Montane forest expansion at high elevations drives rapid reduction in non-forest area, despite no change in mean forest elevation

Aim: At the elevational limit of forest distribution, montane forests show diverse responses to environmental change with upward shifts, increased tree density and lateral expansion reported. To enable informed analysis of the consequences forest advance will have on montane biodiversity, we quantify changes in the area and elevation of the tree line ecotone and identify how patterns of forest advance are modified by topography and over time. Central Mountain Range, Taiwan. Time 1963–2016. Montane Forests. Methods: Changes in the area and elevation of montane forest at the tree line ecotone were quantified using a stratified random sample of aerial photography captured in 1963, 1980, 2001 and 2016. Weighted estimates of habitat area and elevation for


| INTRODUC TI ON
Mountain systems are experiencing higher than average temperature increases (Dirnböck, Essl, & Rabitsch, 2011;IPCC, 2013;Pepin et al., 2015) and high rates of land use change (Haddaway, Styles, & Pullin, 2014;MacDonald et al., 2000). In response to increasing temperatures and land use changes, uphill advances in the position of montane forests have been observed at the elevational limit of forest distribution where montane forest transitions into alpine vegetation (henceforth, referred to as the tree line ecotone) (Améztegui, Brotons, & Coll, 2010;Améztegui, Coll, Brotons, & Ninot, 2016;Harsch, Hulme, McGlone, & Duncan, 2009). Changes in the position and structure of tree line ecotones are expected to impact montane biodiversity . The relative isolation of mountain environments and high habitat heterogeneity means that disproportionately high numbers of endemic and rare species are found at high elevations (Steinbauer et al., 2016;Vetaas & Grytnes, 2002). Consequently, shifts in the position or structure of the tree line ecotone pose a risk to alpine biodiversity through the extirpation of alpine species and contraction of species ranges as lower distribution limits are pushed towards mountain tops (Jump, Huang, & Chou, 2012).
Shifts in montane forest distribution have primarily been reported as changes in the elevation of the upper ecotone limit. Harsch et al. (2009) reported that 52% of studies published between 1900 and 2008 identified advances in the elevational or latitudinal limit of the tree line. Changes in tree density or growth below the altitudinal limit were not defined as advances in the study by Harsch et al. (2009), yet a growing body of evidence reports substantial increases in tree density or range expansion through across-slope movement of montane forests below the elevational limit in response to environmental change (Bharti, Adhikari, & Rawat, 2012;Feuillet et al., 2020;Greenwood, Chen, Chen, & Jump, 2014;Klasner & Fagre, 2002;Liang, Wang, Eckstein, & Luo, 2011;Mathisen, Mikheeva, Tutubalina, Aune, & Hofgaard, 2014). Changes in the tree line elevation only provide partial information on the response of tree line ecotones to environmental change (Feuillet et al., 2020).
Consequently, assessments of change focusing on the assessment of tree line elevation alone potentially provide misleading estimates of forest range shifts and are inadequate to inform how these changes will impact montane biodiversity (Rahbek et al., 2019).
High variability in the response of tree line ecotones to regional warming occurs due to the influence that complex montane topography has on modifying regional climate patterns, resulting in numerous different climate types within close proximity (Rahbek et al., 2019). While temperature has been identified as a primary controlling variable affecting forest position and advance at naturally occurring mountain tree line ecotones (Körner & Paulsen, 2004), topography modifies local climate regimes causing some slopes to experience climatic conditions that may be cooler, drier or more sheltered than neighbouring areas (Malanson et al., 2011;Suggitt et al., 2011). At the local scale, variation in resource availability (e.g. McNown & Sullivan, 2013;Sullivan, Ellison, McNown, Brownlee, & Sveinbjörnsson, 2015), slope morphometry and lithology (Feuillet et al., 2020), radiative stress (Bader, Van Geloof, & Rietkerk, 2007) and drought stress (e.g. Johnson & Smith, 2007;Leuschner & Schulte, 1991;Millar, Westfall, & Delany, 2007) modifies establishment and growth patterns of advancing montane forests. In addition, the structure of a forest stand itself can also act as a feedback mechanism to facilitate or constrain patterns of tree establishment, growth and mortality through increased seed availability, modification of the microclimate and alterations to competitive dynamics (Camarero et al., 2017). Consequently, while regional changes in temperature may exceed thresholds that would be conducive for shifts in the position of the tree line ecotone, controls acting at the local scale can result in tree line ecotones that are structurally diverse and display high heterogeneity in their elevational position and patterns of advance both between and within mountain ranges.
While numerous studies identify controls on the advance of tree line ecotones, there is a fundamental need for standardized and robust approaches that enable the quantification of change in both the elevation and area of tree line ecotones (Feuillet et al., 2020;Hagedorn, Gavazov, & Alexander, 2019). The best estimates of change and the associated implications typically come from long-term repeat surveys of fixed monitoring sites that are distributed across mountain ranges (e.g. Global Observation Research Initiative in Alpine Environments (GLORIA) which focus on alpine flora; Grabherr, Gottfried, & Pauli, 2000). However, repeat survey data of tree line ecotones are scarce and so the majority of studies are based on incidental historical records and a very limited number of field observations (Gottfried et al., 2012;Steinbauer et al., 2018) which may not provide a reliable picture of change over large areas and at worst may lead to bias in forest range estimates (Fisher, Hurtt, Thomas, & Chambers, 2008).
While field observations have steered much of our current understanding of species range shifts and the implications to montane biodiversity, the use of remote sensing data to assess changes in forest distribution is attractive to overcome limitations imposed by poor accessibility to field sites and the lack of historic field data in mountain ranges. Repeat aerial photography data have been used to identify changes in the maximum elevation or tree density of the tree line ecotone (e.g. Feuillet et al., 2020;Klasner & Fagre, 2002;Luo & Dai, 2013;Mathisen et al., 2014;Resler, Fonstad, & Butler, 2004). However, variation in the methods used to analyse aerial photography data alongside a lack of quantitative estimates of uncertainty in the range shifts reported limits the interpretation of the results, and hinders landscape-scale estimation of changes in forest K E Y W O R D S climate change, densification, forest change, migration, mountain, range edge, Taiwan, tree line distribution in mountain ecosystems (Morley, Donoghue, Chen, & Jump, 2018).
On average, elevational changes in species range shifts lag behind upward shifts in isotherms in mountain regions (Rumpf et al., 2018).
However, the rate of biodiversity change in high-elevation areas seems to be accelerating as the rate of temperature increase accelerates, which will likely lead to tipping points in species loss being met sooner with ongoing climate warming (Steinbauer et al., 2018). Therefore, there is a pressing need to quantify variation in patterns of forest advance over time in order to avoid over-or under-stating the extent of tree line ecotone change and subsequent impacts to montane biodiversity. Here, we assess change in the elevation and area of montane forest at the tree line ecotone using repeat aerial photography data, and quantify the rate of forest advance, identifying how patterns of advance vary according to topography and growth stage.  , Greenwood, Chen, Chen, and Jump (2015) found that the high-elevation tree line ecotone in Taiwan is predominantly temperature limited, with topography and local sheltering influencing the position, structure and advance of the ecotone through a modification in regional temperature regimes. The importance of local topographic controls in the Central Mountain Range has resulted in a highly reticulate and structurally diverse tree line ecotone. As a consequence, patterns of forest advance within the study area show a high degree of variation over a short distance.

| Study area
Localized reductions in the elevational position of the tree line ecotone are caused by sporadic, naturally occurring small-scale fires and landslides. However, routine anthropogenic disturbance or grazing by either domestic or wild herds is considered of low impact. the mean error in all image pairs was < 1 pixel).

| Change assessment
Probability-based change estimates are a long-established technique for assessing changes in habitat area and condition that have been widely adopted by forest monitoring programmes interested in quantifying change in forest area and forest degradation (Cochran, 1977;Olofsson, Foody, Stehman, & Woodcock, 2013;Pickering et al., 2019). This approach to change assessment uses manual interpretation of remote sensing data at sample plots to estimate the area of each habitat type at each survey date and within a given terrain feature or geographic region of interest (henceforth, stratum) alongside an uncertainty value for the area estimate.
Probability-based change estimates are recognized as a reliable method for estimating changes in habitat type or condition enabling F I G U R E 1 Aerial photography from 2016 showing the tree line of high elevation conifer forests in the Mt Hehuan study area of the Central Mountain Range, Taiwan (the area above 2,400 m a. s. l. is shown in black in the inset and the study area marked by the red outline) changes in habitat to be identified over time with a high degree of confidence (Olofsson et al., 2013;Olofsson, Holden, Bullock, & Woodcock, 2016;Stehman, 2013).

| Sample design
A proportional stratified random sampling design was used to assess change in forest distribution at the tree line ecotone. To ensure adequate representation of the entire study area, slope orientation was used as a basis for stratification due to the major influence of topography on the patterns of forest advance in Taiwan's Central Mountain Range (Greenwood et al., , 2015. Stratification was based on 12 categories of slope aspect and incline attributes calculated from a high-resolution TanDEM-X Digital Elevation Model (12 m spatial resolution resampled to 15 m pixel size), using four cardinal compass directions (±45° in either direction) and three inclination classes (0-20°, 21-45° and > 46°). The number of samples taken in each stratum was proportional to the area of the study region occupied by the aspect-incline combination (Table S1.1, Figure S1.1).
Following the removal of sample plots that had to be omitted due to a cloud or shadow impairing visual interpretation, a total of 2,785 sample plots were interpreted, equivalent to 1.54% of the total study area.

| Change attributes
At each of the 2,785 sample locations, a sample plot measuring 15 × 15 m in plan view was created and interpreted manually for each year of change analysis (1963, 1980, 2001 and 2016). Each sample plot was assigned one of four vegetation classes for each year in the change survey (Table 1)  with at least 10% canopy cover and trees greater than 5 m in height are classified here as forest. Areas with small trees present within the plot that do not meet the thresholds of a forest as set out by the FAO definition were categorized as establishing forest. The scale of the aerial photography (≤0.5 m pixel size) is sufficient to discriminate differences in tree size based on crown size. Areas with partial removal of the forest canopy between time periods are categorized as disturbed and treeless areas are categorized here as non-forest areas.
The distinction between the forest and establishing forest classes is important. Forest resource assessments rarely comment on areas of forest establishment that do not meet the pre-defined criteria for forest cover. However, ecological and biogeographic studies have a much broader interpretation, with the timberline defined as the upper limit of closed forest, the tree line defined broadly as the line connecting the maximum elevation of trees greater than 2 m in height and the species limit by the uppermost trees irrespective of tree height (Körner, 1998). Therefore, in this study we make a separation between mature montane forest (forest class) and areas of the tree line ecotone above the continuous forest limit that are undergoing increases in tree size or density (establishing forest class). In doing so, we are able to gain a more detailed understanding of the forest-grassland transition that is present at mountain tree lines than a simpler forest/ non-forest vegetation classification. Forest An area of trees that meet FAO (2018) criteria of a forest with at least 10% canopy cover and trees greater than 5 m in height.

| Change estimates
Establishing forest An area of forest establishment, small trees are identifiable in aerial photographs due to their small crown size.
Non-forest An area that lacks trees.

Disturbed
An area of forest with a reduction in canopy cover but some trees remain.

Omitted
Unable to identify vegetation class due to cloud cover or shadow. All uncertainty measures reported are at the 95% confidence intervals unless otherwise stated and area estimates are reported in plan view.

| Landscape-scale change estimates
In the Mt. Hehuan region, c. 20.6% ± 2.3% of the non-forest area in 1963 was establishing forest in 2016 with a further 8.2% ± 1.5% of the non-forest area in 1963 identified as forest by 2016 (Table 2).
Forest disturbance in the Mt. Hehuan region is rare, 1.4% ± 0.4% of the forest area in 1963 became non-forest by 2016 while a further 0.7% ± 0.3% experienced a reduction in canopy cover between 1963 and 2016 (Table 2). There was no evidence indicating anthropogenic causes for forest advance or loss in the Mt. Hehaun region.
In areas of forest loss, complete removal of substrate was visible in the aerial photography suggesting that forest loss is primarily caused by landslide events with no direct evidence in the aerial photography to suggest fire or forest management caused a loss in forest area.

| D ISCUSS I ON
Repeat aerial photographic survey data reveal that forest advance has led to a loss of 29% of the non-forest area present in 1963 in the Mt. Hehuan region of the Central Mountain Range, Taiwan (Table 2).
While there has been an increase in forest area of 295.0 ha between 1963 and 2016 in the Mt. Hehuan study area, the mean elevation of the forest class has not changed over the study period. The

F I G U R E 5
Temporal variation in the rate of forest advance in the Mt. Hehuan study area (Taiwan). The rate of recent forest establishment, defined as the change from non-forest to establishing forest within a single change period (1963-1980, 1980-2001 and 2001-2016) is shown for the study area as a whole (a), by aspect (c) and by incline (e). The rate of advanced establishment, defined as the change from establishing forest to forest within a change period is shown for the study area as a whole (b), by aspect (d) and by incline (f). Uncertainty in the estimates in panels a and b is shown at the 95% confidence intervals substantial increase in forest area, yet stasis in mean forest elevation reported here, (Figure 3) has also been observed globally including in the Khibiny Mountains, Russia (Mathisen et al., 2014), the Tibetan Plateau (Liang et al., 2011), Glacier National Park, USA (Klasner & Fagre, 2002), the Sudetes mountains, Czech Republic (Treml & Chuman, 2015) and the Pyrenees, France (Feuillet et al., 2020). In Taiwan, our findings are consistent with patterns of forest advance previously identified in field assessments by  who identified that forest advance in Taiwan The process of infilling occurs because localized depressions in elevation of the ecotone enable tree recruitment to occur both from below and across slope from the edges of these depressions (Treml & Chuman, 2015). The montane forests of the Central Mountain Range have a highly reticulate tree line owing to strong topographic and micro-climatic controls on tree line position and seedling establishment (Greenwood et al., , 2015. The reticulated nature of the tree line means that the forest has reached its maximum potential elevation in many areas, as determined by mountain ridge tops. As a result, it is not possible for the forest to shift further upslope but the and elsewhere are occurring despite a lack of change in tree line position (Klasner & Fagre, 2002;Liang et al., 2011;Mathisen et al., 2014;Treml & Chuman, 2015). Consequently, it is increasingly important that multiple responses of the tree line ecotone are considered to ensure studies avoid potentially erroneous conclusions of the impact of environmental change on high-altitude ecosystems.
The increases in forest area and mean elevation of establishing forest reported for the study area as a whole mask important variation in patterns of forest advance within the region. East-and south-facing aspects show larger increases in forest area between 1963 and 2016 than west-and north-facing slopes (Figure 4a).
Similarly, forest advance between 1963 and 2016 is greatest on slopes with inclinations between 0 and 45°, with the greatest decline in non-forest area occuring on 0 and 20° slopes (Figure 4b).
This substantial decline in non-forest area is driven by a 6% increase in the area of 0-20° slopes that are occupied by establishing forest. Greenwood et al. ( , 2015 show that seedling establishment in the Central Mountain Range, Taiwan, is strongly controlled by topography, with higher recruitment associated with east-facing or shallow slopes and more sheltered areas. Furthermore, as temperature thresholds are passed at a given elevation, a larger area of habitat is likely to be affected by environmental change when the slope inclination is shallow (Jump, Mátyás, & Peñuelas, 2009). Therefore, in combination with an increased probability of forest establishment on shallow and concave slopes (Feuillet et al., 2020;Greenwood et al., , 2015, environmental changes lead to higher rates of forest establishment and rapid declines in grassland area once shallow slopes become favourable for seedling establishment.  (Greenwood et al., 2015). Consequently, areas that undergo establishment but do not continue to grow sufficiently to be classified as forest may exist in areas where a threshold for establishment has been surpassed, but the necessary conditions for subsequent growth are not met.
This pattern of growth restriction occurs in 21.3% ± 5.3% of the areas identified as establishing in 1963, where the trees failed to grow sufficiently to move into the forest class by 2016 (Table 2).
There is no evidence to suggest that the areas of slow growth exhibit a stunted growth form such as the krummholz tree lines commonly found in other areas . Understanding the mechanisms that lead to variable rates of establishment and growth should be a priority for new research. The rates of change reported here do not match the accelerating rate of temperature change reported (Jump et al., 2012), and so detailed time series data regarding range shift lags and the interactions with biotic and abiotic factors controlling tree establishment and growth is required to enable improved forecasts of forest growth and range expansion at the tree line ecotone.
The ability to identify heterogeneous patterns in the location, structure and rate of ecotone advance, as shown here, is vital to enable the discrimination of areas that are likely to experience rapid changes in habitat structure from those where stasis might be expected under future climate scenarios (Hagedorn et al., 2019). Forest advance is expected to impact biodiversity of the alpine zone, yet local variability in shifts in the tree line ecotone will modify landscape-scale impacts . Forest advance in Mt. Hehuan has reduced the available area of non-forest habitats by 29% and caused a 20-32 m increase in the mean elevation of remaining non-forest habitats through range contraction (Figure 3; elevation gain increases to 32 m when landslide events are excluded).
This reduction in habitat area is a particular problem for alpine flora given the typically high rates of endemism and accelerating rate of warming-induced species richness increase in alpine areas (Jump et al., 2012;Steinbauer et al., 2018). Jump et al. (2012) show that the elevation of mountain plant species distribution has increased, on average, by 3.6 m/yr in Taiwan during the last century, indicating that some forb and shrub species are advancing uphill more rapidly than the montane forest. Furthermore, variation in the rate of forest advance caused by slope morphology and substrate (Figures 4-6;Feuillet et al., 2020;Greenwood et al., , 2015 may enable alpine flora to persist despite tree line shifts and infilling occurring in other areas (Bruun & Moen, 2003). Therefore, it is likely that the presence of refugia in areas of slow change or growth limitation will play an increasingly vital role in the maintenance of alpine biodiversity in mountain systems as forest advance continues. However, even where such refugial areas occur, contraction in population size of alpine species is likely due to a reduction in the non-forest areas, risking population loss and diminishing but not removing their risk of local extinction.
The quantification of forest advance and its variability in mountain regions is vital to improve our understanding of tree line ecotone response to environmental change and the mechanisms that drive the observed variation in response. Such knowledge is of fundamental importance to enable the impacts of tree line ecotone change on montane ecosystem function and biodiversity to be quantified. While the need for standardized methods to quantify variation in ecotone response is recognized, little progress has been made (Feuillet et al., 2020;Malanson et al., 2011). Here, we demonstrate an important methodological application of change assessment using probability-based sampling of repeat aerial photography to enable a precise and unbiased quantification of montane forest range shifts. Aerial photography data with 0.25-0.5 m pixel size (typical of many aerial surveys) are of sufficiently high quality to identify tree recruitment at the tree line ecotone and have been used to enable research in areas where historical field datasets are lacking (e.g. Feuillet et al., 2020;Mathisen et al., 2014).
As such, aerial photography offers an important resource to study contemporary changes in tree recruitment and forest range shifts that complements existing field-based assessment of change and may enhance assessments of longer-term historical patterns that use dendroecology techniques (e.g. Camarero & Gutiérrez, 2004;Liang et al., 2011). Despite locally excellent data availability, repeat aerial photography is not available in many mountain ranges of the world.
However, rapid advances in the spatial resolution of satellite-borne Earth observation data (0.3-0.6 m pixel size) offer significant opportunities to progress towards standardized methods for assessment of forest change and understanding tree line ecotone dynamics (Bader & Ruijten, 2008;Bolton, Coops, Hermosilla, Wulder, & White, 2018;Carlson et al., 2017;Weiss, Malanson, & Walsh, 2015). Integrating the methods that we demonstrate here with rapidly improving Earth observation data for the mountain regions of the world provides a significant opportunity to understand the drivers of mountain forest change, and will enable a robust assessment of the impacts forest advance on biodiversity and ecosystem function from local to global scales. In doing so, we identify that east-and south-facing slopes alongside shallow slopes have experienced the largest declines in non-forest habitat, and 0-20° slopes are at high risk of ongoing loss in habitat area due to an increase in the area of forest establishment. The precise quantification of changes in montane forest elevation and area shown here improves our understanding of the drivers of variation in forest response to environmental change enabling regional-scale assessment of tree line change and facilitating prediction of future forest range shifts and the impacts of forest range shifts on biodiversity and ecosystem function in mountain systems. stir.ac.uk/. The authors are unable to release the original aerial photography or TanDEM-X data used in this study due to licence agreements that restrict the distribution of the data. Aerial Photographs used in this study are subject to access and usage restrictions on the authors. Please contact J-C. Chen (zzzjohn@mail.npust.edu.tw) and

ACK N OWLED G EM ENTS
A.S. Jump (a.s.jump@stir.ac.uk) to discuss access in the first instance.