We analyze meteorological time series of four AWS located close to the GCN Ice Cap and one conventional weather station (WS) at Punta Arenas in Chile (Figure 1). An overview of the pertinent data is given in Table 1. Detailed information about the geographical setting of AWS Paso Galería (PG), AWS Puerto Bahamondes (BH) and AWS Estancia Skyring (SR) is given in Schneider et al. (2003). In addition, a fourth AWS named Arévalo (AR) has operated since September 2007 and is located at the northwestern side of the ice cap at about 58 m a.s.l.

The AWS are manufactured by Campbell Scientific Ltd. (United Kingdom). The similarly designed stations are equipped with sensors of air temperature, air humidity, precipitation, solar radiation, wind speed and wind direction (Table 2, Figure S1). All meteorological variables but precipitation are measured at 2 m above the surface. Precipitation is measured at 1 m above the ground using unshielded tipping-bucket rain gauges. This type of measurement underestimates precipitation by up to 30% at wind speeds of 1.5 m s⁻¹ and even up to 50% at wind speeds of 3.0 m s⁻¹ (Rasmussen et al., 2012; Buisán et al., 2017). Wind-speed induced deviations increase during snowfall due to an intensified drifting of snow (Rasmussen et al., 2012; Buisán et al., 2017). Substantial deviations are further suspected during high precipitation events, since strong precipitation events are often accompanied with strong wind gusts. During periods with very strong winds, the tipping bucket tends to shake which leads to an overestimation of precipitation. Correction factors for this kind of inaccuracy are not available. According to Schneider et al. (2003), the precision of the precipitation measurement at AWS BH and AWS PG is estimated to be about ±20%. We used this number as estimate of uncertainty for all AWS in this study. Wind-speed induced deviations have been minimized due to an additional fixation of the rain gauges in 2010 at AWS PG and in 2016 at AWS BH to avoid the shaking of the tipping bucket. Overall, precipitation data must be interpreted with care bearing in mind that these data are subject to considerable uncertainty in the order of potentially several tens of percent. Meteorological observations at the conventional WS Punta Arenas (PA) hold the longest and most complete time series. These observations (2000-2016) are measured according to the World Meteorological Organization standards. Data are provided by the Laboratory of Climatology, Instituto de la Patagonia, Universidad de Magallanes. Yearly records are published in Anales del Instituto de la Patagonia, Chile (e.g., Butorovic, 2016).

The quality of the time series is assured by performing some basic quality integrity and outlier checks (Durre et al., 2010). In case of unrealistic values, such as relative humidity larger than 100%, identical values for maximum and minimum air temperature, or a repetition of data records within a month, data are excluded. Furthermore, records are checked for daily outliers (Durre et al., 2010) and missing values within a month. Months with missing values of daily observations are not further considered. In case of precipitation, this leads to an exclusion of daily precipitation observations at AWS BH in 2000/01 and at AWS PG in 2008/09. In addition, failures of power supply or sensors due to harsh weather conditions cause substantial data gaps in addition. Figure 2 shows the available observation periods for each station on a monthly basis as used in this study.

For mesoscale weather type classification, daily mean sea level pressure (MSLP) reanalysis data are obtained from the ERA-Interim dataset provided by the European Centre for Medium-Range Weather Forecasts (ECMWF) (Dee et al., 2011). The data sets are downloaded with a spatial resolution of 0.75° × 0.75° grids and 12-hourly temporal resolution (0 a.m., 12 p.m. UTC) between 1979 and 2016 for the area 10°S −80°S and 40°W −110°W covering most of South America and the Antarctic Peninsula. The selected area as well as the choice of MSLP as weather classification variable followed the weather type classification as used by Frank (2002) and Schneider et al. (2003).

To reveal the impact of ENSO and the AAO on climate variability in Southernmost Patagonia, we compare the monthly anomalies of observations with the Oceanic Niño Index (ONI) and the Antarctic Oscillation Index (AAOI). Both indices are provided by the U.S. National Oceanic and Atmospheric Administration (NOAA) - Climate Prediction Center.

ONI identifies El Niño and La Niña events in the tropical Pacific by using a 3-month running mean of ERSST.v5 sea surface temperature anomalies in the Niño 3.4 region (5°N-5°S, 120°-170°W) (Barnston et al., 1997; Huang et al., 2017). Events are defined as five consecutive overlapping 3-month periods with anomalies ≥+0.5°C for El Niño and ≤ -0.5°C for La Niña events. Data is available from http://www.cpc.ncep.noaa.gov/data/indices/oni.ascii.txt.

The monthly AAOI is constructed by projecting the monthly mean 700 hPa height anomalies poleward of 20°S onto the first empirical orthogonal function mode of monthly mean height anomalies at 700 hPa (Thompson and Solomon, 2002). Positive phases are characterized by a strengthening and poleward shift of the westerlies with decreased surface pressure and geopotential height over the fringe of Antarctica. During negative phases, opposite conditions prevail. Monthly data of AAOI can be obtained from http://www.cpc.ncep.noaa.gov/products/precip/CWlink/daily_ao_index/aao/aao_index.html.

Climate extremes of air temperature and precipitation (Table 3) are described using selected indices of extremes (Tank et al., 2009). In case of air temperatures extremes, time series of maximum (TX) and minimum (TN) air temperature are analyzed regarding the number of days with ice (TX ≤ 0°C) (ID) and frost (TN ≤ 0°C) (FD). Percentiles are calculated based on the available observations period (Figure 2) to define the number of days with TX < 10th percentile (P1TX), TN < 10th percentile (P1TN), TX > 90th percentile (P9TX) and TN > 90th percentile (P9TN). Furthermore, we are interested in the maximum length of a cold spell (CS) with consecutive days of TN < 10th percentile, and the maximum length of a warm spell (TWS) with consecutive days of TX > 90th percentile per year.

Precipitation extremes are described by the absolute maximum of 1-Day (PDX) and 5-Day (P5DX) precipitation (mm) as well as the annual amount of daily precipitation larger than the 95th percentile (P95P) and the 99th percentile (P99P). The maximum number of days with daily precipitation higher than 20 mm (HP) and below 1 mm (DD), and the maximum length of wet spells (PWS) and dry spells (DS) per year are also examined.

Weather types are classified by applying the automated classification map-pattern scheme by Lund (1963) based on MSLP. The LUND-algorithm is based on so-called leader patterns. It constitutes a low computational cost predecessor of other cluster algorithms (Hartigan, 1975; Murtagh, 1985). Within the LUND-algorithm, in a first run a leader is defined as the one observation showing most similar cases (observations) of correlations higher than a given threshold with all other observations (Lund, 1963; Philipp et al., 2014a). After removal of the leader and its matching cases, e.g., all observations that correlate with the identified leader above the set threshold Pearson correlation coefficient, the remaining class leaders are called in similar iterative processes (Lund, 1963; Philipp et al., 2014a). After the determination of all leaders, single days are assigned to a specific weather type by returning all observations to the data pool and assigning each one to the nearest leader based on linear correlation (Lund, 1963; Philipp et al., 2014a). The threshold Pearson correlation coefficient for finding leader weather types in this study is set to 0.85. The Lund classification scheme accentuates the dominance of few weather types with westerly air flow. Subsequently, it finds further weather types that occur only rarely but are important and significantly different patterns. In contrast, k-mean leader algorithms tend to locate centroids in such that all classes become equally large (Philipp et al., 2014b). The latter would have resulted in many undesirably similar weather types of principally westerly airflow.

Weather type classifications are calculated from the ERA-Interim MSLP reanalysis data for the period 1979–2016. We apply the cost733class-1.2 software (http://cost733.geo.uni-augsburg.de/cost733class-1.2) originally designed for circulation type classifications in Europe (Philipp et al., 2014b). We checked for the optimal number of classes using a k-means leader algorithm by evaluating the increase of explained cluster variance with increasing numbers of cluster. The number of 10 weather types was found to be optimal which is in accordance with the weather type classification of Frank (2002), used by Schneider et al. (2003). Earlier weather type classifications document only six weather types (Endlicher, 1991; Compagnucci and Salles, 1997). Figure 7 shows the resulting 10 MSLP centroids obtained from the classification procedure after Lund (1963) as used in this study.

Annual and seasonal trends in air temperature based on the 17-year record of AWS PG, AWS BH, AWS SR, and WS PA are analyzed. Air temperature changes over the given time period can further be related to observed recent changes of the GCN Ice Cap outlet glaciers.

Trends are detected using the non-parametric Mann-Kendall (MK) and Sen slope estimator (SSE) trend test. The combined MK and SSE test has been frequently used to quantify the significance and magnitude of trends in climatological and hydrological time series (e.g., Hamed, 2008; Gocic and Trajkovic, 2013; Kisi and Ay, 2014; Onyutha et al., 2016). The MK test is a non-parametric test to identify linear and non-linear trends in time series (Mann, 1945; Kendall, 1975). The test does not require normally-distributed input data and has a low sensitivity to abrupt breaks due to inhomogeneous time series. The magnitude of the trend in terms of slope is the robust estimate of the median following the approach by Sen (1968) and Theil (1950).

The elimination of autocorrelation in time series analysis is essential because otherwise autocorrelation increases the chances of detecting a significant trend even in case trends are absent. Therefore, Yue et al. (2002) proposed the Trend-Free Pre-Whitening (TP) method in case both trend and lag-1 autocorrelation exist in the data record.

To detect trends in the time series we followed the described steps below:

testing the significance of the lag-1 autocorrelation (AC) in each time series (step 1)

The Trend-Free Pre-Whitening procedure (step 2) includes the following steps. First, the apparent linear trend of the time series X_i is removed. Afterward, the lag-1 correlation coefficient r_ac of the detrended time series X_DE,i is determined. The lag-1 AC is then eliminated from X_DE,i by:

The removed linear trend is then added to X_DE,A,i to obtain the final blended time series to which the MK and SSE trend detection test is applied.

Time series of air temperature, relative humidity, precipitation, wind speed and wind direction are analyzed regarding their daily, monthly and annual means, and anomalies. Mean annual and monthly values of the meteorological variables for each station are listed in Table S1 in the supplement. Annual means are averaged based on monthly mean values. The study period of each time series relates to the available observation period shown in Figure 2.

Annual and seasonal trends in mean (T), minimum (TN) and maximum (TX) air temperature are analyzed for AWS PG, AWS BH, AWS SR, and WS PA as described in section 3.3. Significant lag-1 autocorrelation was only detected for TN of AWS BH which was corrected following the TP procedure before trend detection. An overview of the trend analysis results is given in Table 5.

Trends in annual mean air temperature for AWS PG and AWS PA are significant at the 90% level with a trend magnitude (Sen Slope) of +0.05°C a⁻¹ and +0.02°C a⁻¹. An increasing trend of seasonal mean air temperature is only apparent during winter (JJA) at AWS BH and WS PA. A significant upward trend is detected in the annual TN at AWS PG of +0.31°C a⁻¹ between 2000 and 2016. This upward trend is also reflected in different seasons as well. TN increases significantly during spring (SON), winter and fall (MAM) at AWS PG, and during summer (DJF), spring and fall at AWS BH. Three of four stations show upwards trends of TN during spring and fall. In contrast, significant downward trends of TX during the summer season are determined for AWS PG, AWS BH and AWS SR. The largest trend of −0.24°C a⁻¹ (2000 and 2016) was found at AWS PG. No significant trend has been detected in annual and seasonal TX time series for WS PA.

Ten major types of synoptic weather patterns have been classified for Southernmost Patagonia (Figure 7). Weather type 1 is characterized by a strong north-south pressure gradient resulting from a distinct high-pressure cell over the Pacific and low-pressure trough over the Bellingshausen Sea, which leads to intensified westerlies. During weather type 2, the high-pressure cell is weaker and shifted toward the south, while the low-pressure cell is located south of Tierra del Fuego in the Weddell Sea. Cold and humid air masses from the south-west are advected to Southernmost Patagonia (Frank, 2002). A similar circulation pattern to type 2, but with a distinct low-pressure cell located south of Tierra del Fuego in the Drake Passage occurs during weather type 5. A north-westerly advection of humid and warm air occurs when both, the Pacific anticyclone and the low-pressure cell, are shifted toward the north-west of southern South America compared to type 1. Pressure gradients become weaker during weather type 9 while the principal synoptic configuration is mostly the same.

The atmospheric circulation of weather type 4 is dominated by an intensified high-pressure cell in the southern Pacific reaching far toward the eastern side of the Southern Andes. The persistence of this situation leads to warmer and dryer conditions. A southward shift of the Pacific anticyclone with the advection of colder air masses from south-west, is present during weather type 7. A strong high-pressure cell in the southern Atlantic accompanied with low pressure in the southern Pacific is present during weather type 6. The Atlantic anticyclone is stronger and shifted toward the north in case of type 8. Both weather patterns imply a northerly or westerly airflow with increased air temperatures. The circulation pattern of weather type 10 is dominated by a strong high-pressure cell located far toward the south in the Pacific leading to a cold and humid airflow from the south over Southernmost Patagonia.

Major synoptic weather types with distinct low-pressure cells in the Weddell Sea (types 2 and 5) or Bellingshausen Sea (types 1, 3, and 9), causing a prevailing southwesterly, northwesterly or westerly airflow, account for 68% of all days in Southernmost Patagonia. In 32% of all situations, high pressure cells over the southern Pacific (types 4, 7, and 10) and southern Atlantic (types 6 and 8) determine the weather of Southernmost Patagonia. Circulation pattern type 1 with a strong zonal westerly air flow is the most common type of all weather patterns, occurring 30% of the year (Figure 8, Table 6). During the austral summer the frequency of weather patterns with westerly air flow increases to 76% compared to high-pressure influenced weather types (24%). During the austral winter anticyclonic circulation patterns account for 40% of all situations, while low-pressure weather types occur less. This is in accordance with the generally lower wind speeds found in winter at all weather stations in this study.

Anomalies of observed air temperature, precipitation, wind speed, solar radiation, and relative humidity are analyzed according to the major synoptic types of weather patterns for all five stations (Figure 9). Similar types of weather patterns are grouped into classes as listed in Table 6. Large positive air temperature anomalies are accompanied with strong westerlies (class 1) at all five stations. Air temperature anomalies at AWS BH, AWS AR, and AWS PG however are smaller compared to AWS SR and WS PA due to the influence of the nearby GCN Ice Cap and proximity to the Pacific Ocean. The impact of strong westerlies on air temperature anomalies is largest in the austral fall. On days with southwesterly airflow (class 2 and 5), air temperatures slightly drop at the stations located close to GCN Ice Cap, while air temperatures tend to rise during northwesterly airflow (class 3 and 9). The advection of colder air masses from the southwest (class 2 and 5) largely determines the air temperature variability during the summer and fall season.

The impact of circulation patterns with intensified high-pressure cells in southern Pacific and Atlantic (classes 4 and 7, 6 and 8 and 10) on air temperature variability is in general low, but highest for AWS SR and WS PA. The advection of air masses from the north and northwest under high-pressure influence (class 6 and 8) cause warmer conditions at all stations in SON and DJF, while temperatures decrease during the austral winter except for AWS AR which is the only station located clearly west of the Cordillera. The advection of cold air masses from the south (class 10) causes a significant decrease of daily air temperature at all stations with the largest negative anomalies for WS PA and AWS SR. During fall and spring, air temperatures drop significantly below the specific means (class 10).

Circulation patterns highly impact the precipitation variability at the GCN Ice Cap during all seasons. The orographic induced uplift of the humid air masses from west and southwest (classes 1 and 2 and 5) comes along with year-around positive precipitation anomalies at the GCN Ice Cap. Precipitation amounts decrease significantly during high-pressure influences, in case of classes 6 and 8 and 10. The stations located at GCN (AWS AR, AWS BH, AWS PG) obtain more than 80% of daily precipitation during westerly airflow. East of the GCN Ice Cap at AWS SR and WS PA, mean daily precipitation amounts are similar during all weather types. Precipitation increases slightly on days with southwesterly air flow at AWS SR, while at WS PA daily precipitation increases on days with north to northwesterly wind directions (classes 3 and 9 and 6 and 8). The annual precipitation amounts at AWS SR and WS PA are still mainly determined by the airflow from west (class 1) and south west (class 2 and 5) despite the fact that these stations are clearly located on the leeward of the Cordillera during westerly air flow but extreme precipitation rates are rare. Of all daily precipitation extremes (higher than the 90th percentile) 80% can be associated with classes 1 and 2 and 5 at the GCN stations, while the percentage decreases to 70% at AWS SR and 51% at WS PA. High precipitation events at WS PA are also present during the class 3 and 8 with 20 and 13% of the days, respectively. This finding indicates that precipitable water in the air column is carried across the Cordillera by strong upper air westerly winds. During classes 1 and 2 and 5, both stations experience often foehn-type weather conditions with cloud dispersal due to the subsidence of air masses on the lee-side of the Cordillera. High incoming solar radiation can therefore be observed at AWS SR during these types of weather patterns. Since the westerly air flow is also accompanied by higher air temperatures, relative humidity decreases at all stations, but strongest at AWS SR and WS PA.

Wind speed is highest during westerly (class 1) and southwesterly (class 2 and 5) airflow for all stations. AWS AR shows negative wind speed anomalies on days with strong westerlies which are caused by the location of the AWS and local effects due to the surrounding topography as described in section 4.1.2. Wind from north-northwest (classes 3 and 9 and 6 and 8) is in general weaker on days with high-pressure influence (class 6 and 8).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.