PorTree


	Home	Introduction	Methods	Reporting	Downloads	Presentations

Methods

Multiple sources of climate and tree / forest data were required to solve the PORTREE questions at the scale of Switzerland and the European Alps in order to guarantee that enough gradient is sampled for each tree species. These data sets needed to include presence/absence (P/A), growth and stand composition, as well as mortality over the last 10-20 years. Finally, for all sites, spatially explicit climate data are required, ideally as daily, or at least as monthly and seasonal/annual maps of important climate variables including their projections under a range of IPCC scenarios.

Tree and inventory data

We used the following tree/stand data sets (in collaboration with EU-projects, where we were key partners):

For PORTREE models at the scale of the European Alps, national forest inventory data (A, CH, D, F, I, SLO) using presence/absence (P/A) of tree species per plot were compiled. Range statistics on the Swiss scale were calculated within Switzerland from these PORTREE models (See appendix S3).

For models at the European scale: ICP-Forest Level 1 data: data including P/A and "composition" of species per all ~6500 plots. These data were used to calculate the range statistics at the Europe scale.

For growth models, we could only use Swiss data (LFI). We used (a) stand basal area, (b) period length (in ¼ years) between inventory periods per plot, and (c) individual tree DBH as predictors for the growth model calibration in addition to the climate variables listed below. We used individual tree radial DBH increments as dependent variable in the growth models.

Current climate, climate downscaling and topography data

We used three sources for our climate and topography data, namely Worldclim, PRISM precipitation for the Alps, and an SRTM (90m) derived digital elevation model across the Alps resampled to a spatial resolution of 100m.

For topography analyses, we assembled SRTM-based elevation data (Rabus et al. 2003), originally available at 90m spatial resolution, and resampled it to 100m resolution for the European Alps.

Worldclim data (Hijmans et al. 2005) is globally available for monthly mean temperature and precipitation (mean period of 1950-2000) at a ca. 1km spatial resolution. It is an excellent data source for continental scale, contemporary climate data. For the scale of the Alps and the PORTREE model calibration, we used the monthly temperature maps (Tmin, Tave, Tmax,). Precipitation maps were not used due to the low number of precipitation stations used in Worldclim, which resulted in a lack of representing dry interior valleys.

PRISM-derived (Daly et al. 1994) precipitation maps (Prcp) were used for the Alps. This product has been developed at ETHZ for the European Alps, using ca. 6000 rainfall gauging stations (Schwarb et al. 2001). While first only available as annual and seasonal maps, it now contains monthly mean precipitation maps for the Alps and was made available by C. Frei, MeteoSchweiz. The PRISM method is also - in a modified version - the new core of the climate interpolation and mapping method at MeteoSchweiz. The data is available at a 2 km spatial resolution.

Before extracting predictor variables, we downscaled all monthly climate data to the 100m spatial resolution of the topography data, using moving window regression approaches (Zimmermann et al. 2007). The 100m data was then processed to predictor variables and values were extracted for each inventory point location.

Future climate projections and downscaling of projections

Before processing the current climate and topography data to predictors, we processed and downscaled climate change projections to the same spatial resolution of 100m. The following data were used:

RCM climate data from the EU project ENSEMBLES . We primarily were interested in the basic climatology output variables Tmin, Tave, Tmax, Prcp. We used the A1B scenario, since this scenario was available for all models. Other scenarios were only available occasionally. We used data from 6 RCM/GCM model combinations (see table 1) for our analyses.

We downscaled AR4 IPCC scenario simulations for forecasts in the Swiss study areas. IPCC scenarios based on AR4 were not yet easily available in a downscaled form (<=1km resolution), and primarily were only available at very coarse spatial resolution (ca. 20km pixels). The latter was too coarse to be easily downscaled statistically. We thus used RCM output from the ENSEMBLES project, which has made the four basic IPCC scenarios (A1B, A2, B1, B2) available. However, only the A1B scenario was consistently available trough all RCMs, therefore we decided to use this scenario for the "Wald & Klimawandel Program". The other scenarios were downscaled where available.

Predictor variables for model calibration

After downscaling current and potential future climates to a 100m spatial resolution, we derived a series of predictor variables for model calibration. Specifically, we generated the following predictors:

(1) degreeday sum with a 5.56°C threshold, based on the monthly Tave maps (see Zimmermann & Kienast 1999); (2) temperature seasonality (standard deviation of monthly values, according to Worldclim routine), (3) summer precipitation (sum of April to September monthly values), (4) winter precipitation (October to March), (5) potential yearly global radiation (see Zimmermann et al. 2007), (6) slope angle in degree, (7) topographic position (difference between the average elevation in a circular moving window applied to a 100m digital elevation model and the center cell of the window (Zimmermann et al. 2007), (8) aspect value (ranging from 0, i.e. south, to 100, i.e. north; Meier et al. 2014), and (9) distance to running waters (from an European coverage of running waters). These parameters were then used to calibrate predictive models of the spatial distribution of PORTREE species. For some species, we fitted models only for Swiss inventory points, mostly for two reasons: (i) the species was not distinguished well in European inventories (e.g. oaks, maples, ashes), and therefore no sound data basis was available to calibrate models from the whole Alps (see Appendix S3); (ii) the models using the above mentioned predictors did not reveal sufficiently sound model results. For models calibrated for Switzerland only, we added (10) soil depth from the soil suitability map of Switzerland (Eidg. Justiz- und Polizeidepartement (EJDP) 1980), (11) calcareous/non-calcareous bedrock (Allenbach et al. 2008), and (12) distance to running and standing water (thus replacing the less precise layer (9).

Statistical SDM analyses and map projections

We analyzed the distribution of tree species (P/A; = persistence area) along environmental gradients using: [1] Classification and regression trees (CART), [2] Flexible discriminant analysis (FDA), [3] Generalized linear models (GLM), [4] Generalized additive models (GAM), [5] Artificial neural networks (ANN), and [6] Generalized boosted regression trees (GBM) in order to build model ensembles. The selected climate and other environmental variables are (a) known to be relevant to tree species in general, (b) have partly been shown to be of physiological relevance, and (c) had all a Pearson correlation coefficient <0.7 among each other to avoid collinearity problems when fitting the models.

We thus had 6 climate models (for future projections) and six statistical models fitted for each tree species. This resulted in six maps per species under current climate. For future conditions, this resulted in 36 maps per species and time step. Projections were generated at 1km spatial resolution, but can also be generated at higher spatial resolution.

Each map first represents probabilities to find a specific tree species at each landscape pixel. These probabilities are then classified into binary P/A maps for integrating into ensemble maps. Probabilities were converted into binary maps by optimizing the probability cut-level so that the AUC values in a crossvalidation test exercise per species and statistical model was optimized. It thus represents the most accurate reclassification of probabilities into simulated presence/absence values for each species and model.

These maps all represent likely futures of a tree species, and we used all maps to construct ensemble projections that fulfilled the following criteria: First, we excluded models of inferior quality (AUC < 0.8), second we mapped per pixel how many of the 6 SMDs for current, and 36 SDMs for future conditions mapped "presence" of the species. Finally, we reclassified these summed presences and absences so that we generated 3 zones. Namely: (1) a zone were less than 30% of the models project "presence", meaning that there is high agreement of "absence", (2) a zone where more than 60% of all model combinations project "presence", meaning that there is comparably high agreement among models, and finally (3) a zone where 30-60% of the models project "presence", meaning that there is high uncertainty among statistical/climate models as to whether the species is present or absent at these pixels. All models were evaluated based on Kappa, AUC and TSS.

Statistical growth analyses and growth projections

In order to analyze the growth potential along environmental gradients, we fitted the (radial increment or DBH) growth of individual trees as a GLM function with the following variables as predictors: (1) DBH of the dependent tree individual, (2) stand basal area (stand BAI) of the surrounding LFI plot, (3) the period length between the two inventory periods used to assess individual tree radial growth (in ¼ years), and (4) the basic nine climate variables used for the SDM model calibration. By this, the individual tree growth model is sensitive to the size of the tree, the stand density (BAI) and the length of the inventory period, in addition to all climate parameters used.

The growth model was then plotted on top of projected SDM persistence domains (see Figure 12 & 13) by keeping stand and tree size of each plot at a median value among all individuals per species, by setting the inventory period length to 10 years, and by then plotting the modelled growth onto all LFI points with the respective climate values. In a postprocessing, we mapped isolines of maximum simulated growth per plot region, thus representing potential growth rates of a species draped over the persistence niche of the same species. The two niches (persistence & growth) do not fully match on top of each other because the persistence niche is fitted from the PORTREE inventory points across the Alps, while the growth niche is fitted only from LFI data.

References Allenbach K, Maggini R, Lehmann A, (2008). Swiss Environmental Domains: Technical report. FOEN Report, Bern, 55p. Hijmans RJ, Cameron SE, Parra JL, Jones PG, Jarvis A (2005) Very high resolution interpolated climate surfaces for global land areas. International Journal of Climatology 25, 1965-1978. Meier ES, Dullinger S, Zimmermann NE, Baumgartner D, Gattringer A & Hülber K, (2014) Space matters when defining effective management for invasive plants. Diversity and Distribution 20: 1029-1043. doi: 10.1111/ddi.12201. Rabus, B, Eineder M, Roth A, and Bamler R (2003). The shuttle radar topography mission (SRTM) - A new class of digital elevation models acquired by spaceborne radar. ISPRS Journal of Photogrammetry and Remote Sensing 57, 241-262. Schwarb M, Frei C, Daly C, Schär C (2001) Mean seasonal precipitation throughout the European Alps 1971-1990. In: Spreafico M, Weingartner R, Leibundgut C (eds) Hydrological atlas of Switzerland. Landeshydrologie und geologie, Bern, Plate 27 Zimmermann NE, Edwards TC, Moisen GG, Frescino TS, Blackard JA (2007) Remote sensing-based predictors improve distribution models of rare, early successional and broadleaf tree species in Utah. Journal of Applied Ecology 44, 1057-1067. Zimmermann NE & Kienast F, (1999) Predictive mapping of alpine grasslands in Switzerland: species versus community approach. Journal of Vegetation Science 10(4): 469-482.