The current distribution data for 246 taxa were obtained by georeferencing the Atlas of Florae Europaeae . We gathered additional occurrence data for the Mediterranean flora from the Global Biodiversity Information Facility data portal (http://www.gbif.org/). These data were checked and corrected by removing species' presences from botanical and herbaria collections and/or observations with lower spatial resolution than 30′ (∼ 55 km). The final taxa list and the assignment of pollen taxa to and modern taxa distributions is given in Table S1.

The georeferenced geographical distributions were rescaled to the resolution of 30′ (∼ 55 km). The global observed climate data (1950–2000) for January minimum temperature (Tjan), July maximum temperature (Tjul) and annual precipitation (Pann) data were obtained from WorldClim database www.worldclim.org with 5′ resolution (∼ 10 km). The climate data were downscaled to the same spatial resolution of the plant distribution data by aggregating the mean value to the resolution of 30′. All computing was performed using R with the package rgdal .

The climate reconstruction method is based on the PDF of each taxon identified in a fossil dated pollen assemblage. Pollen taxa were assigned to georeferenced plant taxa (see Table S1). This approach was successfully used to reconstruct climate variables from fossil pollen data . Using the PDFs intersection of all taxa identified in a fossil sample we obtain the most likely climate value within which the fossil plant assemblage may occur . It has been observed that normal and log-normal (right skewed) distributions fitted to temperature and precipitation, respectively, tend to better represent the data . To avoid sampling the climate spatial distribution instead of the species tolerance, we corrected for the potential bias by using binned climate within the species range as a weighting factor for each climate value . The chosen bin size is 2 ∘C for temperature variables and 20 mm for precipitation. This procedure decreases the weight of the most frequent climate values and increases those, in the distribution of the species, that occur less frequently in the study area .

The reconstructed climate using the PDF method results from combining the individual PDFs of the species identified in each pollen sample. The product of the PDFs provides the most likely climate value . To identify a taxon as present in the sample, a threshold of three pollen grains was chosen. A minimum of five taxa present is required to reconstruct a climate value for each fossil sample.

Using presence data is seen as an advantage of the PDF method as well as a weakness due to the exclusion of the quantitative data resulting from the pollen abundances . Fluctuations in pollen abundances are related to multiple factors such as the species ecophysiology, differential pollen production, dispersal capacity, and other traits . We have used the pollen proportions to weight the PDFs of the respective taxa. The minimum positive pollen proportion corresponds to the presence of the taxon, while the maximum defines its highest abundance within the fossil record. Using pollen proportions of a taxon within a time series instead of within a sample avoids the bias of differential pollen production and thus allows estimating the presence of a species relative to its maximum percentages in the whole record.

The pollen proportions were converted to alpha values, reducing the species climate tolerance towards the peak density values (Fig. ). We assumed that the pollen proportion has an inverse relation to the proximity of near-optimal conditions. To avoid the selection of a unique climate value from the PDF when the maximum detection of a species occurs – i.e. when its pollen proportion is found to be 1, we use a pollen adjustment value set to 0.9. This means that, at the maximum taxon detection, the PDF will be reduced to the area of the density corresponding to 10 % of the probability (Fig. ). The maximum detection of a taxon indicates a near-optimal climate niche and the adjustment value set to a value near but not equal to one allows some degree of uncertainty in the reconstruction. On the other hand, setting this value to zero will not allow any influence of the pollen proportion, resulting in a binary presence/absence reconstruction . For each sample, the collection of the taxa tolerance intervals built this way are added resulting in a taxon profile, showing where in the climate space the frequency of the taxon is higher, taking into account the proximity to optimal conditions. The final climate reconstruction value is the product of the climate PDF with the taxa profile. The reconstructed value and associated uncertainty are usually extracted from the PDF as the mean and standard deviation . Assuming a normal distribution, we extract the peak density value and the 95 % confidence interval from the density profile. The confidence interval range shows the uncertainty around the reconstructed value and is related to the standard deviation. The reconstructed values for each site were fitted with a smoothing spline to produce continuous time series, from which 1000-year time slices were extracted.

In order to evaluate the robustness of the reconstruction method, we have compared modern reconstructed and observed climate data (1950–2000) from WorldClim database. For reconstructing climate from pollen data, we have used all samples available within the last 500 years. Climate values were averaged for all sites with more than one sample. The correlation between the two data sets was tested using a Pearson's correlation score. To provide the significance of the correlation value, a set of 999 replicates were performed where the observed climate variable was shuffled without repetition. Although this evaluation does not take into account neither the climate oscillations during the last 500 years nor the human disturbances, it still provides a broad evaluation of the reconstruction method because (1) it depicts per site the relationship between observed climate data with reconstructed values and (2) the slope direction of the regression and the related correlation signal indicate that the reconstruction is spatially coherent. A linear regression was used to estimate a baseline for calculating the anomalies at each site using the observed climate. The pre-industrial period around AD 1850 is commonly used as reference climatology to compute anomalies. This period is also often used as a baseline in climate models, facilitating data–model comparisons, and it is less biased with recent climate warming allowing past warming to be better depicted . Although a specific year is selected, the time frame often includes ±500 years , which is equivalent to the period we have used in our study. The regression allows a climate baseline to be built without artificially adding samples to compensate for differential number of samples available for recent periods in a 4-D (spatial plus time) interpolation and the linear equations provide all the information to generate the baseline with the observed climate data.

Thirteen climate grids, ranging from 15 000 to 3 000 calendar years BP (hereafter, “ka”) with a 1000-year interval, were obtained for each reconstructed variable by spatial interpolation of the climate anomalies at each available site. The anomalies were computed for each site between the reconstructed climate and the modern reference climate. Anomalies were projected onto a 30′ (∼ 55 km) resolution grid and interpolated onto a 5′ (∼ 10 km) resolution grid using 3-D thin-plate smoothing splines with two spatial dimensions including altitude. This interpolation method generates accurate climate predictions and it was used for the WorldClim variables .

To further summarise the spatial and temporal variability of the data we applied a functional principal component analysis (fPCA). The fPCA extends the exploratory data analysis of the principal component analysis to functional data , depicting both spatial and time patterns that are then summarised in a few components. applied a fPCA to nearly the same timescale as the present study to depict January temperature patterns from European pollen data. Here we have broadened the approach to each climate time series available in each grid cell to produce gridded spatial components. The functional data were built by combining B-spline basis functions to fit the time series. We have retained the components that explain more than 90 % of the variance and rescaled the range from -1 to 1. We used hierarchical cluster analysis over the produced first components grids of each variable to identify areas in the Iberian Peninsula that share similar climate trends over the past 15 ka. Climate stability was computed for each variable as the mean absolute deviance from the current climate as available in WorldClim data set.

All analysis were performed using the R Project for Statistical Computing with packages fields , rgdal , gstat , and fda . The climate reconstructions were performed with R scripts developed by the authors and available at request.

The 1000-year time interval provides enough resolution to analyse general patterns of climate evolution. However, abrupt climate events are not detectable. The end of the Oldest Dryas (OD; ending around 14.5 ka) is characterised in Iberia by a vegetation change compatible with cold and humid conditions and is followed by the Bölling–Allerød warm period (B-A; ending around 13 ka). Our results show a similar pattern with colder conditions between 15 ka and 13 ka and a higher humidity (Fig. ), particularly evident in the central and southern clusters. Although all clusters show a similar trend, the C1 and C2 are colder than average. The general pattern in the Iberian Peninsula is a contrast between a colder north and a warmer south but, nevertheless, an area dominated by low January temperatures (Fig. , Fig. S3, S5). The evolution of precipitation during the last 15 kyr in the Iberian Peninsula shows a very stable pattern: northern areas comprised in C1 had high precipitation values during the period analysed, while the south was wetter than today (Fig. S3, S5). The increase in the moisture availability during the B-A is in line with the slight increase in precipitation in all clusters between 14 and 13 ka (Fig. ).

As described earlier in Europe , Tjan shows wider changes in amplitude than Tjul. The cold to warm transitions that occurred at ∼ 14.7 and 11.5 ka in Europe had a spatial impact that is noticeable in the reconstructed temperatures (Fig. , Fig. S3, S5).

The B-A warm stage is followed by the cold Younger Dryas (YD; between ∼ 12.9 and ∼ 11.7 ka), marking the beginning of the Holocene. This period records a warming trend for Tjan, while Tjan decreased abruptly (Fig. ) with a reduction of the warmer areas between 13 and 12 ka (Fig. ).

The Holocene warm period (approximately between ∼ 8.2 and 5.6 ka, depending on the location in Europe) is characterised by increasing summer temperatures . Such trend is more obvious in northern Europe and the Alps, while we rather observe a cooling at lower latitudes . Our results point to a slight decrease in Tjan and Tjul around 7 ka, but the overall temperature pattern is rather stable. This is likely affected by the temporal resolution of this study, failing to clearly detect rapid events. Pann shows a slightly wetter climate at 7 ka (Fig. ) which is consistent with earlier reconstruction for the southern European lowlands .

Between 6 and 3 ka, areas with low precipitation expand in the Iberian Peninsula (Fig. ), which allows the expansion of the Mediterranean taxa . The increasing aridity trend in the south is balanced by the high precipitation values in the north (Fig. , Fig. S3, S5), contributing to the shaping of the current Iberia pattern of two contrasting bioclimatic regions: the north is temperate and wet while the south is a dry and warm.

The behaviour of the reconstructed variables at 5 ka is likely to be influenced by non-natural ecosystem changes due to human activities such as the forest degradation that began in lowlands and later in mountainous areas . These human impacts add confounding effects in the fossil pollen record and may lead to slightly biased temperature reconstructions after 5 ka. On the other hand, human impacts at larger scales, capable of leaving noticeable imprints on landscape, were likely to happen later and, furthermore, there is evidence of a cooling and drier stage in the Iberian Peninsula after 5 ka .

The climate change since the LGM in the Iberian Peninsula had an impact on the persistence of temperate species, migrating pathways, and on the overall recolonisation processes during the post-glacial period within the peninsula . During this period, climate favoured migrations and expansion processes that culminated in secondary contacts for several lineages previously isolated in patches of suitable habitat . Particularly the B-A warming phase and the warming stage after the YD, which we show here and which have highly affected the spatial organisation of the climate in the Iberian Peninsula, are likely favouring expansion processes of warmth-dependent organisms. Given the relationship between climate change and biodiversity patterns, the clustering scheme (Fig. ) depicting areas with different climate evolution is consistent with the molecular evidence of a network of putative refugia within Iberia . Refugia have been associated with climate and habitat stability, with both playing complementary roles . However, as shown by large-scale landscape analysis and climate reconstructions , both have a strong dynamic nature in the Iberian Peninsula and likely promoted the formation of patches of suitable habitat during harsh conditions. The highly structured populations that many species exhibit in the Iberian Peninsula have contributed decisively to the idea of refugia diversity . Overall, the information included in the multidimensional climate data allowed us to define areas characterised by a climate evolution during the late Quaternary with smaller amplitude of change (clusters C3 and C4). These areas showed higher stability of both Tjan and Pann (Fig. ). Cluster C4 coincides at a great extent with areas that offered more resilience to change between millennia (Fig. ). Within these areas, temperature and precipitation were suitable to support the survival of temperate trees, likely acting as glacial refugia. On the other hand, the cold areas of the first and second cluster also associated with faster changes cluster likely diminished the suitability for the long term persistence of species. One might infer that the defined clusters are associated with potential isolation or dispersal events of species throughout the studied time span. Particularly, the fourth cluster (Fig. ) includes areas that have already been described as glacial refugia for several animal and plant species (; see chapter 5 for a review of refugia in the Iberian Peninsula). In the area represented by this cluster, the reconstructed Tjan indicate a mild climate with higher precipitation than currently, which is compatible with the persistence of species in these areas. The southern plateau, mostly comprised in the second cluster (Fig. ), also recorded mild conditions which are often associated with southern refugia; however, the rapid Tjul oscillations associated with a cold Tjan and low precipitation may have prevented long-term persistence but are likely compatible with a recolonisation process.

The pattern of stability indicates a southern Iberia with less change, particularly Tjan and Pann. High altitudes offer more resilience to change, particularly to July temperature, and lower areas may be swept rapidly with occurring changes (see Fig. S6). Our data suggest that, at the regional scale and with extensive time-series data, this relation is preserved. Areas of lower velocity of change, which are hence more stable, are associated with high levels of endemicity at global scales , and areas of high velocity are often associated with species extinction . Our results indicate higher stability in the southern part of the Peninsula, similar to other studies based on climate data . However, our studied time frame extends to 15 ka, which does not cover the glacial maximum (∼ 21 ka). At that time period, a higher degree of fragmentation of the stability is expected due to colder conditions, and areas compatible with refugia would be also less contiguous. These could be seen as a macrorefugia, offering conditions for large population during glacial times . Microrefugia are known to occur in the northern areas of the Iberian Peninsula e.g., but the spatial scale used and the number of pollen sites available render microrefugia undetectable in this study.

The distribution data for Mediterranean species were downloaded from GBIF portal (http://www.gbif.org). The raw fossil pollen data is accessible for most used sites at the European Pollen database (http://www.europeanpollendatabase.net). For the remaining sites, data is available by request to the author of the respective publication (see Table 1 for a description of the data sets and sources). The climate data is accessible at the WorldClim website (www.worldclim.org). The output reconstructions, anomalies and variance maps are available in NetCDF format as supplementary material.