Regional and local increases in storm intensity in the San Francisco Bay Area, USA, between 1890 and 2010


  • Tess A. Russo,

    Corresponding author
    1. Department of Earth and Planetary Sciences, University of California, Santa Cruz, Santa Cruz, California, USA
    2. Columbia Water Center, Columbia University, New York, New York, USA
    • Corresponding author: T. A. Russo, ColumbiaWater Center, Columbia University, New York, New York, USA. (

    Search for more papers by this author
  • Andrew T. Fisher,

    1. Department of Earth and Planetary Sciences, University of California, Santa Cruz, Santa Cruz, California, USA
    Search for more papers by this author
  • Dustin M. Winslow

    1. Department of Earth and Planetary Sciences, University of California, Santa Cruz, Santa Cruz, California, USA
    Search for more papers by this author


[1] Studies of extreme precipitation have documented changes at the continental scale during the twentieth century, but few studies have quantified changes at small to regional spatial scales during the same time. We analyze historic data from over 600 precipitation stations in the San Francisco Bay Area (SFBA), California, to assess whether there have been statistically significant changes in extreme precipitation between 1890 and 2010. An annual exceedance probability analysis of extreme precipitation events in the SFBA, coupled with a Markov chain Monte Carlo algorithm, reveals an increase in the occurrence of large events. The depth-duration-frequency characteristics of maximum annual precipitation events having durations of 1 h to 60 days indicate on average an increase in storm intensity in the last 120 years, with the intensity of the largest (least frequent) events increasing the most. Mean annual precipitation (MAP) also increased during the study period, but the relative increase in extreme event intensity exceeds that of MAP, indicating that a greater fraction of precipitation fell during large events. Analysis of data from subareas within the SFBA region indicates considerable heterogeneity in the observed nonstationarity; for example, the 5 day, 25 year event exceedance depth changed by +26%, +16%, and −1% in San Francisco, Santa Rosa, and San Jose, respectively. These results emphasize the importance of analyzing local data for accurate risk assessment, emergency planning, resource management, and climate model calibration.

1 Introduction

[2] Climate change may increase the frequency and intensity of extreme weather events [Trenberth et al., 2007], but few studies have explored the nature of hydrologic change on local to regional scales during the last century. Evaluating temporal changes in precipitation requires examination of a range of spatial dimensions, from local studies that capture the scales of many extreme storms to regional and continental studies that quantify large-scale variations in evapotranspiration, moisture transport, and cloud formation [e.g., Booij, 2002; Smith et al., 2005]. Some recent research suggests that the regional signal of observable precipitation change will not exceed the natural variability until the late 21st century or beyond [Giorgi and Bi, 2009; Mahlstein et al., 2012]. However, significant local changes in precipitation quantity and patterns have occurred over the last 100 years in some areas [Tomozeiu et al., 2000; Douglas and Fairbank, 2011]. Local studies are particularly important for understanding the mechanisms leading to changes to precipitation patterns, which are needed for the creation and calibration of accurate regional climate models. Local studies are also essential for risk assessment, municipal planning, and resource management. Large-scale analyses may not be as directly relevant for regional policy makers and other stakeholders, although these analyses provide broad context and can indicate where evaluation of local data may be especially important.

[3] Several continental and country-scale precipitation studies of data from the twentieth century suggest that the largest storms increased in intensity at rates that exceeded those of annual or seasonal precipitation increases [Karl and Knight, 1998; Groisman et al., 1999, 2004, 2005; Fowler and Kilsby, 2003]. In some cases, annual to decadal variability in large-scale precipitation patterns correlates with global climate phenomena, including the Pacific Decadal Oscillation (PDO) and El Niño-Southern Oscillation cycles [Hanson et al., 2006; Higgins et al., 2007; Arriaga-Ramírez and Cavazos, 2010], higher atmospheric moisture content due to increasing temperatures [Trenberth et al., 2003; Trapp et al., 2007; Allan and Soden, 2008; Min et al., 2011], and the intensity of atmospheric rivers [Leung and Qian, 2009; Dettinger, 2011; Ralph and Dettinger, 2012]. Some assessments of the twentieth century North American observational record show no clear systematic changes in the intensity of large storms [Kunkel, 2003a, 2003b; Bonnin et al., 2011], but this does not preclude significant changes in hydrologic conditions and processes at local to regional scales. In aggregate, earlier studies suggest that there may have been significant changes in extreme precipitation in recent decades, but such changes are heterogeneously distributed in both space and time.

[4] Continental-scale studies of twentieth century precipitation that include the San Francisco Bay Area (SFBA), California, have generated disparate results. Some studies have shown an increase in the magnitude of extreme precipitation events across the western United States (including the SFBA) ranging between ~0.5% and 1.5%/decade [Karl and Knight, 1998; Kunkel et al., 1999; Mass et al., 2011]. Storm events in the western United States have been associated with the arrivals of atmospheric river events, which may be increasingly frequent by the end of the 21st century [Dettinger, 2011]. Other observational and modeling studies have documented a decrease in winter (wet season) precipitation across the same region [Weare and Du, 2008] or found no statistically significant changes [Dettinger, 2005; Snyder and Sloan, 2005; Barnett et al., 2008; Peterson et al., 2008]. Most studies on North America indicate that storm intensity has increased in the last century, although the frequency of storms may have decreased [Brommer et al., 2007] or increased [Karl and Knight, 1998].

[5] The present study was motivated by the wide range of results noted above and by the recognition that long (>100 years), high-quality observational data sets are now available for many areas of interest, including centers of industry, agriculture, and urban development such as the SFBA. Rantz [1971] presented the last comprehensive depth-duration-frequency (DDF) analysis of precipitation in the SFBA, using data collected from 1906 to 1956. These results have been used for decades for technical analysis, planning, and risk assessment [e.g., Tait and Revenaugh, 1998; Keefer, 2000; Crovelli and Coe, 2009], although the observational record [Rantz, 1971] used was relatively short and is now more than 50 years old.

[6] The present study addresses three main questions: (1) Was there a statistically significant change in the intensity of extreme precipitation in the SFBA during the past 120 years? (2) Did the magnitudes of extreme events change relative to changes in MAP during this time? (3) How do local changes in the magnitude of extreme events compare within the SFBA region? We do not assess the mechanism(s) responsible for changes in precipitation intensity over the last 120 years but focus instead on quantifying the magnitudes of changes apparent in the observational record. Additional work will be required to assess causality and mechanisms and to quantify the hydrologic implications of documented changes, as discussed later.

2 Data Sources

[7] Precipitation data were gathered by the California Department of Water Resources (CADWR) and include records from the California Data Exchange Center, California Irrigation Management Information Systems, Remote Automatic Weather Stations, and specifically the San Francisco Water Department (SFWD), Marin Municipal Water District (MMWD), San Jose Water Works, and Sonoma County Water Agency. Historical data were obtained from SFBA precipitation monitoring stations located within a region of 31,000 km2, extending from the Pacific Ocean to the California Central Valley (Figure 1a). Data were compiled at hourly and daily intervals using 336 and 679 stations, respectively. Thirty-one stations distributed across this region include >100 years of data (Figure 1a and Table 1).

Figure 1.

(a) Locations of precipitation stations in the SFBA used in this study. Size of the station marker indicates the record duration for that station, ranging from 15 to 120 years. Stations with >100 years of data are shaded in red. Total number of daily and hourly records shown as a (b) histogram of annual maxima for each decade, daily data in white, hourly data in black, and (c) cumulative density function, daily data in black, hourly data in gray.

Table 1. Stations With >100 Years of Record Between 1890 and 2010, Given With the Data Source, Total Years of Record, and Station Elevationa
Station NameData SourceRecord Length (year)Elevation (m)
  1. aData sources include the following: CD, Climatologic Data of the National Climate Data Center; EBMUD, East Bay Municipal Utility District; MMWD, Marin Municipal Water District; SFWD, San Francisco Water Department. Stations are shaded red in Figure 1a.
Antioch PPCD10718
Chabot ReservoirEBMUD11875
Davis 2 WSWCD10818
Folsom DamCD117107
Lagunitas LakeMMWD119239
Lake MercedSFWD1025
Martinez WPCD10112
Mount HamiltonCD1151282
Napa State HospitalCD10922
Newman 2NWCD10827
Oakland NWSCD1101
Saint HelenaCD10278
San FranciscoCD12016
San JoseCD11220
San Andreas LakeSFWD118115
Santa RosaCD10751
Stockton FS 4CD1124
Upper Crystal SpringSFWD11791

[8] A majority of the precipitation observations come from the standard National Oceanic and Atmospheric Administration (NOAA) 20 cm-diameter rain gauge, an observational standard since the early twentieth century. Data records were assembled and annotated by a former California State climatologist at CADWR who provided the authors with annual maximum accumulations for multiple durations. Annual data coincides with the water year, October through September. These records include demarcation of periods containing uncertain or missing data, which were omitted from the analyses presented in this study.

[9] We split the complete data record into two time periods for initial comparison of DDF characteristics: 1890 to 1955 (referred to herein as the “early period”) and 1956 to 2010 (“late period”). Data used for DDF analysis had at least 15 years of data in one or both of the study time periods. Data used for individual station statistical analyses came from stations that had a minimum number of records within both time periods: 63 stations with >30 years of data and 31 stations with >50 years of data within both time periods. Data used for decadal DDF analysis and sliding interval DDF analysis came from the same 31 stations having >100 years of data.

3 Methods

3.1 Regional Depth-Duration-Frequency Analysis

[10] We applied a standard metric for quantifying the probability of recurrence of large storms: the annual exceedance probability, p. The annual exceedance probability is the probability that, for a particular duration, an event of a given size or larger will be the largest event in a year. The recurrence interval of such an event, which has an intensity of depth/duration, is calculated as RI = 1/p. Exceedance probability analysis is useful for quantifying extremes in hydrologic conditions because it is widely applied for a variety of data types, is simple to implement, benefits from the availability of relatively long data sets, and produces DDF values that are readily interpreted and understood by researchers, resource managers, industry, and the public at large.

[11] Using observational records from 1890 to 2010 (Figures 1b and 1c), we started with the largest annual event depth at each station for 17 durations (event lengths, L) ranging from 1 h to 60 days. The exceedance probability was calculated for each time period using a Pearson type III distribution fitted to the annual maximum depth data, with skew and kurtosis calculated for individual stations. The Pearson type III distribution was used because it provided the most consistent fit across the full range of annual maximum storm sizes, in comparison to other common distributions such as the general extreme value distribution. We report results for six recurrence interval (RI) values: 2, 5, 10, 25, 50, and 100 years. Event depth was calculated for every L-RI pair, resulting in 102 characteristic storm intensities.

[12] We apply this approach to four separate analyses: (1) Preliminary DDF analysis is applied to early and late time periods using stations having >15 years of data in either period, followed by a Markov chain Monte Carlo (MCMC) analysis and nonparametric tests to determine whether the two datasets are significantly different. (2) We repeat the DDF analysis for the same two time periods using only stations with >30 years of data in either period, then calculate the median exceedance depths for each period. (3) We repeat the DDF analysis for decadal time periods using stations with >100 years of data and compare results to synthetic exceedance depths generated from the distribution of observations. (4) We compare changes in DDF characteristics at individual stations with >30 years of data in both periods and repeat the DDF analysis for sliding 25 year time periods using stations with >100 years of data. Analyses (1) through (3) use averaging of multiple stations to increase the signal-to-noise ratio and provide a regional perspective on extreme precipitation changes, whereas analysis (4) allows us to quantify changes at individual stations and assess spatial heterogeneity. These four analyses are described in sequence in the following sections.

3.2 Markov Chain Monte Carlo Algorithm and Secondary DDF Analysis

[13] Early and late time periods were defined with the intent of determining whether exceedance depths had changed in the SFBA between the early time period evaluated previously [Rantz, 1971] and the late time period of roughly equal length that followed. Like Rantz [1971], we completed DDF analyses for the entire SFBA and then identified the linear best fit model for each L-RI pair, based on regressing storm exceedance depth against MAP. To test whether the linear best fit models for the early time period were consistent with data from the late time period, we conducted a Bayesian analysis using an MCMC approach.

[14] The MCMC method develops a distribution of model parameters (in this case, slope and intercept from a regression of exceedance depth versus MAP) generated from a random ensemble of proposals drawn from the data population of each applicable weather station. Each proposal is accepted or rejected based on known constraints, in the form of prior distributions and the resultant model's misfit from accepted data. The resulting ensembles of modeled parameters are interpreted as data-conditioned probability distributions. We determined the distribution of linear models that fit exceedance depth versus MAP data for early and late time periods based on 10,000 trials for each station tested. We ran the MCMC analyses for 72 daily L-RI pairs (L = 1 to 60 days, RI = 2 to 100 years). Event durations <1 day were omitted from the MCMC analyses because there are fewer stations having continuous records of this type, particularly for the early time period. The MCMC approach generates an estimate of uncertainty in the slope relating exceedance depth to MAP and allows for fitting to actual distributions of annual precipitation data for each station. This method is applicable for interannual variations in annual precipitation that are not normally distributed around the MAP, as for many of the data sets assessed in this study, although it also works for normal distributions. This method extends that used by Rantz [1971] by generating a distribution of slope values, rather than a single best fit value, relating precipitation exceedance depth to MAP for each L-RI pair.

[15] We used two nonparametric methods to compare the distributions of best fitting linear models from the early and late time periods for each L-RI pair: Mann-Whitney and Kolmogorov-Smirnov tests. The Mann-Whitney test uses rank analysis, whereas the Kolmogorov-Smirnov test compares cumulative density functions, to determine the probability that data from two sample sets are drawn from a single data set.

[16] We repeated the DDF analysis using all stations in the SFBA with >30 years of data within either the early or late time period. Exceedance depths were calculated for every such station for both time periods and then organized by L-RI pair. We compared the median exceedance depth for each L-RI pair in the early and late time periods.

3.3 Decadal Analysis and Estimation of Variability

[17] Using 31 stations with greater than 100 years of data, the DDF analysis was repeated for 11 decadal time periods between 1890 and 2000. For the decadal analysis, as with the two-period analysis, the exceedance probability was calculated for each individual station using each set of data separately, and the median of the results from all stations was calculated for each L-RI pair. A linear best fit model was calculated, and the slope was used to describe the general trend in exceedance depths over the study period. Breaking the data set into shorter time intervals makes it less reliable for assessing variations in long-RI events, so the decadal analysis was restricted to RI values of 2 and 5 years. This analysis was also restricted to event durations of 1 to 60 days because long records of hourly data were rare.

[18] To account for variability in the DDF analysis using decadal periods, we generated synthetic data by selecting randomly from the full distribution of observations from 1890 to 2010 for every station and event duration. The DDF analysis was then repeated using 10 annual randomly generated maximum storm sizes. Following the same method as for the observations, we calculated the best fit model slope of the calculated exceedance depths over time. We repeated this process 150 times and used the distribution of slopes to represent the natural variability of the data trend over time. The slopes from observations were compared to the distribution of randomly generated slopes to determine if the observed trends of exceedance depth versus time were anomalous relative to natural variability.

3.4 Individual Station Analysis

[19] We compared DDF characteristics calculated during multiple time periods to quantify changes at individual stations. The first analysis used the early and late time periods, whereas the second used sliding 25 year time periods.

[20] Sixty three stations with >30 years of data in the early and late time periods were used to compare storm intensity differences to MAP differences for individual stations. After calculating differences in exceedance depths for individual stations between the two time periods, we calculated the percent difference in depth and the normalized difference for each station (the percent difference in exceedance depth divided by the percent difference in MAP at the same station). The normalized difference value is 1.0 if the percent difference of event exceedance depth is the same as the percent difference in MAP, indicating that the intensity of extreme storms changed commensurately with MAP. Calculated changes are plotted by station location to assess continuity and variability. Data from three major cities (Santa Rosa, San Francisco, and San Jose) in the SFBA were compared to quantify the nature of variations within the area of regional analysis. These three cities represent a variety of geographic settings, being separated by up to 160 km. The cities have differing proximities to the Pacific Ocean, San Francisco Bay, and coastal mountain ranges, all of which may influence patterns of extreme precipitation. Also, all three cities have precipitation records >100 years long.

[21] The second analysis compares exceedance depths calculated using the sliding 25 year DDF analysis for 31 stations with >100 years of data. The exceedance depths calculated for 25 year periods can be compared over time, showing trends not captured in the analysis of only two (early and late) time periods. The series of exceedance depths are presented for the three cities listed above to show how they vary with time.

4 Results

4.1 Changes in Mean Annual Precipitation and Exceedance Depth

[22] MAP in the SFBA changed by +4.5% between the early and late time periods, although there is considerable variability between stations (Figure 2). In contrast, for the state of California overall, there was a change of −4.2% in MAP between the same two time periods [National Oceanic and Atmospheric Administration (NOAA), 2010]. The increase in MAP in the SFBA was accompanied by large changes in storm exceedance depth for most L-RI pairs. Changes in MAP vary greatly between stations (Figure 2b), as do the observed changes in extreme precipitation intensity.

Figure 2.

(a) Annual total precipitation from 31 SFBA stations with annual total records longer than 100 years (grey), 10 year running average for each station (black), and 10 year running average for State of California (red) [NOAA, 2010]. Periods with positive average PDO index [Mantua et al., 1997] are shown highlighted in pink. (b) Median annual precipitation for the early and late time periods, 1890 to 1955 and 1956 to 2010, respectively. Boxes represent 1 standard deviation with whiskers at maximum and minimum values for each period.

[23] As shown by Rantz [1971], the relationship between exceedance depth and MAP for SFBA stations during the early time period was often linear, with greater variability in event depths where MAP was greater. In the late time period, we found greater variability, increased storm intensity (e.g., Figure 3a), and steeper slopes relating storm exceedance depth to MAP (e.g., Figure 3b). Coefficients for equations relating exceedance depth to MAP for all L-RI pairs in the late time period are presented in Tables A1 and A2 in the supporting information. Mann-Whitney and Kolmogorov-Smirnov tests of the two slope distributions show that early-period data and late-period data comprise separate data sets for all 72 L-RI pairs (α < 0.01%). Thus there is a significant difference in the intensity of extreme events during 1890–1955 and 1956–2010.

Figure 3.

Example of exceedance depth versus MAP and distribution of slopes for 10 day, 25 year events, calculated for early and late time periods. (a) Exceedance depth versus MAP, with early period data (1890–1955) in black and late period data (1956–2010) in gray. The early period result is essentially that calculated by Rantz [1971], whereas the late period data show more variability, greater storm intensity, and a steeper slope. (b) The distribution of linear model slopes that fit the data from Figure 3a based on an MCMC analysis, with early period data in black and late period data in white. This histogram shows visually what was revealed through nonparametric statistical analyses: data from the two time periods comprise separate data sets.

[24] To illustrate the difference in storm intensity between the two time periods, we compare the median exceedance depths for 1, 3, and 10 day storm durations (Figure 4a). The increase in median exceedance depth between the early and late time periods for the full SFBA is 0.90 to 6.8 cm/event, corresponding to large decreases in the RI of large storms. For example, a 10 day, 50 year event during the early time period became the 10 day, 10.2 year event during the late time period. Similarly, the exceedance depths for 25 year RI storms during the late time period are equal to or greater than the equivalent depths for 50 year RI storms during the early time period for all durations of 1 to 60 days, indicating a ≥100% increase in the exceedance probability of a given event (Figure 4b). The distribution of exceedance depths for 10 day, 50 year RI events for all stations with >30 years of data is clearly different for the early and late periods (Figures 4c and 4d), with the latter having a higher mode and a large fraction of stations falling above the mode during the late period.

Figure 4.

Median exceedance depths from early and late time periods shown for (a) 1, 3, and 10 day storms plotted for RI = 2 to 100 years, (b) RI = 50 year storm in the early period, and the RI = 25 year RI storm in the late period, and (c and d) distribution of exceedance depths for the 10 day, 50 year events from stations with >30 years of data in the early and late periods, respectively. The arrow in Figure 4a shows that the 10 day, 50 year storm exceedance depth in the early time period is equal to the 10 day, 10.2 year RI storm in the late period.

[25] Median exceedance depths calculated over decadal time periods for RI = 2 year and RI = 5 year storms have increased over the past 120 years (Figure 5). Observed trends in exceedance depths versus time were compared to natural variability by generation of synthetic data using the full population of extreme events for each station. Trends in exceedance depth versus time were calculated from synthetic distributions for 24 L-RI pairs corresponding to L = 1 to 60 days and RI = 2 and 5 years. Nineteen of 24 slopes calculated from the observed exceedance depths were greater than 2 standard deviations from the mean slope generated from synthetic data (examples shown for 1 and 10 day events in Figure 6), indicating that decadal trends are unlikely to result from natural variability (Figure 5 and Table 2). In fact, all 24 L-RI pairs used for decadal analysis show a positive trend of exceedance depth versus time.

Figure 5.

Exceedance depths calculated for (a) 1 and (b) 10 day storms, with 2 year (dashed line) and 5 year (solid line) recurrance intervals, for each decade of the twentieth century. The slope (m) is given for all daily storm durations in Table 2.

Figure 6.

Distribution of slope values for synthetic decadal exceedance depths, based on random selection of annual maxima from the best fit distribution to the full observational dataset (1890 to 2010). (a) 1 day, 2 year RI events, (b) 1 day, 5 year RI events, (c) 10 day, 2 year RI events, and (d) 10 day, 5 year RI events. The slopes of observed storm exceedance depth versus time (Figure 5) are indicated along the x axis with vertical arrows.

Table 2. Slope (m) of Linear Regression Showing Change of Exceedance Depth Analysed Over Decadal Time Periodsa
Duration, L (day)RI (year)m (cm/year)
  1. aRegression slopes are given for the change in precipitation exceedance depth for ≥1 day storms at 2 and 5 year RI. For each L-RI pair, observations with a slope not greater than 2 standard deviations from the mean randomly generated slope are indicated in bold. Decadal data and trends for 1 and 10 day storms are shown in Figure 2.

4.2 Changes in Storm Intensity Relative to MAP

[26] On average, the increase in the intensity of extreme storms in the SFBA is greater than the increase in annual precipitation, with the largest storms showing the most disproportionate increase. These results are consistent with some large-scale analyses [Karl and Knight, 1998; Groisman et al., 1999, 2005; Fowler and Kilsby, 2003]. For storm durations of 1 h to 60 days, the median normalized change in extreme storm intensity ranges from 1.01 to 1.52. On average across the SFBA, ~80% of the precipitation intensity increase for daily interval duration storms can be attributed to an increase in MAP, whereas the rest of this change results from a concentration of a greater fraction of annual precipitation into a smaller number of larger events. More than 20% of SFBA stations show a percent increase in storm intensity four times larger than the percent increase in MAP. Storm intensity changes are often inconsistent with long-term trends in MAP, as seen in some earlier studies [Alpert et al., 2002; Goswami et al., 2006; Leahy and Kiely, 2010; Douglas and Fairbank, 2011]. We also found that the relation between storm intensity and MAP varied greatly between stations in a relatively small geographic region, as discussed in the next section.

4.3 Local Variability

[27] Differences in annual precipitation and in the most intense storms were localized within the SFBA (Figure 7). The spatial scale of changes in extreme storm events appears to occur at ~20 to 50 km, consistent with spatial scale studies of extreme precipitation [Booij, 2002; Smith et al., 2005] and >10 times finer than the spatial resolution of previous precipitation change studies. The MAP (Figure 7a) and change in MAP (Figure 7b) vary greatly across the region. A majority of the stations with >100 year records show increases in MAP (Figure 7b), although there are distinct decreases in the dryer Central Valley (~100 km east of San Francisco). The percent change in storm intensity (Figure 7c) generally has the same sign as change in MAP, but has a greater magnitude for many stations, indicated by normalized change values >1. We found no correlation between changes in precipitation intensity and elevation.

Figure 7.

Individual station analysis results showing (a) MAP from the later period, (b) MAP percent change between the early and late periods, and (c) average percent change of daily (L = 1 to 60 day) storm exceedance depths between the early and late periods.

[28] In a comparison of the three urban subregions, we found that event intensity tended to increase more in San Francisco and Santa Rosa than in the San Jose metropolitan area (Table 3), with the largest local differences being associated with the largest events. For 2, 5, and 10 day events having RI = 50 years, the intensity increase for San Jose was −11.6% to 9.9% (−1.0% to 0.8%/decade). In contrast, the same L-RI events had an intensity increase of 4.0% to 42% (~0.3% to 3.5%/decade) in San Francisco, with exceedance depth increases ranging between 0.4 to 6.3 cm.

Table 3. Average Changes in Storm Exceedance Depth Given for Three Cities in the SFBA Comparing Early and Late Periods Between 1890 and 2010a
LocationRI (year)2 Day Event5 Day Event10 Day EventMean Normalized Change in Depthb
Change (cm)% ChangeChange (cm)% ChangeChange (cm)% Change
  1. aChange and percent change are given for 2, 5, and 10 day events at 2 through 50 year RI and the mean normalized change in depth from all storm durations.
  2. bPercent change in event depth divided by the percent change in MAP at each station, respectively.
Santa Rosa20.767.030.865.991.386.649.94
San Francisco20.407.100.496.040.413.971.38
San Jose20.348.110.609.860.658.000.82

[29] Exceedance depths calculated using the sliding 25 year analysis period for the station with the longest record in each of the three subregions illustrate more detailed trends (Figure 8). Although the overall trend in exceedance depths in both San Francisco and Santa Rosa increased during the period of record, the time when most of the increase occurred varied with location. In Santa Rosa, the most dramatic increase occurred during 1910–1960, whereas the largest increase in San Francisco was during 1940–1990. In contrast, exceedance depths in San Jose remained essentially level from 1930 to 2000.

Figure 8.

Exceedance depths calculated using a sliding 25 year analysis period given for three major cities in the SFBA: (a) Santa Rosa, (b) San Francisco, and (c) San Jose. Exceedance probabilities were calculated for 25 year periods starting with 1890 to 1915 and ending with 1985 to 2010. Exceedance depths are given for the 5, 10, and 25 year RI storm events.

5 Discussion and Conclusions

[30] Our analyses indicate that there have been significant changes in both the frequency and magnitude of extreme precipitation events around the SFBA over the last 120 years. Regional average increases in storm intensity are greater than estimated for the same area by analysis of larger regions [Kunkel et al., 1999; Groisman et al., 2004]. In addition to identifying a significant change in storm intensity within the SFBA, we found greater variability than was suggested from previous analyses [Rantz, 1971; Abatzoglou et al., 2009], particularly for the largest events. For example, 25% of the stations in our study show >5%/decade increase in 25 year RI storm magnitude, more than three times that suggested by larger-scale analyses. Collectively, these results suggest that municipal planning, infrastructure design, and risk assessment should be updated in response to observed historical (and likely ongoing) trends and, in many cases, should emphasize local historical observations.

[31] The issue of variability is central to the discussion of extreme precipitation changes over time. Our objective was to quantify observed trends and determine whether calculated changes in exceedance depths were representative of a multi-decadal trend or could comprise random temporal variability. By generating synthetic values from the full distribution of data in the decadal DDF analysis, we demonstrated that a majority of the exceedance depths based on observations were located several standard deviations from the randomly generated mean, indicating that the increasing trends are unlikely to be random. Confidence that decadal trends were not random was higher for 2 year RI events than for 5 year RI events, but that may result (in part) because DDF characteristics were calculated using 10 year data intervals.

[32] The mechanism(s) responsible for increased precipitation intensity observed during the late time period were not assessed in this study but could include greater atmospheric moisture transport and/or storage due to atmospheric warming or changes to wind patterns or speeds, perhaps in association with annual and decadal-scale global climate phenomena, including atmospheric rivers. There are likely to be many factors involved, as indicated by the irregularity of decadal variations in exceedance probability values for RI = 2 and 5 years (Figure 6). However, the finding of positive slope values for exceedance depth versus time for all 24 L-RI pairs in the decadal analysis (Table 2) suggests that the increase in storm intensity shown in this study is not an artifact of the time periods selected. We cannot determine whether changes are due to natural variability or anthropogenic climate change.

[33] The sliding 25 year analysis provides insight into the multi-decadal changes in extreme precipitation at individual stations. Results support the spatial variability predicted based on the two-period analysis using stations with shorter records. In addition, the sliding analysis clarifies trends over the 120 year period between stations that appeared to have similar increases in exceedance depth using the two time period analysis. Although we calculated a linear trend to quantify change in the decadal analysis, the trends are clearly more complex.

[34] Quantitative changes reported in three cities within the SFBA (Table 3) provide further insight into the variability of storm intensity change with respect to MAP. San Francisco and Santa Rosa both show increases in storm intensity, with the greatest changes associated with the largest storms, consistent with the sliding 25 year period analysis. The increases in storm size observed in San Francisco and Santa Rosa are accompanied by increases in MAP, but not to the same degree. The normalized change in extreme events in Santa Rosa exceeds 9.0 in all cases, whereas it ranges from 1.4 to 4.4 for San Francisco. In contrast, San Jose shows increases for smaller storms (2 and 5 year RI) and decreases for larger storms (25 and 50 year RI). These results imply both that intensity is not changing uniformly and specifically that small and large storms have changed in dissimilar ways within the SFBA. Most of the large-scale analyses have concluded that the largest storms are increasing the most, as the average of our study suggests, but some local results show the opposite trend.

[35] Greater intensity in extreme precipitation and an associated increase in runoff in urbanized areas are likely to contribute to more frequent and extensive flooding, erosion, sediment transport, and other geomorphologic changes [e.g., Baker, 1977; Mulligan, 1998; Osterkamp and Friedman, 2000]. Urban planners, resource managers, homeowners, first responders, and other stakeholders need to reassess whether infrastructure, land-use policies, and agricultural practices are adapted for current and anticipated future hydrologic conditions. Historical daily precipitation records are available for many areas, which would allow analyses similar to those reported herein to be applied elsewhere. Large-scale studies demonstrating changes in extreme precipitation, combined with results emphasizing the importance of local heterogeneity, suggest that other municipal regions should undertake similar analyses and assessments.

[36] The nonstationarity seen in records from the last 120 years of extreme precipitation events in the SFBA shows how historical analyses tend to lag current conditions. Thus, consideration should be given to using trends from precipitation records to project the future probability of major events. Updated records of major event frequency and intensity are also important for assessing how the global hydrologic system continues to respond to anthropogenic forcing [Zhang et al., 2007; Barnett et al., 2008; Min et al., 2011] and for creating data sets needed to test and calibrate the next generation of regional and global climate models.


[37] We thank Jim Goodridge (California Department of Water Resources) for providing thorough and organized precipitation records extending back more than a century. This manuscript benefitted greatly from three anonymous reviews. This work was supported by the National Science Foundation Graduate Research Fellowship Program (ID# 2009083666), the National Institute for Water Resources (grant 08HQGR0054), and The Recharge Initiative.