U.S. hourly surface observations are examined at 145 stations to identify annual and seasonal changes in temperature, dewpoint, relative humidity, and specific humidity since 1930. Because of numerous systematic instrument changes that have occurred, a homogeneity assessment was performed on temperatures and dewpoints. Dewpoints contained higher breakpoint detection rates associated with instrumentation changes than did temperatures. Temperature trends were tempered by adjusting the data, whereas dewpoints were unaffected. The effects were the same whether the adjustments were based on statistically detected or fixed-year breakpoints. Average long-term trends (1930–2010) indicate that temperature has warmed but that little change has occurred in dewpoint and specific humidity. Warming is strongest in spring. There is evidence of inhomogeneity in the relative humidity record that primarily affects data from prior to 1950. Therefore, long-term decreases in relative humidity, which are strongest in winter, need to be viewed with caution. Trends since 1947 indicate that the warming of temperatures has coincided with increases in dewpoints and a moistening of specific humidity. This moistening is especially pronounced during the summer in the Midwest. For the nation, trends in relative humidity show little change for the period 1947–2010, during which these data are more homogeneous. Moistening has occurred throughout the central United States while other regions have experienced drying. Urban-related warming and drying trends are present in the data, but their effect is minimal. Regional changes in land use and moisture availability are likely influencing trends in atmospheric moisture.
Water vapor is an important greenhouse gas that plays a critical role in both the hydrological cycle and surface energy budgets. Recent global warming affects the amount of moisture in the air, as change in atmospheric water vapor is intrinsically linked to climate change. Increasing evidence from observations and climate models indicates that anthropogenic activity is increasing atmospheric moisture (Boucher et al. 2004; Willett et al. 2007; Santer et al. 2007; Min et al. 2008). Through advancing knowledge of how atmospheric water vapor changes are occurring, improvements can be made to understanding its spatial and temporal responses to climate change.
Relative humidity and specific humidity are the primary measures of atmospheric moisture, describing the degree of saturation of the air and the amount of water vapor in the air, respectively. Through the Clausius–Clapeyron relation, if relative humidity is constant then specific humidity increases exponentially with temperature. Most studies of global atmospheric moisture indicate that recent warming is associated with an increase in specific humidity and little change in relative humidity (Dai 2006; Trenberth et al. 2007; Willett et al. 2008). A more recent analysis suggests that reductions in relative humidity have occurred because of a lack of increase in oceanic moisture being supplied to warming land surfaces (Simmons et al. 2010).
The North American Regional Reanalysis also shows that temperatures have a positive relationship with specific humidity and a negative relationship with relative humidity (Lu and Takle 2010). Changes in North American atmospheric moisture have been shown through increases in specific humidity and some seasonal decreases in relative humidity (Gaffen and Ross 1999; Vincent et al. 2007). Regional-scale and local-scale effects, such as airflow regimes, land-use change, and urbanization, influence U.S. dewpoints (Schwartzman et al. 1998; Robinson 1998, 2000), indicating that the factors affecting atmospheric moisture are very dynamic.
To further understanding of atmospheric moisture dynamics, it is desirable to have reliable long-term records. In the United States, the Climate Data Modernization Program (CDMP) has recently digitized hourly observations dating from the late 1920s, in addition to completing records that were only digitized at 3-h frequency. The utility of analyzing trends in these long-term records is contingent on their homogeneity, as biases in the data may distort real changes in the observations. This study aims to investigate the homogeneity and analyze trends in temperature and atmospheric moisture variables for the United States since the 1930s.
U.S. hourly synoptic observations were obtained from version 2 of the Integrated Surface Database (ISD; Smith et al. 2011) and the Applied Climate Information System database of the Northeast Regional Climate Center. Stations that were digitized by the CDMP formed the basis of the initial dataset. This initial dataset was then refined to select the longest and most complete records. Some station records were combined to provide longer records because they experienced subtle, yet distinct, moves. These records are denoted by an asterisk on the station-location map in Fig. 1, and the series were also checked for consistency.
Atmospheric moisture can be calculated by using a number of different observed variables. Theoretically derived relationships among surface air temperature T, surface station pressure p, and surface dewpoint temperature Td or surface wet-bulb temperature Tw were used to calculate surface relative humidity RH and specific humidity q. Although T, Td, and p are the most commonly reported observations, some of the earliest data reported Tw without a Td observation. In these instances in which Tw and T were observed, Td was derived using Bolton’s (1980) psychrometric formula. The RH and q were calculated according to Dai (2006).
Temperature and dewpoint records that were at least 90% complete and had start dates that are prior to 1948 were chosen for analysis (Fig. 2). A threshold of 90% was chosen to include as many stations as possible while eliminating those that contained limited or sporadic data. Most of the stations also observed station pressure; a small number of stations had less complete pressure data, however. When pressure was not continuously recorded, the median station pressure was used, as this is a similar approximation to that used by Willett et al. (2008) to allow RH and q to be calculated. The errors associated with this approximation are negligible (Willett 2007). The final dataset contained 145 stations dating back to 1930. Figure 1 shows the locations of these stations within the National Climate Assessment regions. The United States is divided into eight regions: Northwest (NW), Southwest (SW), Great Plains (GP), Midwest (MW), Northeast (NE), Southeast (SE), Alaska (AK), and Islands (IS). These regions are broadly similar to those used by Gaffen and Ross (1999) in their analysis of U.S. surface humidity.
Quality-control checks have been applied to the ISD (Smith et al. 2011), and any data that did not pass these checks were eliminated. Additional checks were made to incorporate basic physical range checks for temperature (from −63° to 63°C), dewpoint (from −100° to 63°C), relative humidity (0%–100%), and specific humidity (0–99 g kg−1). A visual inspection of the plotted data was also performed to ensure that any further outliers were removed.
It is highly desirable for any dataset that is being analyzed to have accompanying metadata. Such records of site, instrument, or observing-routine changes serve to guide and validate any adjustments for inhomogeneities. Many of the metadata associated with the ISD are stored as images of the original station-history forms and are not yet digitized. The completeness of these records is also uncertain. Only the most recent data have metadata describing changes that may have affected the homogeneity of the hourly data. There have been a number of well-documented dewpoint-instrumentation changes that have occurred in the United States (Ratner 1962; Elliott 1995; Gaffen and Ross 1999; Robinson 2000; Lin and Hubbard 2004b, 2008; Brown and DeGaetano 2009, 2010). These instrument changes are listed in Table 1.
A recent Canadian study into the influence of the change from a psychrometer to the dew cell showed the homogeneity of T to be unaffected whereas inhomogeneity in the Td series only occurred in the northeastern part of Canada, where extremely cold temperatures (below −39°C) affected readings (Vincent et al. 2007). Since these inhomogeneities were associated with dew-cell problems that occur at low temperatures for which q varies little, time series of q were found to be homogeneous (Vincent et al. 2007). Inhomogeneities had the greatest effect on RH, for which adjustments were necessary at 69% of the Canadian stations.
In the United States, Gaffen and Ross (1999) performed four checks and found that instrument changes did not significantly affect U.S. data from 1961 to 1995. The effect of these instrumentation changes has been shown to affect T and Td at some stations (Lin and Hubbard 2004a, 2008), but their influence is inconsistent (Robinson 2000; Brown and DeGaetano 2009). Consequently, a number of studies that were based on U.S. hourly temperatures and dewpoints chose not to adjust the data (Gaffen and Ross 1999; Changnon et al. 2006). It was determined that the analysis of the new long-term dataset would benefit from a homogeneity assessment because of recent humidity-instrument changes as well as the uncertainty associated with changes in the early part of the record.
c. Homogeneity investigation
The homogeneity of the Td series were investigated using the method that was described fully in Brown and DeGaetano (2009). This method utilizes reference series and determines inhomogeneities at annual resolution. Reference series are created for the sparse network of dewpoint temperatures using the comparatively dense Cooperative Observer Program (COOP) network. Reference series were generated from minimum temperatures observed at multiple nearby COOP network stations under specific weather conditions (foggy, precipitating, and clear calm) for which dewpoint and minimum temperature are similar. The correlation between annually averaged temperature series that are based on the hour at which daily minimum temperatures occurred and annually averaged dewpoints series for the corresponding hour under each weather condition generally exceeded 0.8. Such correlation is similar to that which is typically available for conventional homogeneity tests for daily temperature. Three different methods of obtaining reference series and three different methods for determining breakpoints were applied, and breakpoints were identified on the basis of the consistency of the methods (Menne and Williams 2005).
An analogous method was followed to detect breakpoints in the hourly temperature series. Reference series that were based on average daily temperatures were used to assess the homogeneity of the hourly temperatures. Average temperatures in the COOP data are calculated by averaging daily minimum and maximum temperatures. At hourly stations, the minimum and maximum of the 24 hourly temperatures were used to generate comparable averages. The annual averages of the daily average temperatures for COOP data were then used to create the reference series.
Although the Td homogeneity-assessment procedure of Brown and DeGaetano (2009) was able to be adapted to identify breakpoints in T, this method could not be applied to RH and q as suitable reference series could not be developed from COOP data. In addition, the complexity of assessing the homogeneity of these two variables is compounded by potential inhomogeneities in the T and Td series from which they are derived. Therefore the homogeneity of RH and q was not assessed.
Figure 3 shows the frequency of detected breakpoints for annual T and Td since the 1930s. Fewer temporally consistent breakpoints were detected in T when compared with Td, with the most prominent T spike occurring around the time of the Automated Surface Observing System (ASOS) installation. In contrast, there are four pronounced spikes in Td breakpoint frequency that coincide with site and instrumentation changes listed in Table 1. Many of the stations with a breakpoint in the late 1940s experienced relocation around this time, whereas the changes in the early 1960s and 1980s correspond to hygrometer installation/modification. The ASOS transition in the 1990s showed little effect on Td until the early 2000s when this variable was no longer directly measured by the sensor. Breakpoints do not show regional patterns that would suggest that local climatological behavior affects the detectability of the instrument changes.
Despite Fig. 3 showing evidence of a link between instrument changes and inhomogeneities in the data, these links are only detected at a small proportion of stations. The low detection rate could be attributable to site moves that coincide with instrument changes, compensating for the instrument biases. Improved metadata information would have allowed the effect of site changes to be assessed, as these could have a larger effect on homogeneity than did instrument changes. More extensive metadata would also enable a more comprehensive assessment of which breakpoints are being detected and which are not being detected so as to further evaluate the sensitivity of the breakpoint method. As with all homogenization techniques, breakpoint detection is adversely affected in sparsely observed regions as the reference stations are too distant to detect smaller inhomogeneities.
Using the detected inhomogeneities, annual and seasonal adjustments [December–February (DJF), etc.] were calculated from the differences between station averages and the reference-series average temperatures, before and after each breakpoint. For Td, the annual and seasonal averages of daily Td, recorded at the time of minimum hourly temperature, and reference-series average minimum temperatures were differenced to obtain adjustments (Brown and DeGaetano 2009). The average of two different reference-series methods (Menne and Williams 2005) was used to determine the adjustments when both were able to be constructed; otherwise, only one was used. Although a third reference-series method was used to detect breakpoints, its use of a multiple-linear-regression procedure (Vincent 1998) precluded its use for estimating the magnitude of inhomogeneities. The adjustments were applied retroactively relative to the most recent homogeneous segment of data. Adjustments were only made where there were at least 5 years of reference-series data on either side of the breaks; otherwise, the series was not adjusted.
Fixed breakpoints were also assumed in all series on the basis of nationwide instrument changes that occurred in 1960, 1985, 1995, and 2004. An additional breakpoint in 1947 was included as a large number of sites were relocated at this time. These fixed breakpoints were adjusted using the same procedure as was used to adjust the detected breakpoints. Although these fixed breakpoints may not match the exact year that changes were made at a specific station, the effects of such inaccuracies in the annual series are not likely to have a significant impact on trends. It was also expected that if these breakpoints did not cause inhomogeneity in the data their effect on the time series (and hence adjustment) would be minimal. Regardless, in both cases the fixed breakpoints provide a means of assessing the sensitivity of the trend analysis to homogenization.
Long-term trends in T, Td, RH, and q were assessed on the basis of their anomalies calculated relative to a 1947–2010 base period. Because of the varying start dates of the stations used here, this base period was selected to maintain comparable anomalies throughout the U.S when compiling averages dating from 1930. Annual and seasonal trends at 145 stations were analyzed and were subsequently aggregated to regionally averaged time series. Spatial plots of trends generally focus on the period 1947–2010 since data were available at all stations over this time period. Thus, station trends are comparable during this common period. Throughout this section, trends were calculated using a nonparametric Thiel–Sen slope estimator and significance was determined by using a prewhitened nonlinear trend that uses a Kendall test to determine significance (Zhang et al. 2000; Wang and Swail 2001). A significance level of 5% is used to indicate a statistically significant trend.
The adjustments applied to T and Td series are summarized in Fig. 4 and show that most adjustments are in the range of ±0.5°C. Little difference is seen between detected adjustments Tadj and fixed adjustments. On average, T adjustments increase as time decreases, with the earliest temperatures adjusted by 0.5°C. In contrast, Td adjustments are negative, positive, and then negative again, with the adjustments showing slightly larger variability than do those for the T series. The T adjustments exhibit a consistent response, with most adjustments being positive, suggesting that the earliest hourly observations throughout the United States are consistently colder on average than recent observations, relative to the reference series. However, Td adjustments are both positive and negative, indicating that less-coherent adjustments are being applied to the data.
Comparison of unadjusted annual T and Td trends with those based on detected adjustments and fixed adjustments indicates that inhomogeneities have affected the data trends (Fig. 5). A mixed pattern of enhanced and diminished trends occurs in the western U.S. stations as a result of the homogeneity adjustments. Eastern and central interior regions of the United States show the most change in annual T trends, with adjustments causing station trends to decrease.
In contrast to T, more trends in Td increase after adjustments are made, although few station trends are significant. Only in the Midwest are there coherent decreases in the Td trends as a result of adjustments. There is little difference in annual trends of the detected and fixed adjustments series for both T and Td, and the magnitude of the change in trends that result from the adjustments is less than 0.3°C (10 yr)−1. Similar effects are seen in seasonal trends (not shown).
b. Temperature trends
Trends in U.S. annual-average T anomalies for the period from 1930 to 2010 are shown in Fig. 6. Seasonal trend magnitudes are listed in Table 2. Annual-average T shows little change initially, before an increase after the 1970s is apparent. The overall increase in T of 0.072°C (10 yr)−1 is significant, but the trends in adjusted temperature data are one-half of this value and lack significance. The increase in T is most evident in March–May (MAM) for which 0.14°C (10 yr)−1 of warming occurs for the unadjusted data; although this warming is reduced when the data are adjusted, the trend significance is retained (Table 2). Detected and fixed breakpoint adjustments indicate that the earliest T observations have a cold bias that reduces the warming trends in the unadjusted observations (Fig. 6).
Pearson’s correlation coefficients between the homogeneous U.S. Historical Climate Network (USHCN) dataset and the unadjusted, detected adjustments and fixed adjustments were near 0.92, with little difference between the data series. Less variability is present in the annual time series derived from hourly data when compared with those derived from monthly USHCN data. The differences between the unadjusted and adjusted hourly data–based averages are also of much smaller magnitude than the difference between the hourly data and the USHCN data. These comparisons add to our confidence that the inhomogeneity adjustments that are applied are robust. It would not be fair to strictly compare our results with those based on the USHCN as there are substantial differences in the number of stations as well as station character (about one-half of our hourly station network is urban). Geographic biases in station locations also complicate a comparison. Nonetheless the new analyses show to a certain extent that our results are similar to those of the USHCN.
Most stations in the United States are warming at a similar rate for Tadj, although some interior and southern stations have weaker and nonsignificant trends (Fig. 7). The spring warming is especially apparent across the northern United States. A cluster of stations in the central part of the northern United States also exhibits strong warming in DJF. Increased warming at some urban stations—most prominently Phoenix, Arizona (KPHX), and Reno, Nevada (KRNO)—is apparent in the Tadj trends. When grouped by assessment region (Fig. 1) warming is occurring in all U.S. regions except the Southeast. The strongest warming trends are seen in Alaska, the Southwest, and the Northeast. For the common (data exist at all stations) 1947–2010 period, Tadj trends for the combined United States (ALL) are significant and of greater magnitude than those starting in 1930 (Fig. 8). The strongest trends are seen in Alaska, particularly in winter.
Trends in annual-average Tadj are also reflected in the annual extremes. The series of unadjusted (homogenization of individual daily values is not feasible) annual minimum temperature displays the same pattern of warming as does annual-average temperature (Fig. 9). Significant trends are absent only in the Southeast. The trends in annual maximum temperature are more spatially variable, with the West experiencing the greatest warming and the rest of the country showing little change. Significant decreases are confined to stations in the Great Lakes region (Fig. 9).
The character of the U.S. annual-average temperature trend experiences a notable change between the beginning and end of the record (Fig. 6). To account for this variation, two subperiods (1947–79 and 1980–2010) of series were analyzed. These two periods essentially bisect the record temporally, with the break also corresponding to the change in sign of the trends. The trends in Tadj since 1980 indicate warming, whereas the previous 33 years show a cooling trend, except in the West where the majority of stations show warming trends (Fig. 10).
c. Dewpoint and specific humidity trends
No significant change in annual-average or seasonal Td or q is apparent for the United States (Fig. 6; Table 2). In contrast to T, adjustment has little effect on the Td series. Similarities in the variability of Td and q are emphasized by a Pearson correlation coefficient of 93%, with these two variables exhibiting a decreasing trend until the 1970s when trends generally begin to increase. Because of the strong relationship between Td and q, the changes in q can be inferred from those in Td plots henceforth, unless stated otherwise.
Regional differences in Td trends are more apparent than those for Tadj (Fig. 7). The central regions of the United States show the strongest and most significant Td trends. Dewpoints increase through time in this area (Fig. 7). Like for Tadj, the strongest warming occurs in MAM, especially in the central United States (Fig. 7). Decreasing Td trends are confined to the Southeast, mainly during winter.
Over the period with complete data records at all stations (1947–2010), the U.S. average trends in Td increase and are significant during all seasons except winter (Fig. 8). Thus, the magnitudes of these trends, like those for temperature, are greater in this period. The strongest trends are seen in Alaska and, for q, the Islands, albeit only two stations exist in these regions (Fig. 8). The moistening (and warming) in Alaska is most pronounced in DJF. In the other regions, increases in Td are strongest in MAM.
In the continental United States, increases in the annual minimum Td are more widespread and of greater magnitude than changes in the annual maximum dewpoint (Fig. 9). The greatest increases in annual minimum Td occur along the northern-tier states. Like Tadj, there is a tendency for Td to decrease in the early part of the record (pre-1979), with increases more prevalent in recent years, especially the central and eastern United States (Fig. 10) The warming noted at western stations in the early part of the Tadj series is also reflected in the Td record.
d. Relative humidity
Nationally annual and seasonal trends in RH have been decreasing since the 1930s although not significantly (Fig. 6; Table 2). Decreases in RH are the strongest during DJF and MAM, with the changes in June–August (JJA) and September–November (SON) tempering this drying in the annual series (Fig. 7). The RH anomalies feature a sudden drop in the last 5 years that is atypical but corresponds to similar abrupt decreases in Td and q (Fig. 6).
A comparison between hourly derived RH and RH data from the twentieth-century reanalysis, provided by the National Oceanic and Atmospheric Administration (NOAA) Office of Oceanic and Atmospheric Research (OAR) Earth System Research Laboratory Physical Science Division (acquired online from http://www.esrl.noaa.gov/psd/) (Compo et al. 2011) indicated substantial disagreement between the two datasets, particularly prior to 1950 (Fig. 11). There also appears to be a smaller shift during the mid-1980s around the time of the HO-83 installation. In contrast, q (Td is not available from the reanalysis) is relatively consistent between the two datasets. Although the two q data series are only modestly correlated (correlation coefficient r = 0.53), there is little evidence of consistent biases in the data series. Attempts to homogenize the annual RH series by accounting for the detected biases in the temperature and Td series did not change the series too substantially. Consequently, all trends in RH must be treated with caution as evidence suggests they lack homogeneity, particularly in the 1930s and 1940s.
On a spatial scale, annual RH exhibits few significant trends since 1947, with only 16 of the 145 stations showing any evidence of a change (Fig. 7). The trends in RH are increasing throughout the central region of the United States but generally decrease to the east and west (Fig. 7). This feature is also present in Td and q for which the strongest and most significant trends are located in the central U.S. region. On a seasonal basis, the strongest RH trends are negative and occur during DJF along the Eastern Seaboard.
Of the four variables, the trends in RH show the least seasonal consistency, as positive and negative trends are present in most regions (Fig. 9). The Southwest and eastern regions of the United States feature decreases in RH; the timing of these occurrences differ, however, dipping in JJA and DJF, respectively. In contrast, RH features a nonsignificant increase in Great Plains throughout all seasons and peaks at 0.4% (10 yr)−1. The largest increases occur in JJA, primarily in the Midwest and Northeast. Unlike T and Td, RH trends are similar in the 1947–79 and 1980–2010 subperiods, except for a moistening throughout the central United States prior to 1980, coinciding with the region where Tadj displays the strongest cooling (Fig. 10).
e. Urban–rural effects
It was noted that Fig. 7 featured a number of large trends in T and RH that appear to be related to the location of urban stations. Urban stations are warmer and drier than rural stations, with the urban trends often displaying statistical significance. The 145 stations in the dataset were classified as urban or rural according to the specifications used by Brown and DeGaetano (2010); there are 73 stations classified as rural and 72 classified as urban.
Comparisons of annual trends at rural and urban stations are shown in Figs. 12 and 13. In general, station trends in the United States show that consistent changes are occurring at rural and urban stations for all variables (Fig. 12). Strong positive trends in Tadj and Td are present in the north, with only RH featuring some consistent negative trends, especially at urban locations. It is apparent from Fig. 12 that the distribution of rural stations is more heavily weighted toward northern and central U.S. stations while the urban stations are located in coastal and southern portions of the country.
The Midwest, however, contains a more even distribution of urban and rural sites than does the entire United States. Therefore, only stations in the Midwest are used to calculate average time series so as to reduce the effect of geographic bias. A slightly shorter time period for the long-term averages was also required because observations at rural stations started in 1933. Little difference exists between the long-term Td trends (1933–2010) in the two station environments, with neither showing significant changes. This is in contrast to the spatial trends using the shorter time series in which rural stations in the Midwest show significant increases (Fig. 12). An examination of the seasonal effects on T and RH since 1933 indicates that time of year affects rural and urban trends in the Midwest (Fig. 13). There are only subtle differences in rural and urban trends for T, but urban sites are consistently warming at a greater rate. Rural and urban environments have a more noticeable effect on RH. The trends in RH indicate that there is significant rural moistening occurring in JJA but that significant drying occurs at urban stations during DJF.
Homogeneous records are highly desirable for assessing trends in long-term climate data. Although there have been numerous instrument changes in the U.S. hourly dataset, their detectability is low, with less than 12% of stations indicating significant Td inhomogeneities, with lower detection rates for T (Fig. 3). These results are similar to those of the Vincent et al. (2007) homogeneity analysis for Canada, in which T was not adjusted and Td was only adjusted at a small number of northeastern stations where cold temperatures affected observations. Although instrumentation changes in the ISD have been found, their lack of consistent detection suggests that the changes do not introduce systematic bias or that other coincident site-specific factors are masking these breakpoints.
In the United States, instrumentation changes often coincide with station moves, and the lack of digitized station-history information (metadata) hinders elucidation of the true source of these breakpoints. Microclimatic changes associated with station relocations are likely to have a larger effect on temperatures than on dewpoints; more inhomogeneities were detected in the dewpoint series, however (Fig. 3). This disparity could suggest that the station moves have a lesser effect on the temperatures and dewpoints than do instrument changes, which were frequently detected in the dewpoints. Sparse networks and a lack of rigorous studies on detecting and adjusting hourly datasets hinder research establishing the homogeneity of hourly datasets. Nonetheless, with the exception of the homogeneous gridded global Met Office Hadley Centre and Climatic Research Unit Global Surface Humidity dataset (HadCRUH), this is the first study to assess trends in homogenized U.S. hourly temperature and dewpoint data.
The unadjusted T record at most hourly reporting stations has a bias toward enhanced warming. For Td, adjustment has no consistent effect. The warm T bias appears to be associated with the earliest period when the psychrometer recorded cooler temperatures than did subsequent instruments. The reliability of these adjustments is difficult to ascertain, and no other studies have found an effect of the psychrometer change in hourly data. A comparison between psychrometer and platinum-resistance-thermometer observations from ships found differences of only ±0.1°C on dry-bulb temperatures (Kent et al. 1993), but this is in contrast to our results. Detected and fixed adjustments result in similar national-average series that are more similar to each other than to the unadjusted data. Similarity of the adjusted data is driven by the reference series from which they were derived and therefore provides little additional information on biases in the unadjusted data. Future work aims to incorporate data from Canada and Mexico where different instruments and network changes occurred, and it should increase confidence in the veracity of U.S. trends.
b. Comparison with earlier research
On a national basis, warming temperatures are the most significant feature of the changes in the U.S. hourly temperature record. There is strong evidence that unadjusted U.S. hourly temperatures have warmed by 0.072°C (10 yr)−1 over the period of 1930–2010. The adjusted data indicate that this warming is enhanced by inhomogeneities, however, which temper the trends but still indicate warming of 0.039°C (10 yr)−1. These warming rates bracket the temperature increase in the full homogeneous monthly U.S. Historical Climate Network, version 2, of 0.056°C (10 yr)−1 over the same period (Menne and Williams 2009). This is not surprising given the differences in the number, location, character, instrumentation, and limited overlap of stations in these two networks.
Hourly temperatures are warming over all seasons, but the strongest increase is occurring during the spring. Springtime was also found to be the most prominent season for warming by Gaffen and Ross (1999) using a shorter time period. Vincent et al. (2007) and Lu et al. (2005) also found warming in the spring, but stronger warming was seen in the winter months for Canada and the United States, respectively. The lack of wintertime warming seen here could be attributable to the observation periods of the dataset.
Regional patterns indicate widespread warming throughout the United States with the exception of some cooling occurring in the Southeast, a feature that is consistent with previous studies (Lu et al. 2005). Annual-extreme minimum temperatures feature this regional pattern, too, although the rate of warming is higher than is seen in average temperatures. High temperatures that occur in the 1930s, which are also present in the monthly U.S. temperature record (Menne and Williams 2009), affect trends here as well. Most of the hourly observations in the 1930s are located in the coastal and southern regions of the United States and are therefore likely to introduce some region-based biases into the earliest data averages, especially during the coldest months.
Humidity-related variables show that more-subtle changes are occurring in the United States when compared with temperature. Trends and variability in Td and q are closely related, and both measures have experienced no long-term change in the United States. These trends also appear to be less sensitive to adjustments for inhomogeneities. In contrast, regional moistening that is highest in the Midwest has occurred over recent times (1947–2010). A geographic divide in Td and q trends indicates a drying in the west and moistening in central and eastern United States is apparent since the 1980s. Similar patterns in the United States have been seen in global analyses (Dai 2006; Willett et al. 2008). This regional drying pattern is consistent with changes related to limited ocean moisture supply (Simmons et al. 2010).
Although the magnitude of the average decrease in RH since 1930 needs to be viewed cautiously because of data-homogeneity concerns, this decrease continues in the post-1950 period, for which data veracity is more certain. This decrease appears to be driven by significant warming of T since there has been little change in Td and q. An attempt to use adjusted T and Td values resulted in little change to the RH values. Reanalysis data show similar trends over recent periods (Simmons et al. 2010; Lu and Takle 2010). A number of global-scale analyses show increases in q and no change in RH (Dai 2006; Trenberth et al. 2007; Willett et al. 2008). From 1947 to 2010, warming in T commonly corresponds to a moistening in q (Fig. 8). This relationship is contrary to the longer-term U.S. trends. As noted by Robinson (2000), however, trends in humidity are extremely sensitive to the observation period, and these consequently affect comparisons with other studies. The recent large drop in humidity in the most recent years, also seen in other studies (e.g., Willett et al. 2008; Simmons et al. 2010), may have an undue effect on trends.
Since 1950, fluctuations in RH show characteristics that are analogous to those seen in Td and q, with peak wetness in the late 1990s (Fig. 6). Drying of RH in the western United States is also apparent in global studies of trends in atmospheric moisture by Willett et al. (2008) and Simmons et al. (2010). The most pronounced drying occurs in winter and spring, a feature that has also been seen in Canada (Vincent et al. 2007). Although there is little seasonal spatial coherence nationally, there is a wintertime reduction in RH along the Eastern Seaboard and a summertime increase in the Midwest and Northeast.
c. Midwest moistening
Documented effects of irrigation (Mahmood et al. 2008) and land-use change (Sandstrom et al. 2004; Changnon et al. 2006; Diffenbaugh 2009) on Td in the central United States and the Midwest suggest that these changes are contributing to moistening trends. In the present analysis, these areas also experience cooling of maximum temperatures while the atmosphere is moistening. Such trends concur with the work of Mahmood et al. (2004) who found that July and August maximum temperatures showed cooling in irrigated areas. Further evidence of the influence of agriculture is provided by the peaks in q occurring during summer when evapotranspiration and therefore irrigation are at their height. Sandstrom et al. (2004) concluded that advection from the Gulf of Mexico did not affect atmospheric moisture in the Midwest as the South has experienced little change; instead regional moisture sources were causing moisture increases. These conclusions are also valid on the basis of this study, as the South does not show the moisture increases that are similar to those of the Midwest.
Large-scale circulation changes also impact the frequency and moisture content of airstreams, resulting in interdecadal changes in Td on regional scales (Schwartzman et al. 1998; Robinson 2000). Within the central and eastern region of the United States where moistening is occurring, a closer inspection of trends reveals that stations in southern and inland eastern areas are drying (Fig. 7). These stations that show anomalous drying are generally not located near any major sources of moisture and are not located in major irrigated agricultural regions. Simmons et al. (2010) inferred that the recent decrease in near-surface RH over land is due to limited oceanic moisture sources, and this rationale is consistent with the changes that have occurred in the United States.
A lack of spatial coherence in trends can be seen in some of the atmospheric moisture variables. Of particular note are the strong trends seen at Phoenix and Reno, where the strongest warming and drying occurs in T and q, respectively. These two stations exemplify the effects of urban environments, where warmer temperatures and lower humidity prevail relative to rural environments. Both of these locations have experienced increased urbanization, with the urban heat island in Phoenix being particularly well known. Changing land use has been implicated in driving these changes in Phoenix (Balling and Brazel 1986). In addition, the large proportion of impermeable surfaces in cities may serve to limit moisture availability.
Atmospheric moisture trends may also reflect changes in precipitation as a local source of moisture. Schwartzman et al. (1998) suggested that the relationship between precipitation and increased dewpoints is enhanced through convective cycle/precipitation changes. They theorize that lack of evidence for positive feedbacks with increasing temperature may be due to diurnal moistening and drying effects, which are masked by analysis of average trends over longer time scales. Diurnal changes in both temperature and atmospheric moisture are likely to have occurred and may elucidate the changes found in this work. Future work will feature a more detailed investigation of changes at the hourly resolution as well as the effects of different weather conditions.
A homogeneity assessment of hourly U.S. meteorological observations indicates that there is a cold bias in the temperatures during early periods when psychrometers were in use. When adjustments are made for these biases, the rate of warming in the temperature series is tempered, although the effect of the trend magnitude associated with the adjustments is generally small. Inhomogeneities associated with metadata were more consistently detected in dewpoints, but they had no consistent effect on trends. Relative humidity series computed from the hourly temperature and dewpoints showed marked inhomogeneity prior to 1950. Preliminary attempts to adjust this bias on the basis of the adjusted temperature and dewpoint series did not correct this inhomogeneity.
Hourly temperatures in the United States have warmed since 1930, with much of this warming occurring since 1980. Although warming is a feature of all seasons, the increases in temperatures during spring are strongest. The Southeast is the only region of the United States that consistently displays a lack of warming trends. Drying trends were found in average U.S. relative humidity measurements in the post-1950 period, for which the data biases are less problematic. Urbanization does not appear to have biased the observations, despite contributing to enhanced warming and drying trends. There have been no long-term changes in dewpoint temperatures or specific humidity but rather there has been a decreasing (1947–79) and then an increasing (1980–2010) trend in both variables. Increases in atmospheric moisture during the summer are seen in the Midwest. These regional changes in atmospheric moisture appear to be related to changes in land use and moisture availability.
This project was supported by NOAA Grant NA07OAR4310061. Partial support was also obtained through NOAA Contract EA133E07CN0090. Support for the Twentieth Century Reanalysis Project dataset is provided by the U.S. Department of Energy’s Office of Science Innovative and Novel Computational Impact on Theory and Experiment (DOE INCITE) program and Office of Biological and Environmental Research (BER) and by the NOAA Program Office. We thank the anonymous reviewers who provided thoughtful suggestions that improved this manuscript.
Current affiliation: Department of Atmospheric Science, Colorado State University, Fort Collins, Colorado.