The impacts of climate change are felt most strongly through the changes in the extremes (Karl et al., 1999). Many regional studies have analysed changes in climate extremes and have generally indicated that cold extremes are becoming less frequent and warm extremes are becoming more frequent (Karl et al., 1996; Plummer et al., 1999; Easterling et al., 2000; Hyun et al., 2002). In the Asian region, Yan et al., (2002) reported a gradual reduction of the number of cold days in China over the 20th century, and an increase in the number of warm days since 1961. Klein Tank et al., (2006) used 116 stations located in central and southern Asia to show warming in both the cold and warm tails of the distributions of daily minimum and maximum temperature between 1961 and 2000. Using 91 stations from southeast Asia and the south Pacific, Manton et al., (2001) found significant increases in the annual number of hot days and warm nights, and decreases in cold days and cold nights over 1961–1998. Griffiths et al., (2005) reported similar results for 89 stations across the Asia-Pacific region during 1961–2003. More recently, analyses of changes in climate extremes across South Asia have been described by Sheikh et al., (2010). Again, a general tendency toward more warm extremes and fewer cold extremes was found. However, few previous studies for Asia have examined changes in extremes as a function of observation station elevation or latitude.
The issue of elevation dependency of the surface climate-change signal is important for a number of reasons. First, an enhancement of high-elevation changes in surface climate would imply greater impacts on high-altitude ecosystems and hydrologic systems. Second, an amplified response at high elevations could be utilized as an early climate-change detection tool. Third, the capability of reproducing the elevation dependency of the climate-change signal could provide an important aspect of model verification (Giorgi, 1997). Recent studies suggest a more sensitive response to global warming at higher elevations in China compared to other parts of the country (Titan et al., 2006; Yang et al., 2006; Kang et al., 2007). Other investigators have also indicated greater rates of warming at higher elevations in other parts of the world (Beniston and Rebetez, 1996; Giorgi et al., 1997; Fyfe and Flato, 1999; Beniston, 2003; Pepin and Lundquist, 2008; You et al., 2008, 2010; Kang et al., 2010). There are a number of reasons why an elevation dependency of surface climate change can be expected. Beniston and Rebetez (1996) attribute their finding of greater warming at higher altitudes to the fact that high-elevation stations are more directly in contact with the free troposphere than low-elevation ones, and therefore, less affected by ameliorating anthropogenic factors such as urbanisation and pollution. The snow-albedo feedback can also provide a strong elevation-dependent forcing. As snow is depleted at high altitudes under warming conditions, the surface albedo decreases, so that more solar radiation is absorbed at the surface and the high-elevation surface warming is enhanced. However, enhanced warming with elevation is not found everywhere and some observational studies show contrasting patterns. Analysis of high-elevation stations on the Tibetan Plateau over 1961–2005 showed no significant correlations between elevation and trend magnitudes of temperature extremes, except for coldest day temperature (You et al., 2008). In South America, lower elevations to the west of the Andes have experienced the greatest warming, while warming at higher elevations to the east is less marked (Vuille et al., 2003). Analysis of 1084 stations from the Global Historical Climate Network (GHCN) and Climate Research Unit (CRU) datasets by Pepin and Seidel, (2005) did not yield systematic relationships between temperature trends and elevation.
In addition to altitude, it is generally accepted that climate sensitivity increases at higher latitudes as changes in high latitudes are strongly tied to changes in albedo and energy budgets (Chapin et al., 2005). However, this relationship has not been investigated in terms of climate extremes.
The distribution of weather observation stations with respect to latitude and altitude in South Asia is shown in Figure 1(a) and (b). The stations are located from near-sea level to nearly 4000 meters above sea level, and over 30° of latitude from 6°N in Sri Lanka to 36°N in Pakistan. Consequently, the climate observation network of South Asia provides an opportunity to examine variations in trends in extremes, both vertically and meridionally. For South Asia, the southernmost regions in Sri Lanka and peninsular India are surrounded by ocean, and the most northern sections of Pakistan are characterized by the glacier climate of the Himalayas. Consequently, latitudinal differences in South Asia could be confounded by climatological differences. The purpose of this paper is to examine trends in a number of indices of temperature extremes over South Asian stations to determine whether there are indeed significant variations with altitude or latitude.
2. Data and methodology
This study is based on time series of daily temperature extremes from observation stations in Bangladesh, India, Nepal, Pakistan and Sri Lanka resulting from a collaborative research project on changes in climate extremes in South Asia (Sheikh et al. 2010). Daily minimum and maximum temperature data from 197 stations shown in Figure 1 are used for the analysis. Altitudes for the ensemble of the stations are ranging from 1 m to 2343 m. Data of poor quality, undesirably low quantity or low spatial coverage will lead to unsatisfactory data analysis and vague results. Therefore, though many stations in South Asia have longer records, we have selected these 197 stations covering the period of 1971–2000 in order to maximize the number of stations that passed tests for quality and homogeneity. (Aguilar et al., 2005) Data pre-processing has been considered to be an important part of the present study. In order to ensure that extremes are not discarded with an overenthusiastic or a purely automated data cleansing, a well-considered pre-processing is crucial for a meaningful analysis. Exploratory Data Analysis (EDA) is used to explore the data for each station individually. This allows identification of some extreme outliers and give an idea of potential inhomogeneities and missing values in the datasets. Time series plots for maximum and minimum temperatures have been subjected to visual examination. Some quality checks are also applied to station datasets, as there are some potential sources of error with the daily data: (1) Decimal point is out of place; (2) Maximum temperature is less than minimum temperature of the day; (3) Same data value is repeated for several consecutive days; (4) Same extreme is repeated; and (5) Identification of outliers using objective criteria based on the standard deviation of the time series (viz. mean − n*s.d. < x < mean + n*s.d. where x is the data value to be inspected, mean is average for that day of the year, s.d. is standard deviation of the mean and n is an integer set by the user; in present analysis, n is set as 3). Efforts have also been taken to cross-check the suspected/doubtful data and to confirm erroneous values, with the help of parallel datasets.
Eleven extreme temperature indices recommended by the Expert Team on Climate Change Detection Monitoring and Indices (ETCCDMI) are used. These indices are listed in Table I and described more fully at http://cccma.seos.uvic.ca/ETCCDMI/list_27_indices.shtml. Data quality checking and extremes analysis were performed using the RClimDex software maintained by the Climate Research Branch of Environment Canada (http://cccma.seos.uvic.ca/ETCCDMI/software.shtml). Time series for each index were computed by researchers from each participating country for their respective stations.
Annual percentage of nights with minimum temperature below 10th percentile using the 1971–2000 baseline period
TX10 (% days)
Annual percentage of days with maximum temperature below 10th percentile using the 1971–2000 baseline period
TN90 (% days)
Annual percentage of nights with maximum temperature above 90th percentile using the 1971–2000 baseline period
TX90 (% days)
Annual percentage of days with maximum temperature above 90th percentile using the 1971–2000 baseline period
TNn ( °C)
Lowest minimum temperature of the year
TNx ( °C)
Highest minimum temperature of the year
TXn ( °C)
Lowest maximum temperature of the year
TXx ( °C)
Highest maximum temperature of the year
Cold spell duration
Annual count of nights during spells with at least 6 consecutive nights with minimum temperature below 10th percentile using the 1971–2000 baseline period
Annual count of days during spells with at least 6 consecutive days with maximum temperature above 90th percentile using the 1971–2000 baseline period
DTR ( °C)
Diurnal temperature range
Mean annual diurnal temperature range (maximum temperature minus minimum temperature)
Time-series of each index is tested with EDA. For each station, trends per year have been computed for each index of temperature extreme for the period 1971–2000 using ordinary least squares regression and deriving the slope of the linear fit. The statistical significance of the trends has been tested at the 5% level (p < 0.05) using a t-test. These station trends are examined for dependence on altitude and latitude. For this purpose, trend magnitudes were sorted in ascending order of station altitude and latitude. Index station trends (1971–2000) are plotted from low to high station altitude and from south to north station latitude. Index trend values at stations were regressed against latitude and altitude to determine the meridional and vertical variations in index trends.
3.1. Changes in extreme temperature indices versus altitude
3.1.1. Frequency of temperature extremes
Figure 2 presents relationships between the magnitudes of the trends in indices of frequency of cold/warm days/nights (TN10p, TX10p, TN90p, TX90p) and elevation. The results are shown as a function of log10(altitude) due to the large range in station altitudes in South Asia.
The TN10p index is the percentage of daily minimum temperature values below the 10th percentile each year. It is a measure of the frequency of relatively cold nights and intuitively we would expect this index to have decreased as the climate warmed over 1971–2000. Figure 2(a) shows that generally stations below the height of 100 m do indeed show trends that are negative or close to zero. However, above the height of 100 m, many stations show large increases in the number of cold nights. This results in a statistically significant positive slope of the regression line in Figire 2(a). You et al. (2008) found that all Tibetan Plateau stations with elevations exceeding 2 km have negative trends in TN10p. In contrast, both positive and negative trends are seen at these altitudes for South Asian stations west of the Himalayas. This apparent inconsistency suggests something else is affecting the trends in high-altitude stations in South Asia. A mechanism that could contribute to changes in the number of cold nights is the change in nighttime cloud cover (Karl et al., 1993; Kukla and Karl, 1993; Karl et al., 1996; Dai et al., 1997; Dai and Trenberth, 1999; Plantico and Karl 1990). Decreased cloud cover can lead to cooler surface air temperatures overnight, and it is possible that stations with positive slopes in TN10p are locations where cloud cover has decreased.
TX10p is a measure of the percentage of cold days each year. TX10p station trends are plotted against elevation in Figure 2(b). In a warming climate, we would expect fewer cold days. However, many stations are shown with positive slopes in Figure 2(b). There is a small positive slope of the regression line but it is not statistically significant. In Figure 2(b), the majority of stations above 1 km have negative slopes in TX10p, a result that is in general agreement with the trends found over the Tibetan Plateau.
TN90p is a measure of the percentage of relatively warm nights each year. In a warming climate we would expect the trend in this quantity to be positive. However, we see in Figure 2(c) that both positive and negative trends have occurred above the altitude of 20 m. Consequently, the slope of the regression line in Figure 2(c) is negative and statistically significant at the 5% level. Again, this result is different from that obtained by You et al. (2008) in which no station above 2 km on the Tibetan Plateau showed a negative trend for TN90p.
We would expect the percentage of warm days per year (TX90p) to increase in a warming climate. However, many stations have registered a decreasing trend for TX90p, as shown in Figure 2(d). The slope of the regression of the trend with altitude is negligible, and the regression line is positive through the range of altitudes. In general agreement with the results from the Tibetan plateau, the majority of stations above 1 km in Figure 2(d) have positive trends.
Kothawale et al., (2010a) have shown that temperatures (mean, maximum and minimum) increased by about 0.2 °C per decade for the period of 1971–2007, with a much steeper increase in minimum temperature than maximum temperature. Further Kothawale et al., (2010b) have compared the behavior of regional trends in frequency of temperature extremes with mean trends, and seen that the increase in hot days and nights and decrease in cold days and nights are consistent with the increasing trend in monthly/seasonal maximum and minimum temperatures, respectively.
3.1.2. Intensity of temperature extremes
Figure 3 presents relationships between elevation and the station trends in indices TNn (coldest night temperature), TNx (warmest night temperature), TXn (coldest day temperature) and TXx (hottest day temperature). With perfect data and no local feedbacks we would expect these indices to increase in a warming climate. This behaviour is largely confirmed for stations above 2 km on the Tibetan Plateau, where the majority of stations show positive trends in these indices (You et al., 2008). However, these indices show both decreasing and increasing trends at high-elevation South Asian stations.
For trends in TNn (Figure 3(a)) the weak downward slope with elevation is not statistically significant. The variability of the index trends increases with elevation, i.e. at high elevations, stations exhibit large positive and negative time trends in the temperature of the coldest night of the year. For TNx (Figure 3(b)) there is again an increase in the variability of the trend magnitudes as elevation increases, but the dispersion of the points is less than for TNn, and the slope of the regression is significant at the 5% level. In Figure 3(b), we see that the trends are small, below 100 m altitude, and the variability at the highest stations contributes to the observed negative slope.
The positive slope in Figure 3(d) suggests that the warming trend in the hottest day of the year (TXx) is more pronounced at higher elevations. However, the slope of the regression is not significant if calculated using the robust statistics explained in Section 3.3. In Figure 3, we notice greater variability of trend magnitudes for the cold indices (TNn and TXn) in comparison with the hot indices (TNx and TXx); this difference results in the regression being statistically significant for the hot indices and not significant for the cold indices.
3.1.3. Cold and warm spell duration
The cold spell duration index (CSDI) is a measure of events with 6 or more consecutive nights recording minimum temperature less than the 10th percentile. Without local feedbacks, we would expect the cold spell duration to decrease in a warming climate. In Figure 4(a), trend magnitudes for CSDI are generally negative or close to zero for station altitudes up to 100 m, but many stations at higher elevations have large positive trends values. Reasons for increases in cold spell duration at high elevations are unknown but, as for TN10p, could relate to changes in overnight cloud cover. Because of the large positive trends at high elevation, the regression line shows a statistically significant increase with altitude.
Warm spell durations (WSDI), the index of warm spell duration, measures events with at least 6 consecutive days with maximum temperature greater than the 90th percentile. Warm spells are expected to increase in a warming climate. We see in Figure 4(b) that many stations have large decreasing trends for this index. The regression line indicates a weak increase in trend magnitudes with altitude, but the slope is not statistically significant.
3.1.4. Diurnal temperature range
The mean diurnal temperature range (DTR), the difference between mean daily maximum and minimum temperature, shows a decreasing trend over most areas of the world over the period of 1950–1993 (Easterling et al., 1997). A decreasing trend in DTR was noted also for the Tibetan Plateau during 1961–2005 (You et al., 2008). However, a different picture emerges over South Asia in Figure 5; both positive and negative trends are found. Moreover, the dispersion of DTR trends increases with elevation, such that there is a statistically significant positive slope of the regression between DTR trends and altitude. Existence of both positive and negative trends in DTR over South Asia was also reported in Figure 2 of Easterling et al. (1997) over 1950–1993. The slope of the regression in Figure 5 is consistent with the increasing slope of TXx (Figure 3(d)) and the negative slope of TNn (Figure 3(a)), reported above. The large spread of trends in DTR at high elevations suggest a mix of changes in cloud cover at these sites, as DTR is correlated with local cloud cover (Easterling et al.1997).
3.1.5. Mean trends for categorized elevation
Most of the stations analysed in the paper are low-elevation sites (Figure 1) and, therefore, they may not have dependency between the warming trend and altitude in this range. Therefore, average trends are computed for a categorized elevation rank for four different categories: (1) < 500 m; (2) 500–1000 m; (3) 1000–1500 m; and (4) > 1500 m.
Table II show higher magnitudes mean trends for the indices of extremes in maximum temperature. viz. TX10p, TX90p, WSDI and TXx over the high-altitude category indicating enhanced warming over high-altitude stations through extremes in maximum temperature. Indices of extremes in minimum temperature viz. TN10p, TN90p, CSDI and TNn also show the warming over the high-altitude category, however, the magnitudes of mean trends over high altitude category are lower compared to low altitude category. The higher rate of warming through maximum temperature has caused diurnal temperature range to increase over the high-elevation category.
Table II. Mean values of trends for categorized elevation rank changes in index trends with altitude
Categories of Altitude (m)
3.1.6. Difference between low- and high-altitude trends
The results presented above show that, in general, trends at lower-altitude stations are consistent with a warming climate. However, stations at higher altitudes show mixed trends. As an illustrative example, we again consider warm nights (TN90p), which we can expect, to have positive trends in a warming climate. This is seen to be so at stations up to a height of nearly 20 m in Figure 2(c), but at greater heights both positive and negative trends are obtained, although the majority of stations have positive trends. To investigate this question further, Figure 6 displays the geographical locations of stations at altitudes greater than 500 m where the trend magnitude is > 1.5/ year. It is seen that these high-altitude stations in peninsular India and Sri Lanka have positive slopes, i.e. this region shows geographical homogeneity and consistency with global warming. Regional homogeneity of positive slopes is also observed over Nepal in the eastern Himalayas. A relatively high rates of warming in all of Nepal were found (Shrestha et al., 1999) due to the high rates of warming in high-elevation areas of the Himalayas and Middle mountain regions. Similar warming trends are also observed in the Tibetan Plateau. Liu et al., 2002 show that warming is more pronounced in higher-altitude stations than in lower ones in the Tibetan Plateau. Reduction of snow and glacier cover in the Himalayas may also be contributing to higher rates of warming observed in the higher-elevation regions (Kadota and Ageta, 1992; Yamada et al., 1992; Fujita et al., 1997; Jin et al., 2005). Change in the surface albedo of the region due to reduction in snow and glacier cover in high elevation will increase the surface air temperature, thereby acting as a positive feedback mechanism (Meehl, 1994). However, mixed trends are seen over central India and in northern Pakistan in the Greater Himalayas. Cooling history of the Himalayan ranges of northern Pakistan is largely a function of uplift and erosion (Zeitler, 1985).
3.2. Changes in extreme temperature indices versus latitude
As seen in Figure 1, South Asia spans over 30° of latitude, providing a good opportunity to study relationships of time trends in the extreme temperature indices with latitude. Similar to the analysis with altitude, a straight-line regression was fitted to plots of latitude versus trends in extremes at stations. The results are summarized in Table III, with some of the more interesting results discussed below.
Table III. Changes in index trends with altitude and latitude
Changes in index trends per log altitude (m)
Changes in index trends per log altitude based on median regression
Changes in index trends per 10° latitude
Changes in index trends per 10° latitude based on median regression
Figure 7(a) shows the variation with latitude for trends in TX10p (percentage of cold days); the slope of the regression line shows an increase with latitude and is statistically significant. However, it should be noted that about 60% of the South Asian observation stations are located between 20°N and 30°N, and this concentration of data points affects the regressions with latitude. We see in Figure 7(a) that the majority of stations at low latitudes have registered decreasing trends in the percentage of cold days, as we would expect in a period of global warming; however, this signal is mixed at high latitudes. This difference between the low and high latitudes gives rise to the positive slope in Figure 7(a).
A similar behaviour is seen in Figure 7(b) for percentage of hot days (TX90p). Generally, positive trends are calculated for up to about 15°N (i.e. Sri Lanka and south India), with mixed trends at higher latitudes. This results in a negative slope of the regression.
There is a general increase in WSDI at low latitudes in Figure 7(c), with mixed trends at higher latitudes, resulting in a statistically significant slope of the regression between WSDI and latitude. By contrast, the slope of the regression for cold spell durations (CSDI) in Figure 7(d) is negligible because the trends in this index are mixed across latitudes.
In all the relationships with latitude shown in Table III, a statistically significant slope was detected when there was a clear signal of climate warming in Sri Lanka and south India. The stations in Sri Lanka and south India experience greater oceanic influence than stations at higher latitude, and this may partly explain the difference in trends in extreme temperature indices between low and high latitudes in Figure 7.
3.3. Median-based regressions
In regression analysis, an assumption is made that the residuals (i.e. departures in the data from the fitted line) are random variables with zero mean, follow a Gaussian distribution, and the variance of the residuals does not change with the independent variables. It is apparent from Figures 2–5 and 7, that these assumptions are violated in the extreme trend data. We can see that in many of the figures, the variance is not constant, and the data are not symmetrically distributed around the fitted line. Outliers are also seen in some of the figures. The reliability of the fitted slope is therefore questionable and should be investigated further with robust statistics that do not depend on the assumptions of classical statistics (Wilks, 2006; Cleveland, 1979).
We have repeated the regression calculations described above using median-based regression methods to obtain robust estimates of the slopes. The x-axis in each case was divided into a number of equally spaced bins, and the median of the dependent variable values in each bin was calculated. A straight line was then fitted to the medians. The slope values obtained using median-based regressions are presented in Table III alongside the standard calculations. Generally, the results are very similar between the slopes obtained using the standard and median-based regression for the variation in trends with altitude. Both methods identify the same regression slopes as significant, apart from the TXx index, which the median-based method assesses as non-significant. This similarity reflects a relatively even distribution of stations with height. The results between the two regression methods are less similar for the variation across latitude. The sign of the regression remains the same in each case, but different conclusions are reached for the significance of some of the regressions. This reflects the uneven distribution of stations across latitudes. Nevertheless, the discussion of variations in trends across the latitudes described above remains valid in a qualitative sense.
South Asia encompasses weather observation stations located from sea level to nearly 4000 m above sea level, and over 30° of latitude from 6°N in Sri Lanka (Galle) to 36°N in Pakistan (Chitral). This paper has examined trends in a number of indices of temperature extremes over South Asian stations to determine whether there are significant changes in trends in the indices with altitude and latitude.
This study reveals that, in general, trends at lower-altitude stations over South Asia are consistent with a warming climate. However, stations at higher altitudes often show mixed trends. For example, the percentage of cold nights (minimum temperatures in the lowest 10th percentile) per year shows increases at many stations at higher altitudes. Similar results are seen in spell-duration indices which are consistent with warming at lower altitudes, but less coherent at higher elevations. For example, many high-altitude stations show increases in cold-spell duration, and decreases in warm-spell duration. Trend magnitudes for DTR and intensity indices also show comparatively higher spread at high altitudes.
Similar results are seen for changes in trends of extreme indices with latitude. Changes in extremes at lower-latitude stations are often consistent with a warming climate, whereas many stations located at higher latitudes show opposite behaviour. This partly reflects a tendency for high-altitude stations in South Asia to also be located at relatively high latitudes, as seen in Figure 1.
Daily temperatures at mountain locations are influenced by local topography which determines the relative impacts of free air advection and the local radiation budget. It is possible that inclusion of information on local topography can explain some of the inconsistencies found at high-altitude stations. Also, many of the mountain stations are at remote locations, and consistency of data quality is an issue that needs to be investigated further. Sheikh et al. (2010) notes some regional differences in the trends of temperature extremes across South Asia, such as more pronounced warming in the eastern Himalayas compared to the Greater Himalayas.
The results of this study suggest that low-altitude locations in South Asia can expect changes in temperature extremes that are generally consistent with broad-scale warming. However, high-elevation sites appear to be more influenced by local and regional factors and, hence, future changes in temperature extremes may be less predictable for these locations.
The work in this paper was supported by the Asia Pacific Network (APN) for Global Change Research), Kobe, Japan through a 3-year project entitled ‘Development and Application of Climate Extreme Indices and Indicators for Monitoring Trends in Climate Extremes and their Socio-economic Impacts in South Asian Countries’ awarded jointly to the Global Change Impact Studies Centre (GCISC) and Pakistan Meteorological Department (PMD) in 2005. We acknowledge with thanks the role of APN in facilitating this important research work for South Asia.