before and after the break point follows the same probability distribution. However, a non-parametric method, different from previous parametric methods cited above, was adopted to avoid uncertainties in the adjustment dealing with extremes, i.e., only through the original ECDF, instead of parameter estimation by a certain probability distribution (e.g., gamma).

There are two circumstances in which MASH-based homogenized series can be additionally adjusted based on ECDF. If the break points detected by MASH agree with relocation records in each season, the MASH-based adjusted daily series will be additionally adjusted via ECDF, and the final adjusted daily series based on MASH+ECDF are obtained. In contrast, if no break point is detected, or the break points cannot be verified by metadata, the MASH-based series remain unchanged.

The ECDF method includes the following steps:

(1) Define the candidate sub-periods (CSPs) divided by break points, which are detected by MASH and consistent with metadata. CSPs are numbered starting from the most recent sub-period.

(2) One or more reference station(s) is (are) chosen, if they have high correlation with the candidate station on the daily scale and no documented break points during the 10-year window centered on the relocation date of the candidate station, as suggested by (Yan et al., 2010). Accordingly, the 10-year window is defined as the reference sub-periods (RSPs). RSPs are the same as CSPs, starting from the most recent sub-period.

(3) Start the adjustment with the most recent inhomogeneity. The differences between the ECDFs of the reference and candidate series after the break point in RSP 1 are calculated after they are interpolated into the same intervals (1000 intervals between 0 and 1). If there is more than one reference station, their records will be combined together as one series.

(4) The differences before the break point should be the same, if the candidate series is homogeneous. Based on this assumption, the differences are interpolated into the ECDF of the reference series in the RSP 2 before the break point and then interpolated back into the ECDF of the candidate series in the CSP 2, and the final adjustment values are obtained.

(5) CSP 2 is homogeneous with respect to the latest sub-period (CSP 1) after subtracting the adjustment values. Repeat the process for all other CSPs sequentially.

(6) Note that the adjustments are implemented separately for spring (March, April and May), summer (June, July and August; JJA), autumn (September, October and November; SON), and winter (December, January and February).

To classify urban and rural sites regarding variability of summer precipitation, an empirical orthogonal function (EOF) analysis (Deng et al., 1989) was carried out on the standardized anomaly series of summer precipitation at the 15 stations. The linear trend and correlation were assessed to express long-term changes in each original and adjusted series during 1960-2012. A trend was considered statistically significant at the 95% confidence level and the slopes of trends were calculated by least-squares linear fitting.

Figure 2.  Schematic diagram of candidate sub-periods (CSPs) at Beijing, the reference stations, and reference sub-periods (RSPs) for each break point in the (a) summer and (b) autumn precipitation series. The number in brackets is the correlation coefficient of daily precipitation series from 1960 to 2012 between Beijing and the reference station.

Figure 3.  Cumulative density function of (a) daily precipitation and (b) daily precipitation >70 mm for the original (black line) and final adjusted (red line) series during 1960-2012 at Beijing.

First, the MASH method was applied to detect and adjust the inhomogeneity in daily precipitation series at the 15 stations in the BMR. The MASH method detected some documented changes associated with relocations. For annual series, the estimated inhomogeneity (ratio of original/adjusted) among the 15 stations varied between 0.92 and 1.15. Eight stations (53.3% of the total number) were inhomogeneous, with a total of 17 breakpoints and seven explained by relocation records. For the seasons, four, one, eight, and nine precipitation series were inhomogeneous and the estimated inhomogeneity ranged from 0.45 to 1.46, 1 to 1.18, 0.87 to 2.92, and 0.38 to 2.59 in spring, summer, autumn and winter, respectively. The number of break points in each season was different and 24.59 % of them (of a total 51) were caused by relocations. For the BJ station, there were seven, zero, one, three, and four break points in the annual and seasonal series (Table 1). On the seasonal scale, only the break points in the summer of 1997 and the autumns of 1965 and 1969 coincided with relocation records. As a result, relocation was not a major reason for inhomogeneity in the precipitation series. The remaining break points may be due to factors including the new type of screens and instruments, introduction of automatic measurement equipment, and changes in observation times.

Next, the ECDF method was used to additionally adjust the daily MASH-based precipitation series. To demonstrate how ECDF works in homogenization, the BJ station is selected as an example. Figure 2 shows the CSPs of summer and autumn precipitation series at BJ, and the corresponding ASPs of reference stations, according to the definition given in section 2.2.2. For the summer series, CSP 1 (CSP 2) was July 1997 and JJA 1998-2012 (JJA 1960-96 and June 1997); the reference stations were FT (0.85), CY (0.84), and DX (0.82); and ASP 1 (ASP 2) was July and August 1997, JJA 1998-2001, and July 2002 (July and August 1992, JJA 1993-96, and July 1997). For the autumn series, the CSPs were SON 1969-2012, 1965-68, and 1960-64. For the break point in 1969/65, the reference stations were FT and CY/CP (0.68), and the corresponding RSPs were SON 1969-72 and 1965-68/SON 1965-69 and 1960-64, respectively.

Figure 3 provides the CDFs of the original and final precipitation and precipitation >78 mm during 1960-2012 at BJ. It shows that the CDFs of daily precipitation ≤ 0.8 mm were almost identical in the original and adjusted daily precipitation series. The CDF of daily precipitation between 0.8 and 100.7 mm in the adjusted series is greater than that in the original series. However, the CDF of daily precipitation >100.7 mm in the adjusted series is less than that in the original series, indicating that the probability of extreme precipitation is adjusted towards the higher. For example, the CDF at 110 mm is 99.95% and 99.87% in the original and adjusted series, respectively.

Figure 4.  Original and adjusted seasonal and annual precipitation (unit: mm) and summer precipitation extremes (unit: d) during 1960-2012 at Beijing. The corresponding moving t-test with the 0.05 significance threshold (u1, u2) of the adjusted series is also given. A maximum t-value of above the u2 threshold indicates a decreasing jump.

To further understand the impact of inhomogeneity on trend estimation, we compared the annual/seasonal precipitation and summer precipitation extreme series (Fig. 4) and their trends between the original and adjusted series (Table 2) at BJ. Here, the occurrence of an extreme precipitation (day) is recorded if daily precipitation exceeds the 90th percentile of the daily precipitation distribution during 1960-2012 (Jones et al., 1999; Yan et al., 2002). The spring series was homogeneous with increasing trends of approximately 4.82 mm (10 yr)^-1 in the original and adjusted series (Fig. 4a). For the summer series, the precipitation was enhanced in some years during the early period after adjustment. The decreasing linear trend was slightly amplified from -23.22 to -24.01 mm (10 yr)^-1 due to some higher extreme values in the early period after adjustment (Fig. 4b). For the autumn series, the increasing trend was adjusted by 0.07 mm (10 yr)-1, which was less than the estimate of 5.84 mm (10 yr)-1 in the original series (Fig. 4c). In winter, both series show the same trends of approximately -0.23 mm (10 yr)-1, although there are some slight adjustments in the early period (Fig. 4d). The adjusted annual series indicates a stronger decreasing trend of -13.66 mm (10 yr)^-1 than the original series [-12.79 mm (10 yr)-1] (Fig. 4e). The frequencies of summer precipitation extremes have slight differences of 1 day between the original and adjusted series (Fig. 4f). The decreasing trend is -0.25 d (10 yr)^-1 after adjustment, more prominent than the estimate of -0.19 d (10 yr)^-1 in the original series. A moving two-sided t-test (n1=n2=10, degrees of freedom = 18) of the adjusted series indicates a sharp decrease (significant at α=0.05) in the mid-1990s in the annual and summer series, and the summer precipitation extremes series; and a rapid increase in the mid-1970s in the winter series.

Figure 5.  Geographical distribution of linear trends [units: mm (10 yr)^-1, right panels] and regional mean trends [units: mm (10 yr)^-1, left panels, left coordinate] in seasonal and annual precipitation and summer precipitation extremes and corresponding moving t-test curves with the 0.05 significance thresholds (u1, u2, left panels, right coordinate) during 1960-2012 in the Beijing metropolitan region. The solid blue/red circles indicate significant decreasing/increasing trends at the 0.05 significance level. The open blue/red circles indicate decreasing/increasing trends. A maximum t-value of above the u2 threshold indicates a decreasing jump. The 15 districts and counties are also indicated on the map.

Figure 6.  (Continued)

Based on the final adjusted dataset, we calculated the trends in annual and seasonal precipitation and summer precipitation extremes averaged over the BMR. After adjustment, the spring and autumn precipitation series exhibited increasing trends of 4.97 and 5.02 mm (10 yr)-1, which were slightly less than the original series by -0.20 and -0.05 mm (10 yr)-1, respectively. However, the adjusted summer, winter, and annual series showed decreasing trends of -20.04, -0.11 and -10.16 mm (10 yr)-1, which were more prominent than the original values by -0.05, -0.03 and -0.33 mm (10 yr)-1, respectively. For the summer and annual precipitation, the inhomogeneity may be responsible for an artificial trend of 0.3% (10 yr)^-1 and 3.3% (10 yr)^-1, respectively. The trend of the regional mean summer precipitation extremes was approximately -0.30 d (10 yr)^-1 in the original and adjusted series. The influence of homogenization on estimating the regional mean precipitation trend was not large, partly because some stations were not influenced by relocation and partly because effects of local inhomogeneity compensated each other to some extent. The moving two-sided t-tests indicated that there were more pronounced increasing trends in the annual and summer precipitation and summer precipitation extreme series during 1999-2012, fluctuating with linear variations of 173.1, 127.8 and 3.2 d (10 yr)^-1, respectively (Fig. 5, left panels).

The corresponding geographical patterns of annual and seasonal precipitation trends are also given (Fig. 5, right panels). In spring, all 15 stations were found to have experienced increasing trends, especially in the northern part, indicating a southwest-northeast gradient. For TX, the increasing trend was reduced from 4.52 mm (10 yr)^-1 to 1.12 mm (10 yr)-1 after adjustment, with the largest adjustment amplitude among all stations (Fig. 5g, right panels). For the summer series, the pattern demonstrates the southeast-northwest gradient, with the lowest value of -26.23 mm (10 yr)-1 at YQ and the highest value of -11.52 mm (10 yr)-1 at CY, which is within the Central Business District (CBD) of Beijing, Chaoyang District, with the highest urbanization level among all districts and counties (Zhang, 2006). Away from the CBD site, the decreasing trends become stronger, with larger values in the rural area, implying an underlying effect of urbanization in Beijing (Fig. 5h, right panel). In autumn, the increasing trends vary from 1.65 mm (10 yr)^-1 to 9.19 mm (10 yr)-1; greater in the southeast (highest at DX) and less in the west (lowest at MTG) (Fig. 5i, right panel). In winter, the magnitudes of the trends are small, varying between -0.49 and 0.35 mm (10 yr)^-1, showing a northern-increase-southern-decrease dipole pattern (Fig. 5j, right panel). The annual geographical pattern is similar to the summer, because the precipitation in summer plays a dominant role in the overall annual precipitation amount. The adjusted annual data exhibit decreasing trends ranging from -16.22 to -0.73 mm (10 yr)^-1, presenting a southeast-northwest gradient (Fig. 5k, right panel). The magnitude of the decreasing trend is again lowest at CY and highest at YQ. For the summer precipitation extreme (Fig. 5i, right panel), the spatial pattern indicates a northwest-southeast gradient. The trends are negative at 13 stations, especially at YQ (significant decreasing), ranging from -0.6 to -0.2 d (10 yr)-1; whereas, they are positive by approximately 0.1 and 0.2 d (10 yr)-1 at CY and TX, respectively, with the highest urbanization levels and fastest urbanization rates over the past 10 years. This pattern suggests a possible influence of urbanization.

It is well known that the trend and pattern of regional precipitation is influenced by many factors, including global and regional water cycles, climate variation, topography, urbanization, and land use/cover change (Song et al., 2014). (Hand and Shepherd, 2009) found that the north-northeastern regions of the metropolitan Oklahoma City were wetter than other regions, and precipitation modification by the urban environment may be more significant than agricultural/topographic influences on weakly forced days. Over India, there is a significant increasing trend in the frequency of heavy rainfall climatology over urban regions during the monsoon season (Kishtawal et al., 2010). Over the Pearl River Delta metropolitan regions of China, the urbanization signatures in strong precipitation are significantly different from those in weak precipitation over the urban areas (Li et al., 2011b). For the BMR, the long-term decreasing trend in precipitation is consistent with that in northern China and mainly attributed to changes in regional atmospheric circulation (Xu et al., 2006; Wang et al., 2008; Cong et al., 2010). However, the trend and pattern of precipitation and extremes may be influenced by urbanization (Sun and Shu, 2007; Li et al., 2008a; Zhang et al., 2009a, b; Miao et al., 2011; Yang et al., 2013, 2014; You et al., 2014; Zhang et al., 2014). Here, we take summer precipitation as an example to quantify the possible influences of urbanization on the amount, frequency, and intensity of precipitation and extremes. A relevant study for low-temperature precipitation in the cold season can be found in (Han et al., 2014).

Figure 7.  First two empirical orthogonal functions (EOFs) of the standardized anomalous summer precipitation amount in the Beijing metropolitan region during 1960-2012: (a, c) first EOF mode and time coefficient series; (b, d) second EOF mode and time coefficient series.

Figure 8.  Geographical distribution of (a) amount (unit: mm), (b) frequency (unit: d), and (c) intensity (units: mm d^-1) of summer precipitation; and (d) amount (unit: mm), (e) frequency (unit: d), and (f) intensity (units: mm d^-1) of summer precipitation extremes in the Beijing metropolitan region during 1960-2012.

An EOF analysis was applied to the standardized anomaly time series of summer precipitation at the 15 stations during 1960-2012 to help identify whether there are different regional regimes of climate variability in summer precipitation in association with urban or rural divisions in the BMR. Figure 6 shows the first two EOF patterns and time coefficient series. The first EOF mode explains as much as 70.14% of the total variance, showing a coherent phase (negative coefficients) over the whole region, and indicating very similar inter-annual variation and consistent decreasing trends of summer precipitation among all stations in the BMR (Figs. 6a and c). However, the coefficients are larger in the southeast BMR, where is more urbanized and has a greater urbanization rate than the surrounding areas; and smaller in the northwest, where there is a mountainous area, suggesting a possible combined influence of urbanization and topography on the summer precipitation pattern. Based on the EOF 1 mode, four stations (BJ, DX, CY, and FT) are classified as the urban group, and the two stations (MTG, FS) away from BJ (in the upwind of the urban area) and with similar elevations to BJ are defined as the rural group. The second EOF mode explains 7.17% of the total variance and there is slight linear decreasing trend in the time coefficient series. The pattern clearly demonstrates a west-east gradient of summer precipitation across the BMR, suggesting the possible influence of large-scale factors, such as mean sea level pressure (Figs. 6b and d).

Figure 7 shows the geographical distribution of the annual mean amount, frequency, and intensity of summer precipitation, as well as those of extremes. To further investigate the differences in these six indices between the urban and rural regimes, and quantify the possible influence of urbanization, we calculated the mean time series for the four urban sites, as classified by the first EOF mode, and the two rural stations. Figure 8 shows the amount, frequency and intensity series of summer precipitation and extremes in urban and rural areas during 1960-2012. Table 3 lists the trends derived from the series in Fig. 8 and the corresponding possible contributions of urbanization, which are defined as

\begin{eqnarray*} \dfrac{T_{\rm ur}-T_{\rm ru}}{T_{\rm ur}}\times 100% , \end{eqnarray*}

where T _ur and T _ru are the trends of the urban and rural series during 1960-2012, respectively.

Figure 9.  (a) Amount (unit: mm), frequency (unit: d) and intensity (units: mm d^-1) series of summer precipitation; and (d) amount (unit: mm), (e) frequency (unit: d), and (f) intensity (units: mm d^-1) series of summer extreme precipitation in urban and rural regions in the Beijing metropolitan region as outlined by the first empirical orthogonal function mode in Fig. 6a during 1960-2012.

Summer precipitation in the BMR usually occurs when a westerly trough approaches, and is attributed to water vapor transportation through the southwesterly jet in the mid and lower troposphere and/or to the easterly wind at the bottom of the troposphere. Figures 7a and d show that the higher precipitation occurs at the interface between the mountainous area and the plain areas (mainly on the windward side, especially in the southwest and northeast) and downwind of the urban area (especially in the northeast), while lower values occur in the urban core region. This confirms the combined influence of topography and urbanization on the pattern of precipitation amount, as well as extreme precipitation amount, in summer. This result is consistent with the work of (Yin et al., 2011). The annual summer precipitation amount/extreme amount is 425.46/322.83 mm in the urban region, approximately 4.2%/3.0% higher than that in the rural area. The decreasing trends of summer precipitation and extreme amounts are -17.94 and -16.50 mm (10 yr)^-1, respectively, which are approximately 10.4%/6.0% less pronounced than those in the rural area (Figs. 8a and d; Table 3). This suggests that urbanization tends to slightly enhance summer precipitation and extreme amounts, and reduce the magnitude of the decreasing trend in the urban area.

The geographical pattern of precipitation frequencies indicates a southeast-northwest gradient. The lowest values occur mainly in the urban core area and the southeast part of the city (Fig. 7b). This pattern may be shaped by topography and urbanization; while the frequencies of extremes are greater in the urban region, especially at CY, than those in the surrounding area. This shows that the extremes may be more likely to occur in the urban area (Fig. 7e). The frequencies of precipitation in the rural area are greater than those in the urban area, but the reverse of the extremes in certain years (Figs. 8b and e). The frequencies of precipitation and extremes exhibit declining trends of -0.57 and -0.16 d (10 yr)-1, respectively, in the urban area, which are slightly more and less prominent than the rural area by -0.03 and 0.02 d (10 yr)^-1, respectively. Correspondingly, the influence of urbanization is estimated to be 5.7% and -8.9%, respectively, showing that urbanization may reduce the magnitude of the decreasing trend of precipitation frequency, but enhance that of extremes in the urban area.

For intensity, the patterns of summer precipitation and extremes are similar to those of precipitation amount. The highest intensities are located in the northeast part (downwind of the urban area) (Figs. 7c and f). The intensities of precipitation and extremes exhibit decreasing trends at each site, with a lower center in the core and eastern part of the city. For extreme intensity, the decreasing trend is significant at the 0.05 significance level at urban sites [-1.78 mm d-1 (10 yr)-1], while insignificant in rural sites [-1.17 mm d-1 (10 yr)-1]. The differences in the trends [-0.02 and -0.61 mm d-1 (10 yr)-1] between the rural and urban categories were used to estimate the influence of urbanization on the intensity of precipitation and extremes, accounting for around 6.8% and 51.5%, respectively, of the decreasing trend of the rural category, which may represent large-scale change. These results suggest that urbanization tends to promote a decrease in intensity, especially of extremes.

A dataset of daily precipitation observations in the BMR during 1960-2012 was homogenized based on MASH and additionally adjusted by the ECDF method regarding climate extremes. The new daily precipitation dataset provides a more reasonable basis for quantifying precipitation climate changes and the effect of urbanization on summer precipitation and extremes. The major conclusions from the present analysis are summarized with discussion as follows.

Some break points associated with relocation records can be detected by MASH. However, the numbers are different in each season and significantly less than those in the temperature and wind speed series shown in our previous studies (Li and Yan, 2010; Li et al., 2011a), suggesting that the precipitation series is less significantly influenced by the unnatural changes in the observation systems.

The combination of the MASH and ECDF methods provides a new way to homogenize precipitation series on the daily scale from the perspective of both mean level and higher-order moment and shows the potential for improving and quantifying estimates of climate changes dealing with extreme precipitation and events, especially on the local scale. For BJ, the probability of extreme precipitation (>100.7 mm) increased after adjustment and the decreasing trend of summer precipitation extremes was amplified from -0.19 to -0.25 d (10 yr)^-1.

Inhomogeneity has little influence on the regional average long-term climate trends. When averaged over the BMR, the adjusted precipitation data exhibit an increasing trend of 4.82 and 5.77 mm (10 yr)-1 for spring and autumn, respectively, and a decreasing trend of -24.01, -0.23 and -13.66 mm (10 yr)-1 for summer, winter, and the whole year during 1960-2012, respectively. The summer extreme precipitation event series exhibits a declining trend of approximately -0.30 d (10 yr)^-1. The increasing climatic trends since 1999 are notable, both in the summer/annual precipitation and summer precipitation extreme series.

The geographic patterns of trends in precipitation for the whole year and the four seasons are different. In spring and autumn, precipitation of all the 15 stations shows a consistently increasing trend, with the highest values in the northern and southeast part of the BMR. For summer and the whole year, the patterns of decreasing trends demonstrate a southeast-northwest gradient, with the lowest value in the northwest and the highest value in the core urban area, implying an underlying effect of urbanization in Beijing. However, there are decreasing trends in the southern and increasing trends in the northern part for winter, revealing a northern-increase-southern-decrease pattern. For summer precipitation extremes, there are greater magnitudes of decreasing trends at the rural sites and smaller magnitudes of decreasing trends or even positive trends at the urban sites.

Urbanization tends to slightly enhance the amount of summer precipitation and extremes, and reduce their magnitude of decrease in urban-influenced regions. Quantitatively, the influence of urbanization on the amount of summer precipitation and extremes during 1960-2012 was estimated to be approximately -2.09 and -1.04 mm (10 yr)^-1, accounting for 10.4% and 6.0%, respectively, of the large-scale mean decreasing trends. For the frequencies of precipitation and extremes, the influence of urbanization was estimated to be 5.7% and -8.9%, respectively, suggesting that urbanization may reduce the magnitude of decrease of precipitation frequency, but enhance that of extremes in the urban areas. For precipitation intensity, urbanization tends to promote the decreasing trends in both summer precipitation and extremes, by approximately 6.8% and 51.5%, respectively.

It is worth noting that the homogenization of daily precipitation series is a challenging topic, as daily precipitation is not a continuous variable and has larger variability. Hence, there is greater uncertainty in the homogenization of precipitation than that of other climatic variables, such as temperature. Nevertheless, the additional adjustments based on the ECDF method provide an experimental approach to improving the homogeneity of daily precipitation series in terms of probability distribution, showing potential for further application in homogenizing daily climate series.