Modern datasets that cover extensive geographic regions (e.g., Europe) span a wide range of climatic conditions and are thus often challenging for reconstruction methods designed to estimate pollen-climate relationships (i.e., WA-PLS and CREST). In Europe, this difficulty is compounded by the underrepresentation of modern cold climates, limiting their use for reconstructing past cold conditions. The presence of such heterogeneity can reduce the ability of these methods to accurately match fossil samples with modern analogues. Without proper pre-processing of modern analogues, the performance, reliability, and robustness of reconstruction methods are compromised, potentially leading to less precise climate estimates (Cao et al., 2022).

Biomes are large-scale ecological units characterized by dominant vegetation types and the climatic and environmental conditions that support them (e.g., Prentice et al., 1992, 1996). They capture broad vegetation patterns and climate-vegetation relationships over space and time, facilitating comparisons across regions and temporal scales. Biomization processes prepare the data by classifying the pollen assemblages into PFTs and biomes, which are then used in reconstruction methods. Guiot et al. (1993) suggested using the biomization approach of Prentice et al. (1992) to better distinguish between steppe and tundra environments, thereby improving the interpretation of fossil assemblages when good modern analogues are lacking. The scarcity of suitable modern analogues strongly depends on the temporal window considered for reconstruction, particularly during climatic periods characterized by vegetation compositions substantially different from those observed today.

In this study, biome and megabiome classification methods were applied to modern and fossil (LGM) European pollen assemblages. By processing a large European modern dataset (see Sect. 2.3) through the biomization procedure, we minimize the risk of lacking analogues and integrate spatial heterogeneity of the modern samples to reconstruct regional vegetation patterns. The methodology for processing pollen data to extract climatic signals is illustrated in Fig. 1.

Biomization was originally designed to reconstruct past vegetation patterns across specific regions of the world (e.g., Europe, Africa) for key periods such as the Mid-Holocene or the LGM (Prentice et al., 1996). This approach relies on a standardized methodology that uses modern and/or fossil pollen data (Prentice and Webb, 1998; Prentice et al., 1996, 2000; Peyron et al., 1998) and generally proceeds from taxa to PFTs and then from PFTs to biomes through a square-root transformation using pollen taxon abundances (Prentice et al., 1996; Prentice and Webb, 1998). Here, the taxa list defines PFTs following the European biomization scheme from Peyron et al. (1998) and Tarasov et al. (2000). A pollen abundance threshold of 0.5 % (Prentice et al., 1996) – 0.2 % for Corylus – is subtracted from the initial pollen taxon abundances to account for pollen transport effects (Prentice et al., 1998). Counts for Larix and Pinus were multiplied by factors of 15 and 0.5, respectively, to account for their specific levels of production (i.e., underproduction and low preservation) and pollen overproduction, respectively (Bigelow et al., 2003; Binney et al., 2017).

We assigned 98 pollen taxa to 25 PFTs (section “Code and data availability”, from Prentice et al., 1996), using the sum of the square roots of the pollen abundances corrected for thresholds. This approach allows for floristic and functional heterogeneity within key biomes. Then, biome affinity scores were calculated by considering the sum of the PFT scores belonging to each biome. Intermediate biomes (Tarasov et al., 1998) were used to distinguish warm and cold conditions from steppe and desert biomes by re-assigning PFT scores based on the dominant biome. Additionally, PFTs were combined into 16 biomes following Prentice et al. (1996) and using the standardized biome assignment procedure from Harrison et al. (2010) and Dallmeyer et al. (2019). The Taxa-PFT and PFT-biome biomization tables are available in the section “Code and data availability” (csv file) with the following biomes: cold deciduous forest (CLDE), taiga (TAIG), pioneer (PION), cold mixed forest (CLMX), cool conifer forest (COCO), temperate deciduous forest (TEDE), cool/warm mixed forest (COMX/WAMX), xerophytic shrubs (XERO), tundra (TUND), cool/warm steppes (COST/WAST), cold/hot desert (CODE/HODE), anthropogenic (ANTH), and aquatic (AQU) biomes.

Biomization studies have considered three main hypotheses regarding the approach to (i) selecting the pollen-PFT assignment based on vegetation and environmental distributions from specific regions (Prentice et al., 1996, 2000), (ii) prioritizing the least PFT-rich biome when affinity values for multiple biomes are identical (Harrison et al., 2010; Dallmeyer et al., 2019) and (iii) choosing dominant biomes with the highest affinity score (absolute proportion of a particular biome) across different regions (Chen et al., 2010). Generally speaking, biomization may help mitigate the impact of non-analogue situations (i.e., when modern samples are considered too different from the fossil ones to be used for reconstructions) by reducing taxonomic complexity and grouping taxa into climatically constrained plant functional types (Prentice et al., 1996; Birks, 1998).

Megabiomes are broad vegetation categories that group multiple biomes into larger ecological and climatic units (Dallmeyer et al., 2019). This classification, used alongside biomes in this study, relies on a more spatially extensive and global framework, yielding results that are more directly comparable to global vegetation models. Compared to the biomization procedure, only the content of the European classifications (linking taxa to PFTs and PFTs to biomes) is modified, while the methodology for calculating affinity scores remains the same. This classification aims to (i) define a small set of biomes (i.e., called megabiomes) based on shared PFTs, and (ii) cover a broader climatic range than individual biomes.

In this study, we used the megabiome classifications from Dallmeyer et al. (2019) and Cao and Tian (2021). The main aim of the PFTs–megabiome assignment is to track key signals of land-cover changes (e.g., wetland extent, conifer dominance) while minimizing the influence of the chosen biome classification (Prentice et al., 1998). Megabiome assignments are primarily used in some vegetation models (e.g., BIOME4, Kaplan et al., 2003), particularly when coupled with climate model simulations (as GCMs, Haywood et al., 2009). Affinity scores from pollen taxa were only calculated for pollen taxa percentages greater than 0.5 % in order to increase signal-over-noise ratios (e.g., Chen et al., 2010), following the same scheme as that of the previously described biomization approach (see Sect. 3.1.2). For particular species, such as Larix and Pinus, we applied the same pollen production and transport corrections as described above for the biomization.

For Europe, 242 pollen taxa were grouped into PFTs and further assigned to 6 megabiomes (Dallmeyer et al., 2019): temperate forest (TEFO), boreal forest (BOFO), warm-temperate forest (BOFO), tundra and polar desert (TUND), grassland and dry shrubland (STEP), and desert (DESE). We directly used the algorithm implemented in R (Cao and Tian, 2021; R version 4.2.3, https://www.r-project.org/, last access: 6 October 2025). When affinity scores were found equal, the megabiome with the fewest PFTs was selected. The correspondence of biome and megabiome tables are reported in Supplement 3 for typical taxa in Supplement 4. Table 2 summarizes pollen taxa to biome and megabiome assignment schemes used in this study.

The Modern Analogue Technique (MAT; Overpeck et al., 1985; Guiot, 1990) is widely used to reconstruct past climate variations over time scales ranging from the last million years to the Holocene, and has been applied in particular to glacial periods (Guiot et al., 1993; Davis et al., 2024; Zumaque et al., 2025; Charton et al., 2025). It measures the dissimilarity between a fossil pollen assemblage and a set of modern assemblages from the modern dataset (Guiot et al., 1989; Guiot, 1990). Climate reconstructions are performed by estimating a dissimilarity coefficient (hereafter SQChord, for “Squared chord distance”, (Overpeck et al., 1985) based on the closest modern analogues, using an optimal number of analogues “k” that minimizes the RMSEP (i.e., Root Mean Squared Error of Prediction, Juggins, 2020). This parameter “k” is specific to each climatic variable and fossil sequence (“k” ranges from 8 to 10). Generally speaking, the MAT method appears to perform better with pollen taxa that are diverse and poorly resolved taxonomically like family (Birks, 1995; Williams and Shuman, 2008; Viau and Gajewski, 2009). We used the rioja package in R (Analysis of Quaternary Science data, version 1.0-6, R Core Team, https://cran.r-project.org/web/packages/rioja/index.html, last access: 6 October 2025; Juggins, 2020) to perform MAT reconstructions. Climate reconstruction uncertainties were quantified using the root mean square error of prediction (RMSEP) derived from leave-one-out cross-validation.

The Weighted Averaging Partial Least Squares method (WA-PLS; Ter Braak and Juggins, 1993) is a transfer function based on climatic optima and tolerances of a pollen taxon defined by its spatial niches. With this method, the climatic responses of taxa are assumed to be Gaussian (Ter Braak and Looman, 1986). This reconstruction method assumes unimodal (Gaussian) species response. The number of components (i.e., portion of the covariance between species abundance data and environmental variables) for paleoclimate reconstructions was selected based on those that produce the lowest root mean square error (RMSEP), a high prediction linear correlation coefficient (R2), and less biased reconstructions to minimize mean and maximum bias (Ter Braak and Juggins, 1993). A minimization of the number of components (npls; Ter Braak and Juggins, 1993) is performed to weight and integrate potential interactions between taxa, in addition to the link that connects them to the climate. The rioja package was also used to carry out WA-PLS reconstructions. Uncertainties for WA-PLS reconstructions were similarly assessed through RMSEP obtained via leave-one-out cross-validation.

Probabilistic methods provide more accurate and flexible climate reconstructions than traditional statistical methods, with lower dependence on taxonomic resolution (Netzel et al., 2025). CREST describes the conditional responses of taxa assemblages to climate variables by integrating modern pollen–climate relationships (Chevalier et al., 2014; Chevalier, 2022). Most often, when pollen is identified at the genus or family level, the probabilistic method combines the individual species' parametric PDFs into a single PDF representing the pollen observed in the fossil sequence. Once estimated for all taxa, the PDFs for the taxa present in a sample are multiplied together, each weighted according to its observed abundance. CREST is particularly well suited for regions where pollen data are scarce or geographically unevenly distributed (Chevalier, 2019; Chevalier et al., 2020). In detail, this method defines the conditional response of a plant species to a climate as a parametric PDF. The taxon-climate link is established by combining these PDFs, weighted by the abundance species that make up the genera and families of the observed pollen. A composite likelihood distribution is obtained by multiplying the individual PDFs, providing a complete probabilistic representation of the climatic conditions compatible with the observed assemblage – thus accounting for the full ecological range rather than only optimal conditions. This approach allows CREST to quantify uncertainty explicitly, offering a robust framework for evaluating reconstruction reliability and comparing it with other statistical methods (MAT or WA-PLS).

To date, CREST has proven particularly effective in regions with sparse or poorly sampled data where other techniques cannot readily be used (Chevalier et al., 2020) and has been widely applied to several tropical ecosystems such as savannas (e.g., Chevalier and Chase, 2015, 2016; Chevalier et al., 2021). CREST reconstructions were generated using the CrestR package (https://github.com/mchevalier2/crestr, last access: 6 October 2025, Chevalier, 2022).

The modern distribution of biomes (from EMPD2 pollen dataset) across Eurasia shows that these are geographically homogeneous (Fig. 2.1a), with typical frequencies of biomes from 5 % to 20 % (Fig. 2.1b), except for cold deciduous forest (CLDE), desert (both cold (HODE) and hot HODE)) and supplementary biomes (pioneer (PION), anthropogenic (ANTH) and aquatic (AQU)), which are all minor (<1 %) across Europe. Western Europe (e.g., France, Spain, Italy) exhibits the greatest biome diversity.

While cool conifer and cold mixed forests are primarily located in Scandinavia, cool mixed forests, taiga, and xerophytic woodlands surround the Mediterranean region. The warm mixed forest biome is dominant in southern Spain, Italy, and parts of the Near and Middle East. The clustering of different biomes in close geographic regions is attributed to broad climate patterns, altitude gradients, anthropogenic impacts (e.g., deforestation, agriculture, climate change), and/or local effects (e.g., proximity to water sources). The lack of data in Western Siberia results in an underrepresentation of cold modern climate conditions.

Mean annual temperatures (TANN) are spatially variable in Europe and distinguished by (mega)biomes, reflecting ecological and climatic differences across regions (Fig. 2.1c and 2.2c). TANN is generally negative in tundra regions (TUND, medians ∼-9 °C) and significantly higher in temperate forests (TEDE), cool mixed forests (COMX), warm mixed forests (WAMX), warm steppes (WAST), xerophytic woodlands (XERO), and hot deserts (HODE) (Fig. 2.1b). Other biomes, such as cold deciduous forests (CLDE), taiga (TAIG), pioneer (PION), cold mixed forests (CLMX), cool mixed forests (COMX), and cold steppes (COST), exhibit intermediate climatic conditions. The overlap of COMX and TEDE biomes in geographic and climatic spaces is likely due to their physiognomic similarities (Binney et al., 2017). As proposed by Prentice et al. (1998) and Tarasov et al. (2000), using the same biomization procedure as in this study, we observe a dominance of tundra and taiga in Russia, where modern conditions are among the driest in Europe.

The modern results derived from the major megabiome analysis (Fig. 2.2a) indicate a strong dominance of the temperate megabiome (TEDE) (∼65 %, Fig. 2.2b) across Europe, covering much of Western Europe and seemingly obscuring the greater diversity of biomes observed in Fig. 2.1a. While mountain regions exhibit a mix of CLMX, COMX, and COST biomes, the megabiomization analysis yields a predominance of TEDE, TUND and STEP (Fig. 2.1b). The climate space of the temperate megabiome (TEDE, Fig. 2.2c) is more arid (i.e., colder and drier) than the temperate biome (TEDE, Fig. 2.1c). Consequently, steppe megabiomes – including boreal forest (BOFO), grassland and dry shrubland (STEP), warm desert (DESE), and tundra and polar desert (TUND) – are sparse, although they share common locations with their modern biome counterparts (Fig. 2.2c and Table in Supplement 3).

LGM biome score results from the 43 selected European fossil sites reveal a clear dominance of the cool steppes, called COST (colored outlines around the pie charts from Fig. 3a). However, the dominant megabiome reconstructions oppose temperate (TEFO) and tundra and polar desert (TUND) megabiomes (Fig. 3b), with a few instances of steppic conditions around the Mediterranean region. In detail, (mega)biome scores of the LGM samples are very close to each other, indicating a strong competition and therefore only a weak dominance of any particular (mega)biome. While the mean DESE biome score remains null throughout the LGM, the combined scores of arid megabiomes (i.e., grassland and dry shrubland (STEP) + desert (DESE) + tundra and polar desert (TUND)) generally represent a proportion comparable to that of the TEDE megabiome (points with black outlines in Supplement 5). Even when one megabiome appears dominant during the LGM, several other megabiomes also display substantial scores. These results (Fig. 3a and b) indicate an absence of a dominant vegetation ensemble, while transitional vegetation conditions are also apparent through the co-occurrence of multiple (mega)biomes. The selection of the dominant (mega)biome is thus based on marginal differences between competing scores, which underscores the limitation of this approach when several (mega)biomes co-exist within the same climatic period. By accounting for non-dominant (mega)biomes, our (mega)biome affinity score-weighted mean method based on (mega)biome scores can better capture underlying signals and provide more nuanced climate reconstructions across Europe.

By comparing biome and megabiome dominance scores, we show that present-day steppe conditions (COST and WAST biomes) do not systematically correspond to the steppic (STEP) and tundra (TUND) megabiomes. As defined in Supplement 3, these results suggest that the (mega)biome classifications (based on pollen and PFTs) in the two schemes reflect distinct climatic conditions (Fig. 1). This therefore suggests that the choice of classification could influence the reconstructed climate conditions. The search for the best analogues across different climatic spaces grouped by (mega)biome highlights the challenges of distinguishing local climatic conditions and isolating LGM vegetation analogues for climate reconstructions at the European scale.

LGM climate anomalies based on the (mega)biome affinity score-weighting approach and the three reconstruction methods are shown in Fig. 5 and summarized in Supplement 6. For the three different reconstruction methods, results at each core site are presented as anomalies compared to the modern climate estimated from ERA5-reanalysis (Hersbach et al., 2020; Table 2). From both vegetation classification methods, the LGM cooling across all fossil sites ranges from 1 to 12 °C, in line with LGM conditions significantly different from today. The mean European LGM cooling using the three reconstruction methods (MAT, WA-PLS, and CREST) yield 6.5±2.3, 7.1±2.0, and 5.8±2.4 °C (mean SD, 1σ) from the biomization method (Fig. 5a, c and d) and 7.0±2.2, 7.0±2.0 and 6.5±2.0 °C from the megabiomization method (Fig. 5b, d and f), respectively. The largest LGM coolings are observed in mountain ranges, such as the Alps and the Pyrenees (∼10 °C), with an agreement across all methods (standard deviation of ∼1 °C, Supplement 6). The Lake Bouchet (France), Kersdorf-Briesen (Germany) and Mickunai (Lithuania) fossil sites show the largest disagreement (∼2 °C standard deviation across the three methods). We could not detect any spatial gradient, or any influence of the proximity of the Fennoscandian ice-sheet from these two sets of reconstructions (i.e., biomization and megabiomization techniques, Fig. 5).

The first row of Fig. 6 (Fig. 6a–c) presents reconstructed LGM summer temperature (MTWA) anomalies based on the biomization procedure, with values of -3.7±2.3, -3.7±2.2, and -3.5±2.4 °C (mean SD) for MAT, WA-PLS, and CREST, respectively. The three reconstruction methods yield the most negative MTWA anomalies in southeastern Europe (∼-5 °C), whereas values are more homogeneous and closer to zero in northern and western Europe. MTWA anomalies derived from the megabiomization procedure (Fig. 7a–c) are -5.0±2.2, -5.3±2.2, and -5.1±2.4 °C for the same reconstruction methods, respectively, yielding similar spatial patterns in MTWA anomalies (Fig. 7) that from the biome assignments (Fig. 6). The winter temperature (MTCO, second row of Figs. 6 and 7) anomalies are -7.7±4.0, -8.5±2.9, and -6.2±3.5 °C from biomization procedure and -7.0±4.0, -7.1±3.1, and -6.1±4.0 °C from megabiomization procedure, for MAT, WA-PLS, and CREST, respectively.

Seasonality anomalies (i.e., seasonal amplitudes) can be calculated from the difference between MTWA and MTCO anomalies at each site, derived from biomization and megabiomization procedures, as shown in the third row of Figs. 6 and 7 (panels g–i). Negative LGM–modern seasonal anomalies indicate that seasonality during the LGM was weaker than under present-day conditions, while positive values indicate a stronger seasonal contrast relative to the modern climate. Results from the biomization procedure (Fig. 6g–i) show seasonality anomaly patterns that are broadly consistent with those obtained from the megabiomization approach (Fig. 7g–i). Across Europe, the reconstructed gradients indicate a reduction in LGM seasonality from western to eastern Europe relative to present-day conditions. In contrast, several sites located in southeastern Europe (e.g., Lake Xinias, Megali Limni and Dziguta) exhibit marked increases in seasonality since the LGM, reaching values of approximately 8 °C, comparable to those reconstructed for the extreme western part of Europe.

Because TANN, MTWA, and MTCO display spatially homogeneous patterns across Europe, and because the three reconstruction methods yield comparable spatial averages, we averaged the outputs (mean ± SD; Fig. 8) from the three methods for both the biomization (Figs. 8a and 9c) and megabiomization (Fig. 8b and d) procedures in order to reconstruct mean spatial climate conditions during the LGM. Compared to annual and summer temperatures, MTCO results show greater dispersion across European sites despite higher internal errors (Supplement 6), consistent with the validation uncertainties discussed in Sect. 5.2. Seasonal uncertainties are then derived as the quadratic sum of summer and winter temperature errors. Mean errors for these two climatic variables are reported in Supplement 8, also accounting for external uncertainties derived from inter-site dispersion.

Across Europe, LGM cooling averaged 6.4±2.0 and 6.8±3.4 °C (Supplement 7) based on biome and megabiome assignments, respectively (Fig. 8a–b). These results are generally consistent across reconstruction methods, with inter-method standard deviations below +2 °C (Supplement 7). Inter-model mean MTWA anomalies are similar for both reconstruction methods and for biome versus megabiome assignments, with mean multi-method values of -3.6±2.2 and -5.1±2.2 °C, respectively. Inter-model mean MTCO anomalies are -7.4±3.3 °C for the biomization approach (Fig. 7d–f) and -6.7±3.4 °C for the megabiomization approach (Fig. 7d–f). Mean seasonal anomalies across methods are 3.8±2.8 and 1.6±2.7 °C for biome and megabiome assignments, respectively (Fig. 8c–d). Seasonal anomalies are relatively consistent across methods using the biomization approach, with inter-method standard deviations of ∼2 °C (Supplement 7).

To isolate and quantify the impact of local calibration (i.e., (mega)biomization procedures) on global climate reconstructions during the LGM across Europe, Fig. 9a shows TANN anomalies calculated as the difference between climates obtained using the (mega)biome mean-weighting approaches and those derived from global calibration (i.e., by sampling the entire calibration EMPD2 pollen dataset).

As the MAT method is inherently based on dissimilarity coefficients (Guiot et al., 1989), it is theoretically able to directly identify the dominant analogues with no requirement for (mega)biome calibration using all modern pollen spectra. As a result, one expects only a minimal (if any) effect of (mega)biomization on TANN reconstructions from the MAT method (Fig. 9a). Here, only a few fossil sites show a warming effect (∼+5 °C) of reconstructed TANN using local calibrations based on (mega)biomes (Venice, Billerio and Feher) rather than the global calibration. Then, the MAT method may therefore be suitable for large-scale spatial studies using global datasets to reconstruct annual temperatures across Europe, although its performance statistics should be interpreted with caution given its sensitivity to spatial autocorrelation and the potential for over-optimistic skill estimates. While the effects of local versus global calibration TANN anomalies for MAT may depend on the time period considered and the reconstructed climate variables (Xu et al., 2010), the magnitude of these effects remain generally close to zero. However, the WA-PLS method produces generally inconsistent results between local and global calibration results, with a mean TANN anomaly of ∼-5 °C, indicating colder TANN using local rather than global calibrations. Two significantly negative anomalies (∼-10 °C), were also identified using the WA-PLS method (Lake Iznik, Kersdorf-Briesen). Similarly, the CREST method appears more sensitive to calibration, as evidenced by predominantly negative TANN anomalies (from ∼-2 to ∼-12 °C, with a mean of -6.5 °C, Fig. 9a). This underscores the importance of a robust (mega)biomization scheme – climatic and environmental gradients of the study area – when applying CREST in Europe during the LGM to ensure that reconstructed TANN remains sufficiently cold and consistent with other statistical methods (see Figs. 6 and 8). The importance of modern regional calibrations for quantitative climate reconstructions has also been noticed in previous studies (e.g., Trasune et al., 2024).

As shown in Fig. 9b, applying the (mega)biomization procedure to MAT substantially increases reconstructed MTWA by approximately 3 °C. In contrast, applying the same procedure to WA-PLS and CREST results in lower reconstructed summer temperatures, with MTWA decreasing by about 3 °C. In Fig. 9c, the differences in MTCO anomaly results between local and global calibrations are however more variable than MTWA results, across fossil sites. MAT and CREST methods exhibit a strong dependence on the calibration dataset for winter temperature reconstructions, which could be attributed to the spatial variability of MTCO anomalies in the modern calibration dataset. In detail, the MAT method generally produces larger MTCO anomalies with local calibration than with global calibration (+4.0 °C on average), whereas the WA-PLS and CREST methods tend to yield similarly lower MTCO anomalies (-4.5 °C on average). Then, these findings highlight an opposite effect of the (mega)biomization procedure from the MAT versus WA-PLS and CREST methods in reconstructing the MTWA and MTCO variables. Overall, these results highlight the dependence of climate reconstructions on the chosen method, calibration, and climatic variables. The sites most strongly affected by the (mega)biomization approach are predominantly located in Italy and along the Mediterranean coast (from Lago della Costa to Orvenco, Fig. 9a–c), where reconstructions of all three climatic variables show the largest deviations.

For TANN and MTWA reconstructions, MAT appears to be the most suitable reconstruction method, for identifying dominant analogues within a global calibration dataset to infer past climate conditions. However, MTCO outputs are sensitive to the modern subsamples defined by biome and megabiome classifications (Fig. 9), from MAT and CREST methods.

Across reconstruction methods, TANN, MTWA and MTCO results derived from the biomization process are consistent with the megabiomization outputs (i.e., anomalies defined as the difference between the two treatments – are close to 0, Fig. 10a and b). This result highlights the minimal impact of distinguishing between cold and hot arid conditions into steppe and desert biomes (not the case for megabiomes) on reconstructed LGM climates across Europe from three different methods.

Due to the large MTCO reconstruction errors based on the three methods (Supplement 6), quadratic errors between the biomization and megabiomization approaches are greater for MTCO than for TANN and MTWA (Fig. 10a–c). Nevertheless, MTCO anomalies reconstructed using MAT display only limited spatial variability across sites. Interquartile ranges are approximately twice as large as those obtained for MAT-based MTWA and TANN anomalies when comparing the biomization and megabiomization procedures (Fig. 10c).

To study the reliability of our LGM climate reconstructions with results published in literature, we compare our biomized and megabiomized multi-method results for three different climate variables with those of Davis et al. (2024). Davis et al. (2024) also used the calibration dataset (EMPD2 dataset) and some input fossil records, common with our study. Comparison maps (differences) between the two studies are presented in Supplement 8 and summarized as boxplots in Fig. 12. We use error metrics to compare the results of the two studies and assess the effect of methodological differences on climate reconstructions. Unlike R2, which is insensitive to systematic biases, error metrics (RMSE) express deviations in the original units (°C), making them more suitable for comparison.

The TANN and MTWA errors (RMSE) between our multi-method reconstructions and those of Davis et al. (2024) are comparable to the individual reconstruction errors (2–3 °C; Fig. 12a–b), indicating relatively good agreement between the studies regardless of the calibration ((mega)biomization) and reconstruction methods. In contrast, MTCO anomalies show larger discrepancies between the two studies (4–6 °C; Fig. 12c). MTCO results exhibit poorer agreement between the two studies, with RMSE values approximately twice as high as those for TANN and MTWA (Fig. 12c), for both megabiomization and biomization approaches. Overall, across these three methods, the results of Davis et al. (2024) agree more closely with our TANN and MTWA reconstructions than with our MTCO results, with RMSE differences ranging from 0.5 to 2 °C. Supplement 8 further shows that annual climatic conditions reconstructed for sites located in northern Europe are systematically warmer in our study than in Davis et al. (2024), with differences ranging from +5 to +10 °C. Interestingly, our approach leads to substantially higher MTCO anomalies in Northern Europe (∼+3 °C) and lower MTCO anomalies in Southern Europe (∼-3 °C).

These results align with those reported in Sect. 5.4, highlighting the predominant influence of the (mega)biome-type assignment scheme (Figs. 9c and 10b) and reconstruction methods on MTCO anomalies, likely due to the high spatial variability of modern data within the calibration dataset. MTCO reconstructions show the strongest inter-study inconsistencies and are particularly sensitive to methodological choices. The difference in modern calibration samples used in Davis et al. (2024) and in the (mega)biome definitions of our study may explain the observed differences in reconstructed temperature, particularly the large MTCO discrepancies. These discrepancies demonstrate the strong methodological dependence of palaeoclimate reconstructions and the importance of systematic cross-validation frameworks.

The extent of LGM cooling inferred in this study ranges from -1 to -12 °C relative to present (Fig. 5) and appears to fall within the range simulated by climate models (e.g., Kageyama et al., 2021). We also provide new constraints on LGM temperature seasonality indicating an amplification near the Atlantic coasts and a relative reduction inland, compared to present day. This continental pattern of LGM seasonality is in strong agreement with the mean model from PMIP synthesis by Izumi et al. (2013), which shows a gradient from positive anomalies in Spain (+10 °C) to negative anomalies (-4 °C) in Eastern Europe.

However, comparing pollen-based climate reconstructions with climate model outputs is complex due to both the heterogeneous spatial coverage of the data, the variability of reconstruction methods and the dispersion among climate models (e.g., Bartlein et al., 2011; Kageyama et al., 2021). Moreover, proxies record extreme and local climates, while climate models tend to provide spatially-averaged climate conditions. Ongoing improvements in the spatial resolution of regional climate models make comparisons with local proxies increasingly appropriate, although significant data–model discrepancies remain (Kageyama et al., 2021). Similarly, expanding the availability of pollen data in Europe would further improve the robustness of such comparisons.