Negative elevation-dependent warming trend in the Eastern Alps

Mountain regions and the important ecosystem services they provide are considered to be very vulnerable to the current warming, and recent studies suggest that high-mountain environments experience more rapid changes in temperature than environments at lower elevations. Here we analysed weather records for the period 1975–2010 from the Eastern Italian Alps that show that warming occurred both at high and low elevations, but it was less pronounced at high elevations. This negative elevation–dependent trend was consistent for mean, maximum and minimum air temperature. Global radiation data measured at different elevations, surface energy fluxes measured above an alpine grassland and above a coniferous forest located at comparable elevations for nine consecutive years as well as remote sensing data (MODIS) for cloud cover and aerosol optical depth were analysed to interpret this observation. Increasing global radiation at low elevations turned out to be a potential driver of this negative elevation–dependent warming, but also contributions from land use and land cover changes at high elevations (abandonment of alpine pastures, expansion of secondary forest succession) were taken into account. We emphasise though, that a negative elevation–dependent warming is not universal and that future research and in particular models should not neglect the role of land use changes when determining warming rates over elevation.


Introduction
Mountain regions provide several important ecosystem services, like the protection of human infrastructure or constant supply with high quality drinking water. 40% of Europe's fresh water originates in the Alps supplying tens of millions of Europeans in lowland areas (www.eea.europa.eu). They are important centres of biodiversity and also key targets for tourism and recreation. But the provision of these key services is threatened by climatic changes (Beniston et al 1997, Rangwala et al 2013).
The term 'elevation-dependent warming' (EDW) typically refers to different warming rates over an altitudinal gradient that might also deviate from the overall global warming rate. This aspect becomes particularly important in the light of the accelerating global warming recorded during the last decades (Hartmann et al 2013). Mountain regions which are considered as vulnerable ecosystems (Beniston et al 1997) could experience graver changes earlier than the nearby lowlands, if the warming rate is higher in the mountains. For this reason several mountain areas of the world were studied using both, weather station data and modelling approaches, sometimes with contrasting results. Snow/ice cover, clouds, water vapour, aerosols, and soil moisture have been considered as drivers of elevation-dependent warming (Rangwala and Miller 2012). A recent study and metaanalysis by Pepin et al (2015) highlighted a trend towards accelerated warming in high elevations, but the authors also admit that this faster warming cannot be considered as universal. Zeng et al (2015), in a global perspective, reached an opposite conclusion for some particular mountain regions of the world. More detailed and local analyses using observational data also led to contradictory findings. In some cases a positive elevation dependency of warming was found (Diaz and Eischeid 2007, Liu et al 2009, Wang et al 2013, Acquaotta et al 2015, Wang et al 2016 while in other cases opposite results were reported (Pepin & Losleben 2002, Ceppi et al 2012, Kirchner et al 2013, Philipona 2013. Other studies did not find a clear dependency of the warming rate on elevation (You et al 2010, Marty and Meister 2012, Gilbert and Vincent 2013, Gevorgyan 2014, Oyler et al 2015. When regional and global climate models were applied the results seemed to converge towards a positive correlation between elevation and warming rate (Qin et al 2009, Gobiet et al 2014, Rangwala et al 2013, Kotlarski 2015, Rangwala et al 2015. This paper reports a series of observations made in the southern part of the Eastern Alps (Province of Trento, Italy) over the last three decades. More specifically it examines how air temperature varied with the altitude and tries to identify possible drivers which lead to the occurrence of the significant negative elevation-dependent warming. This is achieved by bringing together weather records, energy flux measurements, satellite MODIS-data for cloud cover, atmospheric optical depth and inventorial data concerning recent land cover changes in the region.

Study area and land use
This study was realized in Trentino, a region located in the Central-Eastern Italian Alps, which covers an area of 6212 km 2 comprised between 45°45′-46°30′N and 10°30′-12°00′E (figure 1). This region is characterized by a valley system, the most prominent being the Valley of the river Adige, with a typical elevation range, from bottom to peaks, of about 2000 m. Overall elevation ranges from 70 m to 3769 m. Annual forest inventory data of the last three decades were provided by the 'Servizio Foreste e Fauna' del 'Dipartimento Territorio, Ambiente e Foreste' della Provincia Autonoma di Trento. These data reported the total surface occupied by forests as well as the stocking timber volume.

Weather data
Homogenised weather data series (Eccel et al 2012) for mean (T mean ), maximum (T max ) and minimum (T min ) daily air temperatures, and incoming radiation were obtained from the Meteorological service of the Autonomous Province of Trento-PAT and from Fondazione Edmund Mach. In total a number of 28 weather stations between 84 and 2125 m a.s.l. were considered (supplementary table S1) for the period January 1975 until December 2010. The data for incoming radiation were available for a lower number of stations and years (some low elevation stations have data starting in 1983 while the records for high elevation stations only start after 1999) (supplementary table 1, supplementary figure 3). The weather stations were first divided into low and high elevation using the median of the elevations as separation threshold. Changes in mean air temperature for the considered time interval were studied for the two groups. Furthermore altitudinal gradients of temporal trends were estimated for T mean , T min and T max by linear regression. Two seasonal sub-periods were also considered: from November to April (A) and from May to October (B). November till April (period A), is the period when snow cover prevails at altitudes above the median elevation of the two groups with some obvious variability among different years. Changes in global solar radiation were examined for 3 stations located at low elevation (supplementary table 1, supplementary figure 3) (San Michele-205 m, Arco -84 m a.s.l., Trento Sud-185 m a.s.l) and 3 located at high elevation (Monte Bondone-1552 m a.s.l., Baselga di Pinè-983 m a.s.l., Tremalzo-1560 m a.s. l.). A data-driven procedure was used to calculate changes in the transmissivity of the atmosphere at each station. The values of the first year in each series (1983 for low elevation and 2000 for high elevation) were used to construct the daily courses of theoretical clear sky days month by month. For that purpose we assembled the highest radiation values measured for each hour of the day within a month (excluding the twilight hours) to a daily course and corrected the values according to the Sun declination for each day. Finally, the monthly and then yearly courses of hourly radiation under clear sky conditions were reconstructed for each individual station. The difference (ΔRad ) and the ratio (Rrad) between each daily mean of incoming radiation and the corresponding maximum radiation under clear sky conditions in the first year of the series were calculated for the whole dataset (1983-2010 for the low elevation stations and 2000-2010 for the high elevation stations). This procedure yielded robust data to assess changes in radiative input over the years for different elevations, minimising the effect of different aspects, local topography and possible errors due to imperfect crosscalibration of the solar sensors. in order to obtain half-hourly data. Daily averages of latent and sensible heat flux (LE and H), net radiation (Rn) and albedo were calculated for both ecosystems over a 9 year period. In order to compare the air temperature and their changes at different heights, daily potential air temperatures (T Pot ) were calculated according to equation (1).
where T 0 is the absolute temperature at the higher pressure in K, T is the absolute temperature at the lower pressure in K, p 0 is the higher pressure, and p is the lower pressure. Aerodynamic resistance to water vapour transfer Ra,v was computed using the equation: Where R a m , is the resistance to momentum transfer (s m −1 ), R b the resistance of the quasi-laminar boundary layer resistance (s m −1 ) (Verma 1989, Baldocchi andMa 2013), u is wind velocity in m s −1 , k is von Karman's constant (0.4), u * is friction velocity in m s −1 , Sc is the dimensionless Schmidt number and Pr is the Prandtl number.

Remote sensing data
Monthly time series of MODIS Terra Level 3 data for cloud fraction (Monthly MOD08_M3.051) (Acker and Leptoukh 2007) were obtained from NASA GES DISC (http://disc.sci.gsfc.nasa.gov/giovanni). Day and night time cloud cover was analysed in comparison for different paired valley and mountain areas (see supplementary table S2, supplementary figure 2) for the whole period since MODIS was operational (March 2000-December 2014). Averages were calculated for the mountain and valley areas and the differences between the two categories (mountainvalley) were analysed over time (figure 6). Monthly means for aerosol optical depth (AOD) were retrieved form neo.sci.gsfc.nasa.gov (Product ID MOD04/ MYD04) for the period 2003-2013 and for pixels corresponding to the same location of the weather stations used for irradiance analysis. The AOD trend over time was studied and furthermore correlated with the difference between the measured incoming radiation and the corresponding clear sky radiation in the first year of the series (ΔRad).

Results
Weather data for the study region (figure 1) show that over the last 35 years annual mean air temperature increased both at low and high elevation, but the rate of change was not equal (figure 2). On average the rate of temperature increase was significantly higher at low compared to high elevation (0.49 ± 0.12 and 0.27 ± 0.15°C decade −1 , respectively). A significant negative correlation was found between the elevation and the observed rate of change in air temperature over the last 35 years (figure 3). This trend is consistent for the different seasons and for the maximum, minimum and mean daily temperature values (T max , T min , T mean ). For the two periods considered (A: November-April, and B: May-October 1975 to 2010) the warming rate of the mean temperatures (ω, Τ mean ) decreased with height by 0.26 ± 0.11°C decade −1 km −1 and 0.22 ± 0.08°C decade −1 km −1 , respectively. The decrease in warming rate with height for minimum temperatures (ω Τ min ) was 0.20 ± 0.11 decade −1 km −1 for the snow covered period (A) and 0.15 ± 0.09°C decade −1 km −1 for the snow-free period (B). The corresponding values for maximum temperature (ω T max ) were 0.33 ± 0.17°C decade −1 km −1 and 0.28 ± 0.16°C decade −1 km −1 for the snow covered period (A) and for the snow free period (B), respectively. It is worth noting that in all cases the decrease in warming rate with height was more pronounced in the snow covered period (A). The daily deviations of radiative input from the potential input under clear sky conditions at the beginning of data recording (ΔRad) decreased for the low elevation stations on average by 1.68 ± 0.25 Wm −2 year −1 for the period 1983-2010 (figure 4). Instead, an increase of 2.7 ± 1.3 W m −2 year −1 (figure 4) was observed at high-elevation stations, even though for a shorter period (2000)(2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010).The changes in ΔRad were more pronounced for clear sky days in all elevations (table 1).
Significant differences were also observed regarding the turbulent fluxes above two representative high elevation land use types from Trentino: 'coniferous forest' (covering 50% of the total area of the province) and 'alpine grassland' (17% of the total area of the province) (Servizio Foreste e Fauna PAT) located at comparable elevations (1349 and 1553 m a.s.l, respectively). When averaged over the whole measurement period of 9 years, the latent heat flux was higher in the evergreen forest by 10.8 ± 1.7 W m −2 (figure 5(a)), with the largest positive mean difference in March (56.4 W m −2 ) and the most negative in late May (−33.2 W m −2 ). The sensible heat flux over the forest was higher than over the grassland for the entire measurement period (38.2 ± 2.6 W m −2 ) except for late autumn and winter, when the differences were negligible (figure 5(b)). Overall the total turbulent energy exchange (H + LE) was higher over forest, which clearly represents the rougher vegetation form with the lower resistance to turbulent transfer (figure 5(d)).
The difference in potential air temperature between the forest and the grassland (calculated as forest minus grassland), averaged for the 9 year period on a daily basis, was 0.13 ± 0.04°C (figure 5(e)). T Pot was higher over the forest during late autumn and winter with the highest value in January (+1.60°C) but lower from early spring to mid-November. The highest negative difference was recorded in October (−0.63°C). For the snow free period (B) the difference was negligible (−0.009°C ± 0.1) while for the snow covered period (A) it equalled 0.25 ± 0.16°C. When averaged on a half-hourly basis, the air above the forest stand turned out to be cooler around noon and in the afternoon hours during the snow free period (period B), but clearly warmer at night time all year long. The temperature amplitude was always smaller in case of the forest (supplementary figure 1). The albedo values were consistently lower for the forest, resulting in an overall 37.26 ± 2.03 W m −2 higher net energy gain ( figure 5(c)). During the snow free period mean albedo values of 0.10 ± 0.00 and 0.24 ± 0.01 were obtained for the forest and the grassland, respectively. The biggest difference of 0.72 between the two ecosystems was recorded in the snow covered period (A) ( figure 5(f)).

Discussion
In the period 1975-2010 the air temperature over land increased at a rate of 0.30°C decade −1 (Rangwala et al 2013). Our analysis of air temperature data from 24 meteorological stations along an elevation gradient revealed a significant negative correlation between the warming rate and the elevation for the same temporal span. This finding confirms previous results reported by Philipona (2013) for the Swiss Alpine region. The two studies are similar with respect to the large number of stations located at different altitudes, the elevation range and the time span covered. The decreasing warming rates (ω) over elevation were slightly more pronounced in the Swiss study with 0.52 and 0.35°C decade −1 for low and high elevations, respectively, which is 0.03 and 0.08°C decade −1 higher than in Trentino (table 2). These small differences, which fall into the confidence limits of the regressions, can be a result of the different time intervals considered and/or of the higher mean elevation of the Swiss weather stations for low and high elevation. Despite the robustness of both data sets, however, the concept of a negative elevation-dependent warming cannot be simply generalised to other mountain ranges of Europe and the world. Major inconsistencies remain with other studies as well as with a series of selected modelling approaches (Gobiet et al 2014, Rangwala et al 2013, Kotlarski 2015. The lack of an unequivocal positive or negative sign in ω definitely requires additional efforts to interpret the apparent contradiction in our results and to understand the causes of the observed negative trends. Both Philipona (2013) and Zeng et al (2015) attributed the occurrence of a negative ω to the increase in solar brightening at low elevations, a phenomenon which is generally attributed to an overall reduction in atmospheric pollution sources. In mountainous areas this reduction took place exclusively in the denser populated valleys. A substantial reduction of anthropogenic secondary aerosols and particulate matter in the atmosphere can actually lead to an increase in atmospheric transmissivity and thus to enhanced radiative forcing at the surface leading to warming   Table 1. Rate of change in mean annual difference between the measured incoming radiation and the corresponding clear sky radiation in the first year of the series (ΔRad, W m −2 ) for low and high elevation. s is the slope of linear regression for ΔRad over the studied period. Statistics are computed for three categories. All days (includes all the days of all the years), partially cloudy days (the ratio between the measured incoming radiation and the corresponding clear sky radiation in the first year of the series-Rrad-was >40% of the maximum Rrad), clear sky days (Rrad > 60% of the maximum Rrad). Low elevation: averages of Arco, Trento and San Michele, High elevation averages for Monte Bondone, Baselga di Piné and Tremalzo.

Low elevation
High elevation  Solar brightening at low elevation is per se sufficient to explain the observed negative sign of ω, but only under the assumption that no warming compensatory effects occurred at high elevations. In this context special attention has to be granted to implications associated with the observed land use change at high elevations in the Alps. The extensive reforestation programme after the end of the Second World War (Fuchs et al 2013) and more recently the wide-spread abandonment of alpine pastures led to a significant expansion and regrowth of forests ( Guidi et al 2014, Tasser et al 2007. Baur (2013) reports that the forest area increased by 8% in Switzerland between 1985 and 2005. In Trentino the forest area increased by 13.7% over the period 1976-2010, while the stocking forest biomass nearly doubled (+41.5%). 75% of the increase in forest area occurred at elevations above 800 m (Sitzia 2009).
It is well known that forests have a lower albedo than grasslands and meadows and this unavoidably leads to a higher absorption of shortwave solar radiation (de Wit et al 2014). For instance, the albedo of an alpine grassland in Tyrol ranged between 0.20 and 0.25 during the growing season (Hammerle et al 2008). Betts and Ball (1997) report a mean summertime value of 0.20 for grasslands. On the other hand, literature values for spruce forests, vary between 0.08 and 0.12 (e.g. Eck and Deering 1992, Betts and Ball 1997, Lukeš et al 2013. All the quoted albedo values are in good accordance with our findings (snow free period values of 0.10 ± 0.00 for the mixed Silver fir and Norway spruce forest and 0.24 ± 0.01 for the grassland). This difference could explain the large net radiation difference between an alpine grassland and a subalpine spruce forest in the Bernese Oberland (Switzerland) measured by Eugster and Cattin (2007). In comparison to the grassland the net gain of the darker forest canopy was a 110 W m −2 for the three summer months. So accordingly, a land use shift from pasture to forestry results in a higher net radiation and higher heat flux rates to the atmosphere, which becomes clearly evident from our results.
Compared to other land use types, the exchange of sensible and latent heat between forests and the   atmosphere is indeed much more efficient, as forests have a higher surface roughness leading to an enhanced mechanical turbulence (Baldocchi and Ma 2013). This reduced aerodynamic resistance to heat transfer can be noticed also from the data of the two flux towers used in this study ( figure 5(d)). Eugster and Cattin (2007) observed that most of the higher net energy input at the forest site was used for evapotranspiration resulting in cooler day time conditions. But this effect was more than offset by the fact that open areas like grasslands lose more sensible heat to cold-air drainage flows at night. As a consequence, on a seasonal basis, the mean air temperature above the subalpine coniferous forests was 1.1°C higher than above the grassland.
However, the relations might change with the admixture of deciduous tree species. Results from 27 Fluxnet sites in the temperate zone of Europe and the US show that the mean summertime (June till August) Bowen ratio for deciduous forests (0.43) was clearly lower than that of coniferous forests (1.07) expressing the higher water consumption (i.e. higher latent heat flux) of deciduous trees (Wilson et al 2002). Additionally, forests may have higher transpiration rates over a longer time compared to non-forested areas due to deeper rooting of the trees, which grants them access to water resources that are not accessible to grasses (Pielke 2005, Mahmood et al 2014. The resulting shift in energy partitioning towards latent heat flux is supposed to cool the air (Ban-Weiss et al 2011).
Finally, taking also wintertime into account, the presence of forests drastically reduces the surface albedo (e.g. Betts and Ball 1997) by shortening the overall duration of surface snow cover, while prolonged snow cover occurs at pastures and meadows. In summary the net effect of forest expansion on nearsurface air temperature is a balance between three components: albedo, surface roughness (by enhancing turbulent fluxes) and evapotranspiration. Past studies have shown that air warming effects of afforestation are likely to prevail in dry regions (Peng et al 2014)  The data from the two Eddy covariance stations in Monte Bondone (MBO, alpine grassland) and in Lavarone (LAV, mature coniferous forest), continuously operated for the period 2004-2012, provide reliable information in this respect. Daily means calculated over 9 years indicate that prolonged wintertime snow cover at the grassland site substantially reduced wintertime sensible heat flux (H) compared to the conifer forest in Lavarone ( figure 5(b)). The lower albedo of the forest together with its higher net radiation (figures 5(c) and (f)) led to a higher mean sensible heat flux (H, figure 5(b)). Latent heat flux (LE) was instead not very different between the two land cover types ( figure 5(a)). As a consequence, near-surface air temperature was observed to be higher over the forest by 1.7 ± 0.3°C, even after considering the elevation difference between the two stations by calculating the potential air temperatures (0.13°C ± 0.03). However, it is worth noting that even though the mean air temperature (9 year average) was higher above the forest, the warming rate above the two land use types was similar (data not shown).
Apart from the pronounced carbon sequestration effect of forests that counteracts warming on a larger scale, forests may still affect atmospheric properties and the local climate in other ways. The hypothesis that forests can directly contribute to the formation of clouds (Heiblum et al 2014) and haze is stimulating debates and new studies in the scientific community. The 'biotic pump' concept Gorshkov 2007, Makarieva et al 2013) claims that a physical mechanism leading to reduced atmospheric pressure over forests creates advection of moist air over long distances. The 'bioprecipitation' concept (Morris et al 2014) assumes instead that ice nucleating airborne microorganisms or bacterial/fungal compounds, which are emitted by forests, might have a detectable cloud or haze formation effect. Moreover forests are net emitters of Biogenic Volatile Organic Compounds (BVOCs) (Unger 2014, Maja 2015, which are precursors of secondary organic aerosol (SOA) formation  Table 2. Comparison between the weather data of Trentino and data reported by Philipona (2013) for Switzerland. For both regions, information concerning altitudinal range, number of stations, mean air temperature (T mean ) and rate of change in air temperature are presented separately for high and low elevation.

Region Period
Altitudinal range (m asl)  . BVOC fluxes are proportional to the relative fraction of the land which is covered by forests and to their net emission rates of BVOCs, which have been shown to correlate in turn with forest leaf area (Kulmala et al 2004). In this light the increasing deviation of the incoming radiation from radiation under clear sky conditions (ΔRad) at the three weather stations above 950 m (figure 6) could be a hint to enhanced SOA formation in high elevations, where the forest area had increased most. But facing only a short measurement series (2000)(2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010) and not having direct measurements of BVOC emissions or SOA formation, this interpretation remains speculative.

Conclusions
The data presented in this paper reinforce the view already expressed by Pepin et al (2015) that a positive elevation-dependent warming is not universal in mountain regions. In addition, our findings confirm the opposite hypothesis of e.g. Zeng et al (2015) and Philipona (2013), who attributed a negative elevationdependent warming to solar brightening and increased radiative forcing at lower elevations. Even though a direct relationship between the observed warming trends and increasing forest cover at high elevation could not be identified, this study illustrates the potential of afforestation to compensate or amplify elevation-dependent warming trends by changing the biophysical surface properties of the landscape (roughness, albedo). Eventually, it would be desirable to develop best management practices for forested mountain regions with respect to climate change mitigation strategies, for instance as concerns the forest management form (e.g. single selection system, clear cutting system), the choice of tree species or the stand density. Similar to other EDW studies, the absence of very elevated weather stations with long term records restricted our analyses to altitudes below 2125 m. Covering an altitudinal gradient of 2040 m this study provides a good estimate of the trend for the managed range (forest and pasture), but still we missed roughly 1600 m to the maximum peak height of the region. This limitation calls for better future assessments of meteorological parameters at very high elevations despite the harsh environment and associated methodological problems.