Temporal solar radiation change at high elevations in Hawai‘i



Trends in downwelling global solar irradiance were evaluated at high-elevation sites on the island of Maui, Hawai‘i. Departures from monthly means were assessed for the 6 month Hawaiian wet and dry seasons over the period 1988 to 2012. Linear regression analysis was used to characterize trends in each season. For the dry season (May–October), statistically significant (p ≤ 0.05) positive trends of 9–18 W m−2 (3–6%) per decade were found at all four high-elevation stations tested. Wet season trends were not significant, except at the highest-elevation station, which had a significant negative trend. No consistent trends in aerosol concentrations have been observed at high elevations in Hawai‘i; therefore, the observed dry season brightening is most likely the result of decreasing cloud cover. Supporting this hypothesis, analysis of 15 years (1997–2012) of high temporal resolution Geostationary Operational Environmental Satellite (GOES) imagery over the Hawaiian Islands showed a statistically significant decrease in leeward cloud cover amounting to 5–11% per decade over the stations. In addition, analysis of Moderate Resolution Imaging Spectroradiometer data were in general agreement with the GOES trends, although statistically significant dry season trends were found at only one of the four stations.

1 Introduction

Spatial and temporal variations in solar radiation (Kd) measured at the Earth's surface affect climate, photosynthetic activity, hydrologic cycling, and surface energy budget. Observations over the past ~50 years have shown substantial decadal variations in Kd [Wild, 2012, and references therein]. A number of authors have investigated both the regional and global trends in Kd and have speculated on causes of observed variations [Alpert et al., 2005; Che et al., 2005; Wild et al., 2005; Dutton et al., 2006; Long et al., 2009; Riihimaki et al., 2009; Liley, 2009; Ohmura, 2009]. In general, Kd variations can be caused by either changes in the quantity of solar radiation reaching the top of the atmosphere or changes in the transparency of the atmosphere. The radiation reaching the top of the atmosphere is affected by Earth's orbital parameters, which change relatively slowly and mainly affect seasonal radiation distribution and radiative output from the Sun. The 11 year sunspot cycle has only a small effect on reported variations in solar radiation [Wild, 2009]. As a result, variations in Kd are mostly the result of changes in the transparency of the Earth's atmosphere.

Transmission of solar radiation through the Earth's atmosphere is affected by variations in aerosol amount and type, concentrations of trace gases including water vapor, cloud radiative properties, and cloud frequency. Trend analyses of ground based surface solar radiation measurements have shown a sustained decrease from the 1950s to late 1980s across much of the globe, a phenomenon known as “global dimming” [Wild, 2012]. Measurements from the late 1980s onward suggest that global dimming did not persist and instead solar radiation increased over many areas, i.e., “global brightening” was observed [Wild, 2012]. Alpert et al. [2005] suggested that reported brightening may be a result of reduced aerosol emissions associated with more effective clean air regulations, along with the decline in the Eastern European economy in the late 1980s. However, recent trends have not been uniformly positive across the globe. Some sites in India, for example, have shown sustained dimming effects throughout the period of record, while sites in China have shown a switch from brightening to dimming after 2000 [Wang et al., 2012]. The areas of the globe showing continued or renewed “dimming” after 1980 are also the areas with the highest aerosol amounts [Ohmura, 2009]. While emission histories fit well with reported dimming/brightening, changes in atmospheric aerosols may not explain trends found in remote locations that are not heavily impacted by anthropogenic pollution. Dutton et al. [2006] have shown that dimming or brightening trends were observed at four remote locations with relatively stable clear-sky optical properties over the period of record. One of these remote stations is Mauna Loa Observatory (MLO) on the Island of Hawai‘i (3397 meters above sea level). Because of Hawai‘i's mid-oceanic location and distance from continental aerosol source areas, variability in atmospheric transmissivity is relatively low [Longman et al., 2012]. Annual cycles are seen in aerosols, with a February–June maximum associated with the period of active transport of Asian desert dust and pollution [Holben et al., 2001; Eck et al., 2005]. However, aerosol values are low in Hawai‘i compared with other areas of the world [Holben et al., 2001], and variations in aerosols have only small effects on incoming solar radiation at high elevations [Longman et al., 2012]. No consistent trends in aerosol concentrations have been observed at high elevations in Hawai‘i [Dutton et al., 2006; Longman et al., 2012], which suggests that changes in cloud climatology are the likely driver for changes in Kd. Dutton et al. [2004] reported a decrease in cloud occurrence at MLO from the late 1980s to 2001. Subsequently, Dutton et al. [2006] linked this decrease in cloud occurrence to observed solar brightening during an overlapping period at MLO.

For this research, data obtained from four remote, high-elevation stations in Maui, Hawai‘i, with solar observations spanning the most recent 21 to 25 years and satellite cloud imagery obtained from both the Geostationary Operational Environmental Satellite (GOES) and the two Moderate Resolution Imaging Spectroradiometer (MODIS) instruments for these same locations were analyzed for the 6 month Hawaiian wet and dry seasons. The objectives of this study were to determine, in each season, if significant solar radiation trends are evident and to identify related trends in satellite-derived cloud climatology data.

2 Methods

2.1 Solar Radiation Instrumentation and Observation Sites

At each study site, solar radiation was monitored with an Eppley model 8-48 (Eppley Laboratory, Newport, RI, USA) thermopile pyranometer measuring global solar radiation, sampled at a 10 s interval and recorded hourly using a LI-COR (Lincoln, NE, USA) LI-1000 data logger (before mid-July 1999) and a Campbell Scientific, Inc. (Logan, UT, USA) CR10X data logger (after mid-July 1999). Beginning in January 2011, a Hukseflux (Manorville NY, USA) model NR01 four-component net radiation sensor and Campbell Scientific CR3000 data logger were installed at each HaleNet station. The Eppley model 8-48 has a cosine response of ±2% (0–70°) and ±5% (70–80°) [McArthur, 2005], a broad spectral response range (glass dome has uniform transmission between 285 and 2800 nm, Eppley Laboratory Model 8-48 Instruction Sheet), and a low thermal offset (±1 W m−2) [Reda et al., 2003]. Expanded responsivity uncertainty (standard uncertainty multiplied by a coverage factor), for a single model 8-48 has been reported as −2.2 to +3.3% [National Renewable Energy Laboratory, 2011]. The Eppley 8-48 has been shown to be more sensitive to wind because of its single dome [Gueymard and Myers, 2009] and more difficult to calibrate precisely due to a longer time constant [Michalsky et al., 2007]. The Eppley 8-48 is commonly used to measure diffuse radiation at Baseline Surface Radiation Network sites when the proper shading devices are used [McArthur, 2005] but has been shown to systematically overestimate radiation under clear-sky conditions [Gueymard and Myers, 2009]. Four-component net radiation sensors such as the Hukseflux model NR01 measure the upward and downward fluxes of shortwave and longwave radiation separately. The shortwave downward component is measured by a thermopile pyranometer (type SR01) with a cosine response of ±0.5%, a spectral response range of 305 to 2800 nm, and a maximum thermal offset error of ±5% [Hukseflux Thermal Sensors, 2011]. Although the Eppley 8-48 and Hukseflux SR01 are not regarded by the World Climate Research Programme as first-class pyranometers [McArthur, 2005], after careful screening and homogenization [Longman et al., 2013; see brief description below], data from these sensors are deemed to be sufficiently accurate and consistent for the purpose of assessing temporal trends.

The climate stations used in this analysis are part of the HaleNet climate network (http://climate.socialsciences.hawaii.edu/HaleNet) located on Haleakalā Volcano, Maui, Hawai‘i (Table 1). Three of the four stations used in this analysis are located on the leeward slope of Haleakalā and cover an 870 m elevation gradient (2120–2990 m), and the remaining station is located on the windward slope at 2460 m. These stations represent areas near and above the 2200 m mean trade wind inversion (TWI) level [Cao et al., 2007]. In general, the attenuation of solar radiation increases with decreasing elevation due to greater atmospheric absorption. In addition, the TWI has a significant impact on radiation budgets at high elevations because it limits the vertical development of clouds [Giambelluca and Nullet, 1991]. The effects of the TWI can been seen along the leeward transect as a sharp increase in mean annual Kd (237 at 2120 m to 276 W m−2 at 2990 m) resulting from decreasing cloud frequency with elevation above the TWI (Table 1).

Table 1. HaleNet Stations Characteristicsa
StationElevation (m)OrientationLatitude °NLongitude °WStart DateMean Kd (W m−2)Mean Cloud (1 − Kd/Kd c-s)
  1. a

    Kd is solar radiation; Kd_c-s is clear-sky solar radiation calculated with a clear-sky radiation model described by Longman et al. [2012].

HN-1512120Leeward20.763156.251Jun 1988237.40.24
HN-1522590Leeward20.741156.249Mar 1990267.90.15
HN-1532990Leeward20.714156.259Mar 1990276.10.11
HN-1612460Windward20.734156.143Jun 1992295.00.84

2.2 Homogenizing Solar Radiation Data

In previous work, Longman et al. [2013] developed a new method to correct for inconsistencies in global solar radiation data at HaleNet stations arising from changes in sensor response. The method uses modeled clear-sky radiation [Longman et al., 2012] as a reference against which measured solar radiation on cloud-free days is compared. Abrupt shifts and gradual drifts in sensor response were detected as changes in the ratio of measured to modeled clear-sky radiation on cloud-free days and corrected to produce homogenous time series. It was obvious in the record that observed shifts in the instrument responsivity were not caused by real changes in aerosols, because the shifts occurred abruptly at times when sensors were either recalibrated or replaced. Sensor calibration drift was identified at only one (HN-152) of the stations analyzed. In that case, the trend attributed to calibration drift was deemed not the result of changes in aerosols because (1) the trend was identified at only one of the three stations analyzed, and (2) previous research found no corresponding trend in aerosol activity at high elevations in Hawai‘i [Dutton et al., 2006; Longman et al., 2012]. To determine the extent to which the homogenization procedure may influence the results of the current study, we calculate and compare trends for both the original and homogenized time series.

2.3 Calculating Seasonal Anomalies

Monthly solar radiation anomaly values were calculated as departures from the respective period-of-record monthly means. Removing the annual cycle in this manner minimizes the potential biasing effects of missing data on the calculation of means for individual seasons, enabling better assessment of temporal trends. A baseline period of June 1993 to October 2011 was used for calculating the period-of-record monthly means in Kd. During a preliminary analysis of the solar anomalies a least squares linear regression model was used to assess solar trends for each month at the three leeward stations used in this analysis (R. J. Longman, Homogenization of long-term solar radiation time series using a clear-sky radiation model and assessment of solar radiation variability at upper elevations on Maui, Hawaii, University of Hawai‘i at Mānoa, Honolulu, USA, unpublished data, 2011). A long-term monthly time series was evaluated for each month within the calendar year, and the p value of the slope was calculated and used as an indicator of significance. Of the 36 time series evaluated (3 stations × 12 months), 14 had negative slopes and 22 had positive slopes. Based on the sign of the trends in each month, two distinct seasons emerged, which correspond to the wet and dry seasons in Hawai‘i [Giambelluca and Schroeder, 1998]. Negative slopes were more common (11/18) during November to April (wet) season, while in the May to October (dry) season, slopes were predominantly positive (15/18). The 14 negative trends identified in this analysis were not statistically significant. For the dry season, results at all three stations showed statistically significant (p ≤ 0.05) positive trends in Kd (1.5–2.1 W m−2 per year) during the month of September. Significant positive trends were also shown during the month of July at stations HN-152 and HN-153 with increases in Kd ranging from 1.6 to 2.3 W m−2 per year. Subsequently, wet and dry season time series used in this analysis were produced by averaging the 6 monthly anomalies within each season year.

2.4 Analyzing Temporal Trends

Temporal trends for wet and dry seasons at each station were estimated using generalized least squares “weighted” linear regression model, which was executed using the nlme package in the R statistical software platform, with weights set equal to the estimated error for each value. Two types of data errors were evaluated for each season-year value: (1) instrument measurement errors and (2) data aggregation errors resulting from missing data. The regression model accounts for the effects of temporal autocorrelation, and weights are set equal to the combined instrument and aggregation errors in each year. The combined error (ei) can be expressed as

display math(1)

where ea is the aggregation error and em is the measurement error.

For measurement error (em), for periods in which the Eppley 8-48 model pyranometer was used to measure solar radiation (prior to 2011), the maximum expanded responsivity (±3.3%) was included in the error term. For periods using the Hukeseflux pyranometer (after 2011) the maximum thermal offset error (±5%) was included in the error term. To determine aggregation errors, two complete 6 month periods of daily anomaly data were randomly selected from each of the three leeward stations used in this analysis (six 6 month periods in total). The average of each 6 month period (183 days) was derived (Aall). A randomly selected daily value was then removed from each 6 month period, and the average was derived from the remaining data (A1). The error associated with 1 missing day (e1) for each 6 month period was calculated as

display math(2)

This process was repeated 1000 times so that a majority of the possible combinations of missing days were included. All e1 errors for the six 6 month periods were compiled to form a single distribution of e1, based on which the error associated with 1 missing day per season was defined within 95% confidence as 2σe1, where σe1 is the standard deviation of the distribution e1. These steps are repeated for 2, 3, …, up to 68 days (the maximum number of missing days in a single season) of missing data from the time series (Figure 1), e.g., for error e2, two daily values are randomly selected and removed from the time series. On average 95% of the data was available for each seasonal value used in this analysis.

Figure 1.

Map of the four high-elevation climate stations, part of the HaleNet system, used in this analysis. Note: GOES pixels have a spatial resolution of about 4 km and those of MODIS about 1 km.

2.5 Satellite-Based Cloud Detection

2.5.1 Geostationary Operational Environmental Satellite Data

A satellite-based cloud analysis was developed using multispectral imagery available from the National Oceanic and Atmospheric Administration Geostationary Operational Environmental Satellites (GOES). Imagery was collected and archived for the Hawaiian Islands at 4 km, 15 min (time) resolution for the period 1997–2012. Cloud detection was based on the contrast in observed emitted and/or reflected radiation compared with a computed clear-sky background (CSB) value, defined as the satellite-observed radiation emitted and/or reflected from the surface when no clouds are present, for each pixel [Wojcik et al., 2005]. The CSB varies spatially and temporally and is influenced by the radiative properties of the surface, surface temperature, terrain height, soil moisture, and solar illumination angle. Because of these variations, the CSB was calculated for each pixel, for different times of day, and for each band, and was updated continuously through time. CSB was estimated for albedo (visible), reflectivity (shortwave infrared, SWIR), a fog product, and longwave infrared (LWIR) emission, which were the basis for four tests of cloud presence. For each pixel, data from high probability clear times over the previous 30 days at a given analysis time were used to estimate CSB. This approach provides sufficient clear-sky data and reduces the effect of diurnal and seasonal cycles of temperature and illumination, in particular, on the calculated CSB. The database from which clear times are determined includes the satellite imagery, ancillary surface, and ship observations collected by the National Weather Service (NWS), World Meteorological Organization (WMO), and at several in situ observations at the Summit of Maui.

The albedo (visible) and reflectivity (SWIR) CSB values are each calculated by identifying and averaging the darkest 10% of albedo and reflectivity values from the previous 30 days of images. For nighttime periods, the fog product CSB is computed by identifying and averaging the warmest 10% of LWIR values for each pixel over the previous 30 days. The LWIR CSB is determined as the average of the difference between the LWIR temperature from the satellite for a given pixel and clear-sky LWIR brightness temperature estimated for that pixel from a linear regression model. The regression model is developed with data from clear-sky pixels that are used as prototypes. These prototype pixels are selected by a series of tests that find pixels with a high probability of being clear, even without the benefit of any of the cloud tests. The predictors include satellite data, time, terrain, and regional observations such as cloud cover and air temperature from the NWS and WMO. The warmest 10% of the LWIR residuals are averaged to determine the LWIR residual CSB that is used in the LWIR cloud tests. Comparing satellite images with CSB values for the different spectral products, four tests of cloud presence (albedo, SWIR reflectivity, fog product, and LWIR) were done, with the results combined to produce a composite probability of cloud occurrence for each pixel. The pixel was set to cloud when the estimated cloud probability exceeded 50%. The cloud algorithm was run on the entire satellite archive, consisting of nearly 500,000 images over the study period. The cloud fraction for the four solar radiation observation sites between the hours of 9:00 A.M. and 5:00 P.M. was extracted from the GOES-derived climatology, and the mean cloud occurrence was computed for the wet and dry seasons in each year.

Evan et al. [2007] presented evidence that global cloudiness trends seen in the International Satellite Cloud Climatology Project data set derived from GOES imagery are as a spurious result of changes in the satellite viewing angles during the study period. This issue is not relevant to our analysis, because all of the cloud imagery was taken from GOES satellites positioned at 135°W, providing a constant and relatively high viewing angle for Haleakalā (156°W).

2.5.2 Moderate Resolution Imaging Spectroradiometer Data

The MODIS instruments aboard the National Aeronautics and Space Administration's Terra and Aqua satellite platforms provide observations with the high spatial resolution necessary to determine patterns of cloud cover over the Hawaiian Islands. The MODIS instrument has a spatial resolution of 250 m in bands 1–2, 500 m in bands 3–7, and 1000 m in bands 8–36 and acquires data continuously, providing global coverage every 1–2 days. The sensors have two overpasses each per day, one at night and one during the day. The Terra and Aqua MODIS cloud mask products, MOD35 and MYD35 respectively, provide the likelihood that clouds obscure a given pixel at a 1 km resolution [Frey et al., 2008]. MOD35 and MYD35 were obtained for the entire MODIS time series (2001 to 2012) over the major Hawaiian Islands. Twelve years of data were obtained from Terra MODIS (2001–2012), and 10 years of cloud mask data were obtained from Aqua MODIS (2003–2012). Monthly statistics were generated from the daily cloud mask data, including mean cloud cover frequency at the two daytime overpass times. The mean daytime overpass times for the Hawaiian Islands, for Terra MODIS and Aqua MODIS, were 11:10 and 13:50, respectively. The cloud fraction for the four solar radiation sites was extracted from the Terra MODIS and Aqua MODIS-derived climatology, and the mean cloud occurrence was computed for the wet and dry seasons in each year. Linear trends are assessed for both the MODIS and GOES data using a generalized least squares regression model.

3 Results

Trends in solar anomalies for the wet and dry seasons at each station are presented based on the entire record (Figure 2). During the wet season, trends were negative at all sites (−0.7 to −0.1 W m−2 yr−1), but only at the leeward summit station (HN-153) was the trend statistically significant (p ≤ 0.05; Table 2). For the dry season, significant positive trends in solar radiation were found at all four stations (0.9 to 1.8 W m−2 yr−1; Table 2). Wet season GOES cloudiness had slight, nonsignificant negative slopes at the three leeward sites and a positive slope at the windward site (−0.3 to 0.1% yr−1; Table 3, Figure 3a). Analysis of GOES-derived dry season cloudiness showed downward trends at all four solar radiation stations, of which the three leeward sites were statistically significant (−1.3 to −0.5% yr−1; Table 3, Figure 3b). For the wet season MODIS data, regression slopes were positive at three of the four stations for Terra MODIS and negative at three of the four stations at Aqua MODIS platforms. For the dry season, consistent downward trends in cloud fraction were observed for both Terra MODIS (−0.3 to −0.7% yr−1) and Aqua MODIS platforms (−1.3 to −0.3% yr−1) at all stations. (Tables 4 and 5; Figures 3c–3f). However, only the dry season Aqua MODIS trend at the leeward summit stations (HN-153) was statistically significant.

Figure 2.

Time series and linear trends for (a) wet and (b) dry season solar radiation anomalies from 1988 to 2012 for four high-elevation climate stations located on Haleakalā, Maui, Hawai‘i.

Table 2. Temporal Trends in Seasonal Solar Irradiance Over the Period 1988–2012 for Four High-Elevation Stations, Maui, Hawai‘ia
Station/SeasonTrend (W m−2 yr−1)nr2SEp
  1. a

    W and D indicate wet and dry seasons, respectively; Trend is the slope of the regression line; n is the number of seasonal values in the time series; r2 is the coefficient of determination; SE is the standard error of the slope (W m −2); p is a measure of statistical significance.

Table 3. Temporal Trends in GOES Cloud Frequency Over the Period 1998–2012 for Four High-Elevation Stations, Maui, Hawai‘i
Station/SeasonTrend (% yr−1)nr2SEp
  1. SE is the standard error (%).

Figure 3.

Time series and linear trends of seasonal mean satellite-based cloud frequency at four high-elevation climate stations derived from 1998 to 2012 GOES imagery for the hours of 0900–1700 (LST) for (a) wet and (b) dry seasons, 2001 to 2012 Terra MODIS imagery at 1110 (LST) for (c) wet and (d) dry seasons, and 2003 to 2012 Aqua MODIS imagery at 1350 for (e) wet and (f) dry seasons.

Table 4. Temporal Trends in Terra MODIS (Mean Overpass Time 11:10 A.M.) Cloud Frequency Over the Period 2001–2012 for Four High-Elevation Stations, Maui, Hawai‘i
Station/SeasonTrend (% yr−1)nr2SEp
Table 5. Temporal Trends in Aqua MODIS (Mean Overpass Time 1:50 P.M.) Cloud Frequency Over the Period 2003–2012 for Four High-Elevation Stations, Maui, Hawai‘i
Station/SeasonTrend (% yr−1)nr2SEp

In general, GOES and MODIS time series were in reasonably good agreement for the overlapping periods analyzed, considering the different temporal and spatial resolutions of the two sources (Figure 4). Comparing MODIS and GOES data shows that they are positively correlated with slopes of 0.8 (r2 = 0.32) and 0.6 (r2 = 0.18) for the Terra and Aqua platforms, respectively (Table 6). In addition, we also analyzed the solar time series for the 15 year (1998–2012) period that corresponds with the GOES time series. The sign of the regression slopes is in agreement with those of the longer period (Table 2). However, for the 1998–2012 period as compared with the full period-of-record results, slopes were higher for the wet season and lower for the dry season. Dry season trends, which were statistically significant at all four stations during the full record, were significant at only two of the four stations for the 1998–2012 record (Table 7).

Figure 4.

The relationship between cloud frequency from high temporal resolution (15 min), low spatial resolution (4 km) GOES imagery and low temporal resolution (once daily), high spatial resolution (1 km) Terra and Aqua MODIS satellite imagery.

Table 6. Regression Parameters for Terra MODIS and Aqua MODIS Versus GOES Satellite Cloud Dataa
  1. a

    The slope of the regression line is b; the y intercept is a; MODIS and GOES are respective cloud frequency means of all season values at the 4 HN stations during periods of overlapping data.

Terra MODIS960.7514.80.3242%36%
Aqua MODIS800.6432.10.1855%35%
Table 7. Temporal Trends in Seasonal Solar Irradiance Over the Period 1998–2012 for Four High-Elevation Stations, Maui, Hawai‘i
Station/SeasonTrend (W m−2 yr−1)nr2SEp
  1. SE is the standard error of the slope (W m −2).


To verify that the apparent trends were not an artifact of the homogenization procedure, the original time series (before homogenization) were analyzed for comparison. Positive trends in both seasons were higher for the original data, indicating that the homogenization produced more conservative results with respect to the observed increases in solar radiation.

4 Discussion

Observed global dimming and brightening trends can be traced to variations in aerosols, clouds, and aerosol-cloud interactions, with regional differences in the dominant source of change [Wild, 2012]. Brightening trends of recent decades have been observed under both clear-sky and all-sky conditions suggesting that decreases in aerosol concentration have been the predominant driver of this change globally [Wild et al., 2005]. For the high elevations of Hawai‘i, however, above which aerosols loads are very low [Holben et al., 2001] and relatively invariant [Longman et al., 2012], cloud changes unrelated to aerosol variations are probably the cause of observed changes in solar radiation. Possible exceptions to that generalization may occur due to the local effects of active volcanism in Hawai‘i, e.g., the current phase of the Kīlauea eruption which began in 1983 and large volcanic events elsewhere, such as the eruption of Mount Pinatubo in June 1991. The effects of the Kīlauea eruption are confined almost entirely to the layer below the TWI and are negligible at high-elevation stations on Maui. During the 2 years following the Pinatubo eruption, the increased aerosol load did cause a detectable reduction in clear-sky radiation at high elevations on Maui [Longman et al., 2013]. However, Dutton et al. [2006] found that Pinatubo had very little effect on global all-sky radiation at Mauna Loa Observatory despite large short-term effects on aerosol optical depth, because of low absorption and high forward scattering of the volcanic aerosols.

All satellite-derived dry season cloud frequency trended downward for the period of available data, confirming that changes in cloud cover are responsible for the observed positive dry season radiation trends. The analysis of the MODIS data is in general agreement with the GOES analysis despite the much more limited temporal resolution (once daily for each MODIS sensor). One possible explanation for the higher cloud percentages for the MODIS data is that GOES images are observed from a stationary, near vertical vantage point, while the images are acquired from several different angles, thus capturing the influence of low-middle clouds not detected with the GOES imagery [Minnis, 1989; Evan et al., 2007]. In addition the higher spatial resolution of the MODIS data does better at capturing the cloud conditions along the steep inhomogeneous mountainside topography. This is especially apparent at the windward station (HN-161), which was the only station where the GOES cloud trend was not statistically significant. At a 4 km resolution, the GOES pixel undoubtedly often extends into the orographic cloud zone, the upper boundary of which is generally found about 200–300 m down slope (horizontal distance of approximately 0.9–1.6 km), perhaps damping the actual changes in cloud cover at the station. We also note that the horizontal distances between the leeward stations, 2400 m (HN-151–HN-152), 3300 m (HN-152–HN-153), and 5550 m (HN-151–HN-153) are sufficient to ensure that they fall in separate MODIS pixels (1 km resolution) for any given image. For GOES pixels (4 km), it is possible to have two or even three stations in a single pixel for some images.

The reduction in dry season cloud cover may be associated with changes in tropical atmospheric circulation resulting in fewer disruptions of Hadley cell subsidence over Hawai‘i. Descending air in the central subtropical Pacific maintains a temperature inversion (the TWI) over Hawai‘i [Cao et al., 2007]. The TWI is a stable atmospheric layer that inhibits uplift and halts cloud development at its base. Cao et al. [2007] found that the persistence of the TWI increased from 1979 to 2003, but it is uncertain if this trend persisted. Model projections of future changes in TWI characteristics over Hawai‘i indicate increases in the frequency and intensity of the TWI and a decrease in the mean TWI base height under the two warming scenarios tested (Representative Concentration pathway's 4.5 and 8.5) [Lauer et al., 2013]. Any changes in TWI occurrence frequency or height would obviously have effects on cloud cover for locations near or above the mean TWI level, including the stations used in this analysis.

While the cause of the apparent increase in TWI occurrence frequency is uncertain, shifts in Hadley cell circulation and/or changes in interactions between tropical and midlatitude circulations provide possible explanations. Cao et al. [2007] suggested that a global warming-related poleward shift in midlatitude storm tracks [Yin, 2005] resulting in few interruptions of Hadley cell subsidence over Hawai‘i might be responsible for greater TWI persistence. However, trends in solar radiation and clouds are seen only in the dry season, whereas the change in storm tracks is of consequence mainly in the winter (wet season). Northern Hemisphere Hadley cell subsidence has increased since 1950 but primarily during the winter (December-January-February) and spring (March-April-May) seasons [Quan et al., 2004]. Perhaps, the best explanation is that the zone of Hadley cell subsidence is expanding due to climate change [Seidel et al., 2008], making interruptions of subsidence less frequent in Hawai‘i. Global simulations of future climate, for example, show that the margins of Hadley cell circulation move poleward in both hemispheres and all seasons as a result of global warming [Kang and Lu, 2012]. It has been hypothesized that the upward motion in equatorial convective regions has intensified while the subsidence regions (such as Hawai‘i) have become drier and less cloudy [Chen et al., 2002], which is consistent with the widely held view that global warming will cause wet regions to become wetter and dry regions drier [John et al., 2009]. Climate fluctuations associated with ENSO (El Niño–Southern Oscillation) might explain some of the observed variations in cloudiness. The effects of ENSO on wet season precipitation in Hawai‘i are well known [e.g., Chu and Chen, 2005]. Analysis of Hawai‘i rainfall versus the Multivariate ENSO Index [Wolter and Timlin, 1998] shows that both wet and dry season rainfall respond to ENSO, though in opposite directions [Frazier et al., 2012]. Dry season rainfall anomalies tend to be positive during El Niño and negative during La Niña, and vice versa for the wet season. Assuming rainfall and cloud cover are positively correlated, one would expect to see solar anomalies with signs opposite of the rainfall anomalies, but this is not always the case for the period of record analyzed here. We are currently investigating the relationship between ENSO and high-elevation cloud cover; results are not yet available.

In this study we focused on the analysis of seasonal rather than annual values because, as previously mentioned, we believe that both natural climate variability and effects of global climate change will have distinctly different signatures in the two 6 month Hawaiian seasons. Also, rules for excluding years or seasons with a certain number of missing days from the analysis result in the annual time series being significantly more limited by missing values. However, we can obtain an estimate of the trends in annual means by analyzing the time series generated by averaging the anomaly values of the two seasons in each year. Doing so results in positive trends at all four stations (0.25–0.71 W m−2 yr−1), but only the trend at station HN-152 (2590 m) was statistically significant (p ≤ 0.05). The trend at HN-152 remained significant even with the large positive anomaly in 2012 excluded.

The ecological impacts of increased solar radiation at high elevations in Hawai‘i may already be occurring. In the immediate vicinity of our study stations, increases in Kd and the number of zero rain days during the dry season, along with higher temperatures have been linked to the decline of an iconic high-elevation endemic species, the Haleakalā silversword [Krushelnycky et al., 2013]. Continuing reductions in dry season clouds and rainfall will inevitably have other effects on the subalpine ecosystem atop Haleakalā and other high mountains in Hawai‘i.

Instrument error can significantly affect the assessment of solar radiation trends. Gueymard and Myers [2009] showed that 7 year global radiation trends for the month of December calculated with measurements obtained from five different pyranometer types at the same location varied widely. The Eppley 8-48 was not assessed in their analysis. However, considering the inherent instrument errors discussed earlier, it is reasonable to assume that this instrument would be subject to measurement errors that could affect the estimated trends. A second source of uncertainty is the way in which global radiation is estimated. The most robust way to calculate global Kd is the summation of independent measurements of direct and diffuse radiation measured separately with a shaded pyranometer and a pyrheliometer, respectively [Gueymard and Myers, 2009]. Wang et al. [2013] found that while estimates of monthly global Kd by the two methods were in agreement, the trend derived from global measurements by a single pyranometer was 2–4 W m−2 per decade lower than that calculated with the summation method. A third source of uncertainty is related to the fact that including or excluding a few years at either end of a time series can have dramatic effects on the resulting trend estimates especially for relatively short records [Liebmann et al., 2010]. We tested this type of sensitivity by removing 1, 2, 3, and 4 years from the start and end of each solar radiation time series and recalculating the trends for all possible combinations. We found that regardless of the start or end date used, dry season trends in Kd remained statistically significant in all cases tested. Despite the uncertainties discussed above, this result, the consistency of the solar trends among stations, and the agreement with satellite-based observations of cloud cover trends during the overlapping periods lend confidence to the results of this analysis.

As previously mentioned, interannual climate variability in Hawai‘i is influenced by ENSO [e.g., Chu and Chen, 2005]. Prior analysis has also shown that decadal-scale changes in air temperature [Giambelluca et al., 2008] and precipitation [Diaz and Giambelluca, 2012; Frazier et al., 2013] are strongly coupled to the Pacific Decadal Oscillation (PDO) [Mantua et al., 1997]. The solar radiation trends reported here are derived from time series limited to 24 years or less and are therefore subject to possible influences of PDO, i.e., the trends could be partly explained by natural variability. Because of this, the results here should be interpreted with caution, especially regarding possible links with long-term changes in climate related to global warming.

5 Conclusions

Dry season solar radiation increased during 1988 to 2012 at upper elevations on the island of Maui, Hawai‘i. Change in aerosol loading is not considered to be an important contributor to the observed radiation changes. Trends in satellite-derived cloud fraction are consistent with observed solar radiation changes; statistically significant, positive Kd trends of 9–18 W m−2 (3–6%) per decade are in agreement with decreasing cloud amounts amounting to 5–11% per decade along the gradient. The decrease in dry season cloud frequency might be related to greater persistence in the trade wind inversion, possibly as a result of expansion of Hadley cell circulation extent in response to global warming. Hawai‘i's fragile montane ecosystems are sensitive to increasing solar radiation, which is likely to exacerbate the negative impacts of increasing temperature [Giambelluca et al., 2008] and declining rainfall [Diaz and Giambelluca, 2012].


This work was partially supported by the Pacific Island Climate Change Cooperative (PICCC) and the Department of Interior Pacific Islands Climate Science Center (PICSC). Additional support of HaleNet field observations and data management provided through NSF EPSCoR 0903833. We thank the staff of Haleakalā National Park and the Pacific Island Ecosystem Research Center (PIERC), USGS, for their long support of the HaleNet system. Special thanks go to Lloyd Loope and Gordon Tribble of PIERC, Henry Diaz of the University of Colorado, and Camilo Mora and Abby Frazier of the University of Hawai‘i at Mānoa. The data used in this manuscript can be made available by contacting Ryan Longman at the University of Hawai‘i at Mānoa (rlongman@hawaii.edu).