Temporal and spatial trends in air temperature on the Island of Oahu, Hawaii


  • Mohammad Safeeq,

    Corresponding author
    1. Department of Natural Resources and Environmental Management, University of Hawaii at Mānoa, Honolulu, HI, USA
    Search for more papers by this author
    • Present address: College of Earth, Ocean, and Atmospheric Sciences, Oregon State University, 104 CEOAS Administration Building Corvallis, OR, USA
  • Alan Mair,

    1. Department Geology & Geophysics and Water Resources Research Center, University of Hawaii at Mānoa, Honolulu, HI, USA
    Search for more papers by this author
  • Ali Fares

    1. Department of Natural Resources and Environmental Management, University of Hawaii at Mānoa, Honolulu, HI, USA
    Search for more papers by this author


We examined trends in minimum and maximum temperatures in the Oahu during the period of past 39 (1969–2007) and 25 (1983–2007) years. We found a strong spatial and temporal variability in the temperature trends on Oahu. During the past 39 years, island-wide minimum temperature has increased by 0.17 °C/decade and shows a considerable variability in trends at individual location. There was no detectable trend found in maximum temperature over the same time period. The year 1983 was identified as the change point in the island-wide minimum temperature. During the recent 25 years annual and summer maximum temperature showed a decline while minimum temperature continued to increase. Trend in diurnal temperature range (DTR) shows a decline during the past 39 years with a stronger decreasing trend during the recent 25 years. The trend in DTR for Oahu is much higher compared to the global DTR trend indicating a rapid warming in minimum temperature. Extreme temperature indices show a general warming during the past 39 years. There has been significant increase in tropical and warm nights at the two urban stations. Maximum temperature generally followed the Pacific Decadal Oscillation (PDO) except the period when there is an increase in Hawaii Rainfall Index (HRI). In contrast, minimum temperature showed better agreement with HRI compared to the PDO, at least up until 1999 after which it showed an increase. Despite the relative cooling in PDO during the recent decade an increase in minimum temperature can be attributed to a decline in HRI.

1. Introduction

Recent worldwide concerns among the scientific community on global climate changes have triggered interest to study the retrospective and projected global changes in surface air temperatures (IPCC, 2007). While the most dramatic warming is projected to occur over continental land masses and extreme northern latitudes, increases in surface air temperature in tropical settings like the Hawaiian Islands are also expected, which could have significant effects on terrestrial and coral ecosystems, water resources, agriculture, and economic health. The geographic isolation from major landmasses makes Hawaii vulnerable to rising temperature.

Many existing studies have already shown an overall warming in Hawaii (Nullet and Ekern, 1988; Karl et al., 1996; Giambelluca et al., 2008; Diaz et al., 2011). Giambelluca et al. (2008), hereafter referred to as GDL, conducted the most comprehensive analysis of state-wide temperature trends to date by deriving a Hawaii Temperature Index (HTI) using 21 low-elevation (<800 m) and high-elevation (>800 m) stations on five islands to analyze state-wide trends in minimum (Tmin), maximum (Tmax), and mean surface air temperature (Tavg) during 1919–2006. Their analyses showed state-wide warming rates in Tmin, Tmax, and Tavg of 0.094, 0.005, and 0.043 °C/decade, respectively. During 1975–2006, the rates for Tmin, Tmax, and Tavg increased dramatically to 0.275, 0.054, and 0.164 °C/decade, respectively. Most of the observed increase in Tavg was attributed to a much larger increase in Tmin compared to Tmax (Tmin increased five times faster during 1975–2006), resulting in a reduction of the diurnal range. GDL reported enhanced warming at elevations >800 m during 1975–2006 at rates of 0.441, 0.085, and 0.268 °C/decade for Tmin, Tmax, and Tavg, respectively. Lower but still elevated rates of 0.153, 0.032, and 0.087 °C/decade for Tmin, Tmax, and Tavg, respectively were reported for elevations <800 m during 1975–2006. The rapid rise in Tavg warming rates for Hawaii in recent decades is consistent with the rise of global Tavg warming rates, which increased from 0.074 °C/decade during 1906–2005 to 0.130 °C/decade since the 1950s (IPCC, 2007).

GDL reported that urbanization in Hawaii has dampened daytime warming (Tmax), enhanced nighttime warming (Tmin), and lessened the mean overall warming rate (Tavg). However, Karl et al. (1996) reported a warming rate for Tavg of 0.267 °C/decade at Honolulu airport during 1900–1990, which is more than six times the long-term state-wide rate reported by GDL and more than three times the long-term global rate (IPCC, 2007). The area around Honolulu airport has experienced significant urbanization during 1900–1990 and the high warming rate for Tavg suggests that urbanization effects enhanced overall warming at this location, which is counter to the finding of GDL.

The temperature variation before the late 1980s was tightly coupled to the Pacific Decadal Oscillation (PDO), which implies a strong linkage between air temperature and regional sea surface temperature (SST) (Giambelluca et al., 2008). Indeed, Nullet and Ekern (1988) reported warming trends at Honolulu, Oahu, and Hilo, Hawaii from the 1950s to 1980s that were consistent with increases in regional SST. Secular warming since the 1980s has dominated so that despite recent cooling of the PDO, surface air temperature has remained elevated (Giambelluca et al., 2008). However, Diaz et al. (2011) noted that SST in the near-shore waters around the island of Oahu during 1956–2008 warmed at a rate (0.194 °C/decade) that was greater than the rate recorded over open ocean areas away from the Hawaiian Islands but similar to the Tavg rates reported by GDL. Thus, the recent secular warming observed by GDL may have been induced, at least in part, by near-shore SST trends. Diaz et al. (2011) also reported a rise in the freezing level surface on the upper slopes of the islands of Maui and Hawaii, and an upward shift in the lifting condensation level (LCL) since 2000. Cao et al. (2007) noted increased persistence of the trade-wind inversion (TWI) since 1979 at Lihue, Kauai, and Hilo, Hawaii, and detected a weakly significant lowering of the TWI base height at Lihue, Kauai. Thus, the observed warming trends in Hawaii, particularly at higher elevations, are consistent with the upward trend in near-shore SST, increased persistence of the TWI, downward shift in the TWI height, and rise in the local LCL.

Regional climate assessment in Hawaii remains a difficult task because of high spatial variability in air temperature despite the small land area. The existing studies of Tmin, Tmax, and Tavg show trends using either an overall state-wide index based on multiple stations (Giambelluca et al., 2008) or an analysis of only one or two selected locations (Nullet and Ekern, 1988; Karl et al., 1996). A comparison of intra-island temperature trends that further investigates the effects of urbanization has yet to be conducted Hawaii. In terms of climate extremes which are known to have higher economic and social impacts in the United States (Kunkel et al., 1999), existing studies in Hawaii have mostly focused on extreme precipitation (Chu et al., 2010). Trends in extreme temperature indices, particularly in low-elevation (<800 m) areas, are largely unknown.

While the multi-station temperature index based analysis used by GDL is useful in determining the regional trend, it does not provide additional explanation of spatial variability. Similarly, temperature trends drawn from only one or two stations (Karl et al., 1996; Nullet and Ekern, 1988) can be influenced by microclimate. Thus, there is a need to perform the retrospective trend analysis on measured air temperature including the extremes and characterize how the trend varies spatially. The specific objectives of this study are to: (1) evaluate the spatial and temporal trends in measured air temperature and selected extreme indices on the island of Oahu and (2) quantify the relationship between local temperature and regional climate indices, i.e. Hawaii rainfall index (HRI), PDO, and Niño-3.4.

2. Materials and methods

2.1. Study area and data

The island of Oahu, Hawaii is located between 21°5′N–157°35′W and 21°45′N–158°20′W (Figure 1). Oahu is the third largest Hawaiian island with a total land area of 1553 km2 and an elevation ranging from sea level to over 1226 m at the top of Mt. Ka'ala. The island's residential population has grown from 500 000 in 1960 to 950 000 in 2010 (State of Hawaii, 2012; US Census Bureau, 2010). Agricultural and urban lands (impervious) occupy 6.5 and 14% of the total area, respectively (NOAA-CCAP, 2005). Tavg at Honolulu is 27.4 °C in August, the warmest month, and 22.4 °C in January, the coolest month (Lau and Mink, 2006). The northeastern trade winds dominate the surface airflow. Oahu's annual rainfall is mostly dictated by topography and location with respect to trade winds (leeward vs windward). The highest annual precipitation (>7000 mm) occurs over the Ko'olau mountain range near Kahana and the least (525 mm) near the coast (Giambelluca et al., 2011).

Figure 1.

Map of the Island of Oahu and the geographic locations of the climate stations used in this study.

Daily measurements of Tmin and Tmax collected during 1969–2007 from 12 stations across the island of Oahu were downloaded from the National Climate Data Center (NCDC) (Figure 1) (http://www.ncdc.noaa.gov/oa/climate/stationlocator.html). Relevant information about the selected stations including map ID, name, state key number, latitude, longitude, elevation, and length of record are presented in Table 1. These 12 stations were selected based on length of record and amount of missing data. Among the 12 stations, only stations 3 and 6 were part of the earlier study by GDL; station 3 was also used in the study by Karl et al. (1996). These earlier studies used datasets that extended back to 1900 (Karl et al., 1996) and 1919 (Giambelluca et al., 2008). However, we chose to confine our analysis to recent time periods (i.e. 1969–2007) where time series data from five or more stations are available.

Table 1. Locations, elevation and duration of record for the 12 climate stations used in the study
Station IDaCOOP IDSKNStation nameLatitude (°N)Longitude (°W)Elevation (m)Record duration
  1. SKN, state key number.

  2. a

    Only stations in bold were used to calculate OTI and OTA and also for analysis of trends in extreme indices.

1510305841.16CAMP ERDMAN21.58−158.181.51983–2007 (25)
2511918702.2HONOLULU OBSERV21.32−158.001.51983–2007 (25)
3511919703HONOLULU INTL AP21.32−157.932.11969–2007 (39)
4514500911KII-KAHUKU21.70−157.984.61983–2007 (25)
5516128785.2MANOA LYON ARBO21.33−157.80152.41983–2007 (25)
6517150870OPAEULA21.58−158.03304.81969–2007 (39)
7518500725.6PUU MANAWAHUA21.38−158.12509.91983–2007 (25)
8519195847WAIALUA21.57−158.129.81969–2000 (32)
9519397717.2WAIKIKI21.27−157.8231969–2007 (39)
10519523795.1WAIMANALO EXP FARM21.33−157.7218.31969–2007 (39)
11519603892.2WAIMEA ARBORETUM21.63−158.0512.21983–2007 (25)
1299999-22519840KANEOHE BAY MCAS21.45−157.7831969–2007 (39)

On average, about 10% of the daily Tmin and Tmax data were reported as missing. Missing data were not estimated because interpolation for quality control correction and missing data estimation can introduce errors into a temperature dataset (Robeson, 1994). However, data were excluded from the dataset when Tmax was less than the Tmin during the same day. Unusual high and low values were compared with data of nearby stations and subsequently removed in case of inconsistency.

We also downloaded daily Tmin and Tmax measurements collected during 1919–2007 for 17 of the 21 stations used by GDL including their four Oahu stations and their 13 outer island low-elevation (<800 m) stations (locations and station information not presented). The period 1971-2000 was used in this study as a common base period, whenever feasible. However, calculation of monthly temperature anomalies using the calendar month means during 1971–2000 was not possible for the gages used by GDL as several did not have data during this period. Instead, we calculated the monthly temperature anomalies at these stations using the calendar month means during 1944–1980. Missing data at these stations were also not estimated.

2.2. Trend and shift analysis

Long-term trends and shifts in Tmin, Tmax, and Tavg (calculated as the arithmetic mean of Tmin and Tmax) were analyzed at five individual stations (stations: 3, 6, 9, 10, and 12) with concurrent data during 1969–2007 (39-year period) (Table 1). Along with trends at individual stations, an analysis of island-wide trend in Tmin, Tmax, Tavg, and diurnal temperature range (DTR) was also performed. DTR was calculated as the difference between maximum and minimum annual temperature anomalies. For the island-wide trend analysis, we added a sixth station (station 8) with a shorter length of record (1969–2000). For each station, monthly temperature and DTR anomalies were calculated as departures from the calendar month means during 1971–2000. The time series of monthly temperature and DTR anomalies from the six stations were averaged to calculate an Oahu Temperature Index (OTI). The approach used to compute the OTI is similar to the approach used by GDL to compute the state-wide HTI. For comparison, we also computed two additional indices based on subsets of stations used in the analysis by GDL. First, we computed a monthly index based only on GDL's Oahu stations, which we refer to as GDL-O. Next, we computed a monthly index based only on GDL's outer island low-elevation (<800 m) stations, which we refer to as GDL-NO.

Long-term secular trends and shifts for individual stations (5 stations), OTI (6 stations), GDL-O (4 stations), and GDL-NO (13 stations) were analyzed on an annual and seasonal (summer and winter) basis during 1969–2007 (39-year period). We used a calendar year to group annual data with a summer period extending from May to October, and a winter period extending from November to April. Short-term trends and shifts in Tmin, Tmax, and Tavg during 1983–2007 (25-year period) were also analyzed at each of the 12 stations, station 8 is only gage that does not extend to 2007: five long-term stations plus seven additional short-term stations (Table 1). No OTI, GDL-O, or GDL-NO indices were computed for the short-term analyses during 1983–2007.

The nonparametric Mann–Kendall test for trend (MK-1) and Sen's slope estimates were used to assess trends in Tmax and Tmin (Mann, 1945; Kendall, 1955; Sen, 1968). The MK-1 test was used for determining the significance of a monotonic trend, while Sen's method was used for quantifying the magnitude of the trend. In this study, a two-tailed test was used to test for statistically significant upward or downward monotonic trends at a significance level, α, of 0.05 and 0.10. For statistically significant trends (p ≤ 0.10 for two-sided MK-1 test), we also performed trend-free pre-whitening (TFPW) on time series data and recomputed the MK-1 and Sen's method tests. For significant trends (p ≤ 0.10), we also used the sequential version of the Mann–Kendall test (MK-2) to identify change points in Tmax and Tmin (Sneyers, 1990). The MAKESENS computer programme was used to aid the trend analyses (Salmi et al., 2002).

2.2.1. MK-1: nonparametric Mann–Kendall test for trend

The null hypothesis of no trend, H0, is tested against the alternate hypothesis of a monotonic (i.e. continuous) increasing or decreasing trend, Hl (Salmi et al., 2002). MAKESENS uses two different approaches to test for trend based on the number of observations. If the number of observations is less than 10, MAKESENS uses the S statistics (Gilbert, 1987); otherwise, it uses Z statistics (normal distribution). The MK-1 test statistic S and its variance are calculated using the following equations (Salmi et al., 2002):

display math(1)

where xj and xk are the annual values in years j and k, respectively, and j > k, and

display math(2)

The variance of S is calculated using the following equation:

display math(3)

where q is the number of tied groups and tp is the number of data values in the pth group.

The values of S and var(S) are used to compute the test statistics Z as follows:

display math(4)

The presence of a statistically significant trend is evaluated using the Z value. The Z value is a normally distributed random variable, and a positive Z value indicates an increasing trend. When the monotonic trend is either upward or downward, H0 is rejected if the absolute value of Z is greater than Z1-α/2 in a two-tailed test at α level of significance.

2.2.2. Sen's slope estimates

The MK-1 test gives information on the trend, but it does not quantify the magnitude of the change. To estimate the slope, the Sen's nonparametric method is used. Sen's slope for a monotonic increasing or decreasing time (t) series f (t) was calculated as:

display math(5)

where Q is the slope of the trend f(t) and B is the intercept. To determine Q in Equation (5), slopes between each data pair are calculated first using the following equation:

display math(6)

If there are n values of xj in the time series, then there will be N = n (n − 1)/2 slope estimates of Qi. The Sen's estimator of slope is the median of these N values of Qi.

The advantages of using the nonparametric MK-1 test and Sen's slope estimates are that missing values are allowed and the data need not to confirm to any particular distribution. In addition, commonly used simple linear regression for trend analysis is vulnerable to gross errors and sensitive to non-normality of the parent distribution (Sen, 1968). Further, the Sen's method is not greatly affected by gross data errors or outliers because it is derived from the median of all slopes as explained earlier.

2.2.3. Trend-free pre-whitening

TFPW is a technique for removing the effects of serial correlation on time series analyses. However, studies have shown that TFPW can also remove some of the trend from a time series with autocorrelation (Yue et al., 2003). In this study, a modified version of the TFPW technique by Burn et al. (2004) was applied to a time series when the trend estimated prior to pre-whitening using MK-1 was statistically significant (p ≤ 0.10 for a two-sided MK-1 test). TFPW was performed as follows:

After estimating the monotonic trend, Q, using Equation (5), each time series f(t) was detrended using:

display math(7)

Lag-one serial correlation r1 was calculated using yt; if the value of the serial correlation r1 is significant at the 0.05 level, the pre-whitening was performed through:

display math(8)

Monotonic trend Q was added after pre-whitening:

display math(9)

The MK-1 test statistics and Sen's slope estimate were recalculated using the temperature time series yt″.

2.2.4. MK-2: sequential Mann–Kendall test for shift

To identify the change point (or year) in a time series with a highly significant monotonic trend (p ≤ 0.10 for two-sided MK-1 test), the sequential version of MK test based on rank statistics (MK-2) was used (Sneyers, 1990). The MK-2 test uses the rank of time series data. A parameter ti is generated by comparing the xj preceding each term xi (i > j), such that xi > xj. The sequential values of the forward u(ti) and the backward u′(ti) Gaussian normal variate were estimated using the mean and variance of ti as described by Kadioglu (1997). Both u(ti) and u′(ti) Gaussian normal variates are normally distributed and their sequential behaviour fluctuates around zero. The intersection of the curves u(ti) and u′(ti) localizes the change and allows the identification of the year when a trend or a change starts. Points where the two lines for u(ti) and u′(ti) intersect are considered as approximate potential trend change points. When either the absolute values of u(ti) or u′(ti) exceeds a certain confidence limit before or after the crossing points (e.g. Z = ±1.96 for 95% confidence level), the intersection point is considered significant.

2.2.5. Indices of local climate extremes

To characterize the impact of changes in temperature on island climate, we analyzed the trend for a total of 13 climate change indices (Table 2) using daily Tmin and Tmax data from the six stations (3, 6, 8, 9, 10, and 12) (Table 1) during 1969–2007 (except station 8 which ended in 2000). We selected indices that reflect climate impacts on residential heating and cooling, ecological consequences, as well as indices related to temperature extremes. Specifically, we estimated summer days (SU25), tropical nights (TR20), annual extremes of Tmin (TNx and TNn) and Tmax (TXx and TXn). We also estimated changes in cold and warm spells as manifested by the cold spell duration indicator (CSDI) and warm spell duration indicator (WSDI), DTR, as well as the number of warm (cold) nights/days, defined as the number of days when the daily Tmin/Tmax was higher (lower) than the 90th (10th) percentile of the Tmin/Tmax during the 1971–2000 base period (Table 2). SU25 and TR20 were estimated using base temperatures of 25 and 20 °C, respectively. CSDI (WSDI) was estimated as the annual number of periods with at least six consecutive days when the daily Tmin (Tmax) was lower (higher) than the 10th (90th) percentile based on the 1971–2000 base period. The percentile, frequency, and magnitude based indices have been recommended by ETCCDMI (http://cccma.seos.uvic.ca/ETCCDMI/) and analyzed for trend globally (Zhai et al., 1999; Klein-Tank and Konnen, 2003; Vincent et al., 2005). Detailed descriptions of these indices are given by the World Climate Research Programme (WCRP, 2010). Calculation of indices and trend analysis were performed using the RClimDex package (available for download from http://cccma.seos.uvic.ca/ETCCDMI).

Table 2. Selected extreme indices of climate change
IDIndicator nameDefinitionsUnit
SU25Summer daysAnnual count when TX (daily maximum) >25 °CDays
TR20Tropical nightsAnnual count when TN (daily minimum) >20 °CDays
TXxMaximum of TmaxAnnual maximum value of daily Tmax°C
TNxMaximum of TminAnnual maximum value of daily Tmin°C
TXnMinimum of TmaxAnnual minimum value of daily Tmax°C
TNnMinimum of TminAnnual minimum value of daily Tmin°C
TN10pCool nightsPercentage of days when TN < 10th percentile%
TN90pWarm nightsPercentage of days when TN > 90th percentile %%
TX10pCool daysPercentage of days when TX < 10th percentile %%
TX90pWarm daysPercentage of days when TX > 90th percentile %%
WSDIWarm spell duration indicatorAnnual count of days with at least six consecutive days when TX > 90th percentileDays
CSDICold spell duration indicatorAnnual count of days with at least six consecutive days when TN < 10th percentileDays
DTRDiurnal temperature rangeMonthly mean difference between TX and TN°C

2.3. Comparison to regional indices of climate

Previous studies reported a link between temperature trends and some weather indices, i.e. PDO and El Niño-Southern Oscillation (ENSO) (Giambelluca et al., 2008; Chu et al., 2010). To characterize the temporal variation in Tmin and Tmax, we used PDO indices, defined by Mantua et al. (1997) (available online: http://jisao.washington.edu/pdo/) and Niño-3.4 index defined by Trenberth (1997) (available online: http://www.cgd.ucar.edu/cas/catalog/climind/TNI_N34/). The negative (cold) PDO phase prevailed during 1890–1924, 1947–1976, 1999–2002, and 2006-May 2009, while the positive (warm) PDO phase dominated during 1925–1946, 1977–1998, and 2003–2005 (Chu and Chen, 2005; CIG, 2009). Since 1999, the decadal cycles in the PDO have broken down, and it is not yet clear if the 1998 shift was a true shift to a negative phase (CIG, 2009). In addition, the monthly HRI (Chu and Chen, 2005) was used to compare Tmin and Tmax with regional rainfall. For comparing rainfall variations in the Hawaiian Islands with ENSO, Chu and Chen (2005) used the Niño-3.4 index because it is more representative of ENSO than other indices (Barnston et al., 1997). Chu and Chen (2005) reported a significant negative correlation between HRI and PDO. In addition, they found that winter rainfall anomalies tend to be enhanced during El Niño/+PDO (low rainfall) and La Niña/−PDO (high rainfall) conditions. Their results confirm that regional SST has a significant effect on Hawaiian rainfall.

For comparison with regional climate indices (e.g., PDO, Nino-3.4, HRI), a separate Oahu Temperature Anomaly (OTA) was computed using Oahu's six stations (Table 1). First, monthly Tmin, Tmax, and Tavg time series data were standardized to produce time series with zero mean and unit variance as follows:

display math(10)

where Ti,j* is the standardized air temperature value for year i and month j, Ti,j is the original monthly temperature time series value, math formula is the mean and σT the standard deviation of the original time series. The residual or the trend-free stochastic component of the original time series for year i and month j, Zi,j*, was then calculated by subtracting the computed mean standardized values for month j, Tj*, from each standardized value, Ti,j*, as follows:

display math(11)
display math(12)

The monthly residual time series of Tmin, Tmax, and Tavg from the six stations (3, 6, 8, 9, 10, and 12) were then averaged to calculate the island-wide monthly OTA.

2.3.1. Trend visualization, cross-correlation, and wavelet analysis

The OTA was used to visualize the relationship between Oahu's air temperature and regional rainfall (HRI) and SST (PDO) by plotting the cumulative rescaled departures as a function of time (Garbrecht and Fernandez, 1994). During extended periods of above average temperature, rainfall, or SST, the cumulative departures increase with time (positive slope), whereas during extended periods of below average temperature, rainfall, or SST, the cumulative departures decrease with time (negative slope). Plots of cumulative departures are analogous to mass diagrams for storage. Periods of increasing and decreasing values of the cumulative departures are respectively analogous to periods of increasing and decreasing storage. However, our approach differs from that of Garbrecht and Fernandez (1994) in that station data representing different time periods were combined to compute the OTA, and indices used to compute the HRI and PDO were derived from different base periods.

To understand the relationship between the OTA and the HRI, PDO, and Niño-3.4 over time, cross-correlation and wavelet analysis were performed. Cross-correlation is a measure of correlation between values in two different series separated by variable time lags. Values of the cross-correlations were computed between the OTA indices of Tmin and Tmax and the indices of HRI, PDO, and Niño-3.4 using SAS 9.2 (SAS Institute, Inc, Cary, NC, USA). Cross-correlations were considered to be significant if they fell outside the ±2 standard error (∼95% confidence interval) against the null hypothesis of no correlation. Annual OTA indices of Tmin, Tmax, and the HRI were used to explore the temporal variability during 1950–2007 using wavelet analysis (Torrence and Compo, 1998). Wavelet analysis provides the temporal variability in the climate signal (e.g. temperature) and has been widely used for analysing the localized variation of power within the time series (Torrence and Compo, 1998).

3. Results and discussion

The station average Tmin and Tmax over the duration of record show strong variability with elevation (Figure 2a). More than 50% of the spatial variability in Tmin (R2 = 0.57, p < 0.001) and Tmax (R2 = 0.53, p < 0.001) can be explained by elevation. Although both Tmin and Tmax show strong linear relationship with elevation, there is a significant variability among stations located at lower (<25 m) elevations (Figure 2b). The Tmin and Tmax decrease by 0.0084 and 0.0054 °C m−1, respectively. The average temperature lapse rate (0.0069 °C m−1) is consistent with average temperature lapse rate of 0.0064 °C m−1 measured at the Island of Hawaii (Juvik and Nullet, 1994) and 0.006 °C m−1 measured at the island of Maui (Loope and Giambelluca, 1998). However, there is a significant difference in the lapse rates of Tmin and Tmax. Temperature lapse rate derived from PRISM data (Daly and Halbleib, 2006) is consistent with observed data and showed similar variability in temperature among the stations at lower elevation.

Figure 2.

Minimum (circle) and maximum (diamond) temperatures lapse rate with elevation.

3.1. Long-term trends and shifts (1969–2007)

3.1.1. Tmin and Tmax

Long-term trends in Tmin exhibited highly significant annual and seasonal warming (p ≤ 0.05) at stations 3 and 9 (Table 3). Both of these locations are located in highly urbanized leeward areas of Honolulu (Figure 1). Highly significant summer warming in Tmin was also evident at station 10, a suburban windward location. Non-significant warming in Tmin was reported at station 6, a rural site, and station 12, a windward coastal site largely surrounded by ocean. Although three of the five long-term stations showed significant warming, the warming rates vary substantially. The rate of warming in annual Tmin at stations 3 and 9 is nearly threefold higher compared to station 10. The OTI for Tmin also showed highly significant annual and summer warming at rates that exceed both the GDL-O and GDL-NO. However, the GDL-O did report significant summer warming in Tmin (p ≤ 0.10) which is consistent with our findings. The low-magnitude non-significant annual and seasonal trends in the GDL-NO for Tmin suggest that other low-elevation sites in Hawaii did not experience similar nighttime warming. Indeed, the greater rates of warming near Oahu's high population centres indicate that urbanization has enhanced nighttime warming (Tmin). The OTI for Tmin during the summer season showed an increase of 0.250 °C/decade during 1969–2007 which is higher than the increase in low-elevation HTI at a rate of 0.119 °C/decade during 1975–2006 reported by GDL. However, seasonal warming trends at all five long-term Oahu stations were greater in the summer than in the winter (Table 3), which contrasts with the results of GDL who found warming trends in winter, in general, were higher than in the summer. Indeed, the warming rates for GDL-O and GDL-NO were also higher in the summer than in the winter during this period, which suggests that warming was greater during the summer at low elevations (<800 m) across the state. The discrepancy in seasonal warming rates may be due to the more rapid warming at higher elevations (>800 m) reported by GDL.

Table 3. Seasonal (summer and winter) and annual trends (°C year−1) for minimum and maximum temperatures at five long-term stations. Island-wide temperature trends were calculated on spatially average temperature index. Dark grey represents the significant trend at p ≤ 0.05, and light grey at p ≤ 0.10 significance level
  1. a

    Trends during 1969–2007 period calculated using the four low elevation (<800 m) Oahu gages persented in GDL.

  2. b

    Trends during 1969–2007 period calculated using the 13 low elevation (<800m) non-Oahu gages presented in GDL.

  3. Sen's slope estimates after pre-whitening are presented between brackets.


Long-term trends in Tmax showed non-significant warming and cooling in the five long-term Oahu stations, OTI, and GDL-O (Table 3). However, highly significant annual and winter cooling in Tmax (p ≤ 0.05) was evident in the GDL-NO. The results suggest that the low-elevation, urbanized locations on Oahu experienced less cooling in Tmax than other locations across the state, and are contrary to the findings of GDL who found that urbanization dampened daytime warming (Tmax).

Rapid warming for Tmin resulted in a highly significant decreasing trend 0.18 °C/decade for DTR (p ≤ 0.05) (Figure 3). This is consistent with the findings by GDL who also showed decline in DTR during 1919–2006 as a result of rapid increase in HTI for Tmin compared to the Tmax. Also, the more rapid long-term warming in Tmin compared to Tmax is consistent with the global temperature trend (IPCC, 2007). The differential decreasing trend in Tmax and Tmin (i.e. Tmax trend − Tmin trend), which is comparable to with the trend in DTR calculated from MK-1, (−0.17 °C/decade) is greater than the decline for GDL-O (−0.11 °C/decade) and GDL-NO (−0.12 °C/decade). However, the decline in DTR from the differential trend in GDL-NO for Tmax and Tmin is resulting from a decline in Tmax as opposed to an increase in Tmin. The higher decreasing trend in DTR from OTI compared to those reported by GDL can be attributed a higher increase in OTI for Tmin compared to GDL-NO during the concurrent 1969–2007 period (Table 3). Comparing with the long-term DTR trend, the magnitude of decreasing trend in OTI for DTR is nearly twofold larger than those reported by GDL for low elevation (<800 m) during 1919–2006. Also the magnitude of decreasing trend in OTI for DTR is sixfold larger than the global DTR trend (−0.032 °C/decade) during 1976–2004 and three times as large as the DTR trend during 1950–2004 (Vose et al., 2005).

Figure 3.

Trend analysis results from nonparametric Mann–Kendall test (MK-1) and sequential version of Mann-Kendall test based on rank statistics (secondary y-axis) (MK-2) for Tmax (a) and Tmin (b) at five long-term stations. Slopes marked with asterisk sign indicate significant trend at p ≤ 0.05 except for Tmin trend at station 10 where it represent p ≤ 0.10 significance level.

The analysis of change points from the MK-2 test showed the beginning of upward trend in Tmin began as early as in 1993 at station 3 (Figure 4). The year 1988 was identified as the change point in the upward trend in Tmin at station 9. This suggests that not only the magnitude (i.e., slope of trend) of Tmin warming at station 9 is large (∼20%) compared to station 3, warming began 5 years earlier. The shift in Tmin at the station 10 (suburban windward location) began in 1983, 5 and 10 year prior to urban leeward stations 3 and 9, respectively. An earlier change point at a suburban location (station 10) compared to the two urban locations (stations 3 and 9) indicates that factors other than urbanization are responsible for the beginning of the shift in Tmin. The shift in Tmin could have been a result of the beginning of the most recent warming in global near-surface temperature in 1976 (Folland et al., 2001) which also coincides with change (from negative to positive) in the phase of PDO in 1977. In addition, Whysall et al. (1987) reported an easterly shift in the trade winds over the tropical pacific between 1950 and 1981, which could have brought the warm equatorial water into the region by increasing the southerly component of the North Equatorial Current (Nullet and Ekern, 1988). Similar to station 10, change point in OTI for Tmin also began in 1983 (Figure 3). Whereas 1976 was identified as the change point in OTI for Tavg which was also the start of most recent global near-surface temperature warming. The year 1999 was identified as the change point in the downward trend in OTI for DTR indicating that much of the trend in DTR has been recent.

Figure 4.

Trend analysis results from nonparametric Mann–Kendall test (MK-1) and sequential version of Mann–Kendall test based on rank statistics (secondary y-axis) (MK-2) for island-wide maximum (a), minimum (b), and average (c) temperatures and DTR anomalies (d). Slopes marked with asterisk sign indicate significant trend at p ≤ 0.05.

3.2. Recent trends and shifts (1983–2007)

3.2.1. Tmin

The magnitude and direction of recent trends in Tmin are spatially variable (Figure 5). Among the 12 stations, 6 stations showed significant increasing trend and only station 12 showed significant decreasing trend (Table 4). The highest warming in Tmin (1.02 °C/decade, p ≤ 0.10) was at station 7, a rural leeward site (Figure 1). The lowest significant warming in Tmin (0.29 °C/decade, p ≤ 0.05) was at station 5, a suburban leeward site. While stations 1, 4, and 6 showed non-significant decline (p > 0.10), a significant increasing trend in Tmin (0.54 °C/decade, p ≤ 0.05) was observed at nearby station 11. These results do not show any consistent pattern of increasing and decreasing trend between leeward and windward sites (Figure 5). However, with one exception (station 11), all the five stations (2, 3, 5, 7, and 9) showing significant (p ≤ 0.10) increasing trend are located around the southern side of the island. The southern part of the island is the most populated area and has experienced significant urbanization during the study period. This indicates a strong urban bias associated with Tmin trend on the island. In the past, Jones et al. (1989) attributed nearly one fifth of the global land-based temperature increase to the urbanization bias in the northern hemispheres. However, in developing region (e.g. South America), Camilloni and Barros (1995) attributed nearly 60% of the temperature trend to urbanization. Much of the decline in annual Tmin at station 12 is occurring during winter season. Despite the recent warming of SST in the near-shore waters around the island of Oahu during 1956–2008, the decreasing trend (−0.51 °C/decade, p ≤ 0.05) at station 12 is distinct. As Tmin at station 12 showed a non-significant (p > 0.10) increase over the long-term (1969–2007) particularly during the summer (Table 3), recent decline in Tmin could have been possibly due to a change in the microclimate as the site is largely surrounded by ocean.

Figure 5.

Spatial variability of the magnitude of annual minimum temperature trends and their levels of significance during 1983–2007. Station 8 was excluded because of short record length. Upward (downward) hollow triangles indicate positive (negative) direction of trends, and their size corresponds to the magnitude of trends. Black triangles indicate trends significant at p ≤ 0.05.

Table 4. Seasonal (summer and winter) and annual trends and Sen's slope estimatesa for minimum temperature during 1983–2007b from OTI and 12 individual stations across the Island of Oahu, Hawaii
  1. Dark grey represents the significant trend at p ≤ 0.05 and light grey at p ≤ 0.10 significance level.

  2. a

    Sen's slope estimates after pre-whitening are presented between brackets.

  3. b

    Except station 8 which ended in 2000.


The recent trends in OTI for Tmin showed highly significant (p ≤ 0.05) annual (0.24 °C/decade) and significant (p ≤ 0.10) winter (0.27 °C/decade) warming (Table 4) at rates that exceed the past long-term (1969–2007) OTI for Tmin trends (Table 3). However, this enhanced warming in OTI for Tmin during 1983–2007 was not evident during the summer. Among the five long-term stations (3, 6, 9, 10, and 12), only urban station 3 showed enhanced warming during 1983–2007. At station 3, Tmin increased 1.5 times more rapidly during 1983–2007 than 1969–2007 (Table 4). In contrast, warming in Tmin at station 9 stayed nearly the same between the two periods. This finding indicates that enhanced warming due to urbanization at station 3 has been more recent compared to station 9. Similar to long-term trends in Tmin, no significant (p > 0.10) decline was observed at station 6 during 1983–2007. Annual and summer trends in Tmin at station 10 during 1983–2007 were also not significant (p > 0.10) despite significant increases during 1969–2007. Thus, not all stations show enhanced warming during 1983–2007. Enhanced warming in OTI for Tmin most likely reflects enhanced warming at station 3.

Recent seasonal trend analysis shows that the Tmin has increased more rapidly during the summer than during the winter except for stations 1 and 5, which is consistent with long-term trends. However, the percentage of stations with significant trend (p ≤ 0.10) is higher during winter (67%) compared to that during the summer (42%). Similar to GDL, the magnitude of statistically significant (p ≤ 0.10) increasing trends increased with elevation, indicating an elevation dependence in temperature trend. Although the stations on Oahu are well below the low-elevation threshold (<800 m) used by GDL, our results show enhanced warming with elevation among low-elevation stations. Hence, these results are consistent with enhanced warming at higher elevations in Hawaii reported by GDL and Diaz et al. (2011). The correlation between Tmin trend and elevation was stronger during winter than summer (summer, R2 = 0.31, p < 0.01; winter, R2 = and 0.58, p < 0.01). As all the stations used in this study are well below the TWI (∼2000 m), the elevation-dependence of the Tmin trend could be due to local microclimatic variations.

The MK-2 test identifies two change point periods (mid 80s and 90s) when significant upward shifts in Tmin began. The change points for three (5, 7, and 9) of the eight stations located significant (p ≤ 0.10) trends were located around 1985. Four stations (2, 3, 9, and 12) were selected to illustrate the variability in the change point across the island during 1983–2007 (Figure 6). The shift in the cooling of Tmin at station 12 began around the same time (year 1993) as the shift in the warming of Tmin at station 3.

Figure 6.

Trends in the annual average minimum temperature at four different locations across Oahu. Slopes marked with asterisk sign indicate significant trend at p ≤ 0.05. Corresponding Gaussian normal variate from sequential version of Mann-Kendall test (MK-2) is shown on secondary y-axis.

3.2.2. Tmax

Percentage of stations with significant trend (p ≤0.10) in Tmax was smaller (25%) compared to Tmin (58%). Opposite to Tmin, the recent trends show an overall cooling in annual and winter Tmax and warming in summer Tmax (Table 5). The annual and seasonal trends in OTI for Tmax show an overall cooling despite the \nobreak significant (p ≤ 0.10) summer season warming at four stations. Highly significant (p ≤ 0.05) warming in Tmax was evident at station 10, a suburban windward location. A similar significant (p ≤ 0.10) warming of summer Tmax was also evident at station 4, a rural windward station. The two urban sites (3 and 9) exhibited cooling of annual Tmax but only trend at station 3 was statistically significant (p ≤ 0.05). Although annual and winter Tmax at station 3 show recent cooling, summer temperature increased by the similar magnitude. Although the majority of the stations show decreasing trends in annual Tmax during 1983–2007, there are only two stations (3 and 12) with significant downward trends (Figure 7). Similar to Tmin, the highest decreasing trend (−0.44 °C/decade) was observed at station 12 (Table 5). Station 12 is the only station that showed highly significant (p ≤ 0.05) decreasing trends for both Tmax and Tmin. The five stations (2, 3, 5, 7, and 9) located around the southern side of the island all show downward trends in Tmax (Figure 7). Only station 3 showed highly statistically significant (p ≤ 0.05) decline in Tmax.

Table 5. Seasonal (summer and winter) and annual trends and Sen's slope estimatesa for maximum temperature during 1983–2007b from OTI and 12 stations across the Island of Oahu, Hawaii
  1. Dark grey represents the significant trend at p ≤ 0.05 and light grey at p ≤ 0.10 significance level.

  2. a

    Sen's slope estimates after pre-whitening are presented between brackets.

  3. b

    Except station 8 which ended in 2000.

Figure 7.

Spatial variability of the magnitude of annual maximum temperature trends and their levels of significance during 1983–2007. Station 8 was excluded because of short record length. Upward (downward) hollow triangles indicate positive (negative) direction of trends, and their size corresponds to the magnitude of trends. Grey (black) triangles indicate trends significant at p ≤ 0.10 (p ≤ 0.05).

The opposite trends in recent Tmax and Tmin might be resulting from the possible impact of urbanization, which has countered daytime warming and enhanced nighttime warming as reported by GDL. However, the rate of decrease in Tmax is small compared to Tmin at individual sites, probably because urbanization, in general, tends to impact Tmin more than Tmax (Karl et al., 1988; Saaroni et al., 2000). Saaroni et al. (2000) attributed this difference to ‘negative urban heat island effect’ or ‘urban heat sink’. Canyon geometry, which characterizes cities, is a major contributor to ‘urban heat sink’ as insolation absorption decreases at the street level with the relatively high heat capacity of the buildings (Saaroni et al., 2000).

The differential trend in OTI for Tmin (0.24 °C/decade, p ≤ 0.05) and Tmax (−0.20 °C/decade, p ≤ 0.05) resulted in a highly significant (p ≤ 0.05) decline in recent DTR by −0.44 °C/decade. Also, the decline in the winter season (−0.51 °C/decade, p ≤ 0.05) DTR is large compared to the summer season (−0.38 °C/decade, p ≤ 0.05). This recent decline in annual DTR is nearly twofold faster than the trend during 1969–2007 (−0.18 °C/decade) (Figure 3). No significant trend was observed in Tavg (0.05 °C/decade, p > 0.10) despite the overall warming and cooling of Tmin and Tmax, respectively. Majority of the stations show change points in Tmax during the positive PDO phase of 1977–1998. The shift in Tmax at station 12 is consistent with that in Tmin indicating that the cooling in Tmax and Tmin began in same year (Figure 8). In contrast, the shift in Tmin began four years before to the shift in Tmax occurred at station 3. On average, the shift in Tmax began 1.5 years earlier than the shift in Tmin. An earlier change point in Tmin occurred only at station 3.

Figure 8.

Trends in the annual average maximum temperature at four different locations across Oahu. Slopes marked with asterisk sign indicate significant trend at p ≤ 0.05. Corresponding Gaussian normal variate from sequential version of Mann-Kendall test (MK-2) is shown on the secondary y-axis.

3.3. Trends in extreme indices (1969–2007)

Indices of cold extremes (TXn, TNn, TN10p, TX10p, and CSDI) predominantly show significantly decreasing trends; however, some stations do show a statistically significant increasing trends (Table 6). Three of the six stations showed decreasing trend in the temperature of the coldest days (TXn) but none were statistically significant (p > 0.10). Similarly, the number of cool days (TX10p) showed non-significant (p > 0.10) decreasing trends at 50% of the stations. These results are not surprising as Tmax shows no significant trends during 1969–2007 (Figure 4) and mostly decreasing trends during 1983–2007 (Table 5). The temperatures of the coldest nights (TNn) show significant decreasing trendsat two stations: −1.04 °C/decade (p ≤ 0.05) at station 6 and −1.4 °C/decade (p ≤ 0.10) at station 12. The two urban sites (stations 3 and 9) show an increase in TNn with a highly significant (p ≤ 0.05) trend of 0.5 °C/decade at station 3. Trends in the number of cool nights (TN10p) at four stations showed both significant and non-significant decreasing trends. Significant (p ≤ 0.05) declines in TN10p occurred at urban stations 3 and 9 at rates of −2.7 and −3.5%/decade, respectively. These decreasing trends in TN10p indicate that cool nights have diminished by 10 and 13 d/decade. CSDI has also decreased at stations 3 and 9 by a rate of ∼1.0 d/decade. Station 8 showed a statistically significant (p ≤ 0.05) increasing trend during 1969–2000. Although there is no consistency in spatial pattern in all indices of cold extremes, the trends at the two urban stations, 3 and 9, clearly indicate urban influence.

Table 6. Annual trends in extreme indices at six locations around the Island of Oahu during 1969–2007a
  1. a

    Except station 8 which ended in 2000.

  2. Dark grey represents the significant trend at p ≤ 0.05 and light grey at p ≤ 0.10 significance level.


Trends in the indices of hot extremes (SU25, TR20, TXx, TNx, TX90p, TN90p, and WSDI) show significant increasing trends, indicating the warming of both daytime and nighttime hot extremes (Table 6). The only decreasing trend was found at station 12 in all indices of hot extremes except some insignificant increase in tropical nights (TR20) and number of warm nights (TN90p). Decreasing trends in hot extremes at station 12 is not surprising given the non-significant trend in Tmin and Tmax during 1969–2007 (Figure 4) and significant decline in both Tmin and Tmax during 1983–2007 (Figures 6 and 8). Annual numbers of hot days (SU25) has increased over the analysis period by as much as 11 d/decade, whereas the annual number of warm nights (TR20) has increased between 10 to 24 d/decade. The two highest increasing trends in TR20 were 13 and 24 d/decade at the urban stations 3 and 9, respectively. Stations 3, 9, and 10 show marginal increasing trend in the temperature of the warmest days (TXx) but none of them were statistically significant (p > 0.10). An increasing trend in the temperature of the warmest nights (TNx) was reported at all stations except station 6. The maximum increasing trend in TNx was 0.5 °C/decade (p ≤ 0.05) at station 9. Five of six stations show an increasing trend in number of warm nights (TN90p) with greatest increase at urban stations 3 and 9. The rates of increase in warm nights at these two locations (14–16 d/decade) are slightly larger than the rate of decline in cool nights (10–13 d/decade). In both cases, increasing trend in TN90p and decreasing trend in TN10p were slightly larger (2–3 d/decade) at station 9 compared to station 3. Trends in the number of warm days (TX90p) and WSDI at stations 3 and 9 were not significant. Only stations 8 and 10 showed significant increasing trend in TX90p (p ≤ 0.10) and WSDI (p ≤ 0.05), with both having least urban influence. This is in contrast to a non-significant (p > 0.05) trend in annual average Tmax during the same period (Figure 4), indicating a warming of extremes but not the mean temperature.

Comparisons of the average indices of warming nights (TR20 and TN90p) during the first (1969–1987) and second half (1988–2007) of the analysis period (1969–2007) at stations 3 and 9 show a contrasting pattern. The average number of the tropical nights (TR20) at station 3 has increased from 250 d year−1 during 1968–1987 to 275 d year−1 during 1988–2007. At station 9, the average TR20 increased from 192 d year−1 during 1969–1987 to 256 d year−1 during 1988–2007. During 1969–2007, TR20 at station 9 increased at nearly twice the rate of station 3 (Table 6) but still on average lower by ∼38 d. At station 3, the average number of warm nights (TN90p) has increased from 20 d year−1 during 1969–1987 to 50 d year−1 during 1988–2007. Similarly, the average TN90p at station 9 increased from 10 d year−1 during 1969–1987 to 55 d year−1 during 1988–2007. While the average TN90p at stations 3 and 9 during 1969–1987 were different, a more rapid increase at station 9 resulted in an average TN90p that was very similar to station 3.

Five stations show decreasing trends in DTR during 1969–2007 with statistically significant trend at three locations (Table 6). Station 8 show a statistically significant (p ≤ 0.05) increasing trend in DTR during 1969–2000. The two urban stations 3 and 9 show highly significant (p ≤ 0.05) decreases in DTR of −0.28 and −0.45 °C/decade, respectively. Decline in DTR can be attributed to a rapid rise in Tmin combined with no significant change in Tmax during 1969–2007 (Table 3). Irrespective of the station, these decreasing trends in DTR are large when compared to the state-wide DTR trend reported in the study of GDL during 1919–2006 (nearly threefold at stations 3 and 12 and fivefold at station 9). There could be two possible justifications for this difference. First, there has been a higher increase in Tmin on Oahu compared to other islands as shown in Table 3 (trend in HTI from Oahu gages is large compared to trend in HTI from non-Oahu gages). Second, increase in Tmin has been more rapid compared to Tmax during the recent times as shown by GDL. Trends of DTR on Oahu are higher than the global DTR trends of −0.066 °C/decade (1950–2004), −0.084 °C/decade (1950–1993) , and −0.032 °C/decade (1976–2004) (Vose et al., 2005; Easterling et al., 1997). Although Easterling et al. (1997) showed no significant urban influence on DTR trend, greater decline in DTR on Oahu due to rapid increase in Tmin at urban sites suggests urbanization does have a localized influence.

3.4. Effect of pre-whitening on trend

Serial correlation coefficient at lag-1, r1, calculated from temperature time series f (t) ranged between −0.05 and 0.70 for Tmin and between −0.25 and 0.75 for the Tmax. Tmax shows greater serial correlation during summer than during winter for the majority (58%) of stations. The effect of AC on temperatre trend analysis is evident in the magnitude of trends as well as their significance level. Applying TFPW in the trend analysis changed the significance level at a few locations (Tables 3-5). However, Sen's slope estimates were only marginally affected by the TFPW, mainly in time series with weaker serial correlation. Theoretically, the sequence of values in a pre-whitened time series are considered independent. However, the residual time series yt′ displays the presence of significant lag-1 AC at several stations. This indicates that the effect of AC on trend can be minimized after pre-whitening, assuming the temperature time series is a first-order autoregressive process, but cannot be eliminated. These results concur with findings of other studies which reported that pre-whitening in the presence of AC removes some of the trend from the data (Yue et al., 2003).

3.5. Comparison of air temperature anomalies and climate indices

To compare trends and shifts in temperature with large scale climate indices, actual and cumulative departures of monthly Tmax, Tmin, and Tavg anomalies (OTA) were plotted along with cumulative indices of the PDO, HRI, and OTA (Figure 9). The Tmax anomaly followed the PDO closely until 1973 after which Tmax anomaly showed no further decline up until 1979 (Figure 9(a)). Starting in 1979, Tmax anomaly started increasing after the shift in the phase of the PDO (from negative to positive phase) began in 1977. After the shift in the PDO phase, HRI began to decline while increase in Tmax anomaly continued. Both Tmax anomaly and HRI were in phase and followed closely, except around 1983–1988 during which a rapid increase in PDO occurred. This indicates that Tmax is more closely tied to PDO compared to HRI. Tmin anomaly did not respond to the change in the phase of PDO in 1977 until 1981 and continued to decline (Figure 9(b)). HRI started declining during the rapid increase in PDO during 1983–1988 but Tmin anomaly was relatively stable compared to corresponding Tmax anomaly. However, beginning of rapid increase in PDO in 1983 coincides with the shift in Tmin from change point analysis (Figure 3). Overall Tmin showed better agreement with HRI compared to the PDO until 1999. HRI and Tmin anomaly show strong disagreement between 1999 and 2007. In addition, Tmin anomaly showed rapid increase beginning in 1999 despite the rapid decrease in PDO. This out of phase relationship between Tmin and PDO was not observed during the rapid decline in PDO during 1970–1976. We conclude that the local climate controls (e.g. heat island effect) might have greater influence on recent warming of Tmin. These findings are consistent with the findings reported by GDL, who found similar out of phase relationship between HTI and PDO during the recent decades. These results suggest that although there is a difference in the magnitude of trends between Oahu and other islands, the effect of large climatic controls (i.e. PDO and HRI) on island temperature is quite similar. Similar to Tmin, Tavg anomaly varied coherently with HRI until 1999 (Figure 9(c)). The shift in Tavg from change point analysis began a year prior to the switching of PDO phase (from negative to positive) in 1977. Effect of increasing PDO between 1983 and 1998 was not evident in Tavg which continued to decline.

Figure 9.

Monthly (bar) OTA and cumulative departures in the HRI, PDO and OTA for maximum temperature (a), minimum temperature (b) and average temperature (c).

The differences in the periodicities of Tmin and Tmax anomalies are highlighted using wavelet power spectrum and global power spectra (Figure 10). Significant variability in the power of Tmin and Tmax lies between 2 and 4 year time scales. However, the periodicity in the power of Tmin is more prominent compared to that of Tmax which only shows significance variability between 1975 and 1987. Similar to Tmin and Tmax, wavelet power spectrum and global power spectra show that most of the variability in the HRI lies in 2–5-year period (Figure 10) except during 1980 and 1990. Between 1980 and 1990 the HRI fluctuates on a time scale of 2–8 years, which is similar to the variability in Niño-3.4.

Figure 10.

(a) Wavelet power spectrum and (b) global wavelet spectrum for maximum (top), minimum (middle) OTA and HRI (bottom). Contours in black show significant power, while blue dashed lines show the significant region at p ≤ 0.05.

The cross-correlogram of Tmin anomaly and the HRI shows significant positive correlation at zero-lag (R = 0.32, p < 0.01) and at positive and negative lags of 1 and 2 months (Figure 11(a)). The highest correlation between Tmin and the PDO (R = 0.16, p < 0.01) and Niño-3.4 (R = 0.17, p < 0.01) was at a positive lag of 8–36 months. There was no significant correlation between the Niño-3.4 and Tmin at zero-lag (R = −0.06, p >0.10). Tmax anomaly and the PDO showed highest positive correlation (R = 0.40, p < 0.01) at zero-lag (Figure 11(b)). Niño-3.4 and Tmax are positively correlated between the lags of 0 to −20 months with highest correlation (R = 0.18, p < 0.01) at negative lag of 18 months. Opposite to Tmin, the HRI shows an overall significant negative correlation with monthly Tmax at zero-lag (R = −0.21, p < 0.01). Similar findings were also reported in some other studies on the temperature–precipitation relationship (Isaac and Stuart, 1992; Zhao and Khalil, 1993; Trenberth and Shea, 2005). Trenberth and Shea (2005) attributed the negative correlation between Tmax and precipitation to increased sunshine and less evaporative cooling under dry conditions. On the other hand, presence of clouds under wet condition can increase the Tmin (positive correlation between Tmin and the HRI) by enhancing downward longwave radiation (Dai et al., 1999).

Figure 11.

Cross-correlograms of the minimum (a) and maximum (b) temperature anomalies (OTA), PDO, Niño-3.4 and HRI. Grey dashed line corresponds to two-standard errors.

4. Summary and conclusions

While all global climate models show consistency in the warming signal, it is difficult to forecast regional changes and their potential effect due to increasing downscaling uncertainty at smaller scale (Boer, 2001). Analysing the trend in observed climate can provide considerable insights into changes in temperature and its potential impact on Hawaii's unique ecosystem, i.e. biodiversity and native habitants. Long-term (1969–2007) and recent (1983–2007) trends and shifts in Tmin and Tmax were analyzed using the nonparametric MK test (MK-1) and its sequential version based on rank statistics (MK-2). The effect of AC on the temperature trend analysis was also evaluated using a TFPW approach. The link between OTA for Tmax, Tmin, and Tavg, HRI, and climate indices (i.e. PDO and Niño-3.4) were also evaluated.

Island-wide trends in OTI for Tmin during 1969–2007 showed an overall warming with well pronounced seasonality. During this period, there was a non-significant (−0.005 °C/decade, p > 0.10) decline in Tmax; whereas the Tmin increased by 0.17 °C/decade (p ≤ 0.05). The OTI and GDL-NO for Tmin shows more rapid warming on Oahu when compared to other locations in the Hawaiian Islands. During 1983–2007, a highly significant (p ≤ 0.05) increase in OTI for Tmin and a decline in OTI for Tmax were found. Although annual and winter OTI for Tmax and Tmax at individual stations showed decline during 1983–2007, Tmax during summer indicated indicated. The enhanced warming in Tmin during 1983–2007 compared to 1969–2007 was evident only at one location (station 3). Trends at individual stations are in the range of warming rates reported earlier (Nullet and Ekern, 1988; Karl et al., 1996; Giambelluca et al., 2008). We also found overall decline in DTR during 1969–2007, with much greater decline during 1983–2007. Island-wide temperature anomaly (i.e. OTA) is closely related to the regional HRI. The HRI shows a significant positive correlation with Tmin and significant negative correlation with Tmax. The PDO shows greater influence on Tmax compared to Tmin. Tmin continued to rise since 1999 despite the recent cooling associated with PDO, possibly due to greater influence of local climate controls (e.g. urban heat island effect) and continued decrease in the HRI.

Using a relatively dense network of air temperature gages, we were able to highlight the spatial heterogeneity in temperature trends on Oahu. Spatial heterogeneity in temperature trends and effect of urbanization (particularly on Tmin and related extreme indices) on the island are intriguing. Although, the potential effects of night time urban warming on energy demands are largely unknown, the rapid increase in number of warm (TN90p) and tropical (TR20) nights may have significant implications for Oahu's population centres (i.e. Honolulu metropolitan area). Describing the temperature trend on the island and evaluating the climate sensitivity of water resources without considering the spatial heterogeneity and enhanced urban bias can be erroneous. Long-term weather stations, which are mostly located in cities, may show disproportionately enhanced temperature trends that may bias any regional means. Similarly, temperature trends calculated from multi-station index based analysis can be over- or under-represented by spatially varying increasing and decreasing trends.

The Hawaiian Islands have limited fresh water resources and rely heavily on rainfall and subsequent recharge to ground water aquifers. Decreasing trend in annual streamflow and baseflow have already been observed (Oki, 2004) which are in agreement with downward trends in precipitation (Chu and Chen, 2005; Timm and Diaz, 2009). Projected increase of summer precipitation by 5% under the AR4 A1B emission scenarios (Timm and Diaz, 2009) could be potentially offset by the increase in evapotranspiration as a result of increase in temperature (Safeeq and Fares, 2012). Effect of enhanced warming in Tmin and subsequent reduction in DTR can manifest the issue of alien species invasion which have already threaten the island ecosystem. Although we were able to highlight the spatial heterogeneity in temperature trend, most of gages were located around the coastline and majority of them are below 500 m elevation. A better monitoring of temperature in island interior and at higher elevations is necessary for accurate assessment of hydro-ecological impact of climate change on the island.


The project was partially supported by two grants from the US Department of Agriculture: (1) the Cooperative State Research, Education and Extension Service (grant no. 2004-34135-15058) and (2) the McIntire-Stennis Formula (grant no. 2006-34135-17690). The authors wish to thank Pao-Shin Chu from the Department of Meteorology, University of Hawaii, for providing the monthly HRI data. The 1-dimensional wavelet transform software used was provided by C. Torrence and G. Compo and is available at URL: http://paos.colorado.edu/research/wavelets/. Finally, the authors wish to thank Tom Giambelluca and one anonymous reviewer for their constructive comments and suggestions.