The study region covers the north-central Mediterranean Basin (Fig. a) and is divided into two study zones, i.e., southern France and Italy.

Zone 1 (southern France) extends from the southern French coast to the north of the Massif Central, including part of the Pyrenees and the Alps and extends from longitude 1° E to longitude 8° E. This region is influenced by a Mediterranean climate, associated with cool mild winters and hot dry summers (Fig. b and c). Mean annual precipitation ranges from 500 to 800 mm, with a maximum recorded in autumn and a minimum in summer. Rainfall seasonality is influenced by altitude, resulting in an increase in spring precipitation with altitude while temperatures decrease . Because of its latitude, southern France is directly influenced by large-scale atmospheric circulation patterns over the North Atlantic and Europe, i.e., a persistent low-pressure area located over Iceland, alongside the semi-permanent high-pressure system associated with the Azores . Low-pressure conditions over Iceland, associated with an anticyclonic pattern in the eastern Atlantic, result in cold, dry, northerly winds over southern France, particularly in the Gulf of Lion. Meanwhile, high pressure brings warm, humid, southeasterly winds in the region . This dynamic interaction between Iceland and the Azores influences westerly airflow patterns, affecting thermal exchanges across the North Atlantic and Europe, especially during the winter months at mid to high latitudes. The westerly flow also varies inter-annually, shifting between northern and southern paths. This variation, known as the North Atlantic Oscillation (NAO), plays a significant role in climate variability in the western Mediterranean. The strength of the NAO is determined by the difference in normalised sea level pressure (SLP) between high and low latitudes in the Atlantic . A positive NAO indicates a stronger meridional pressure gradient and more intense westerly winds, while a negative NAO suggests a weaker gradient and weaker westerlies. During winters with a positive NAO, subtropical atmospheric pressure tends to rise, while Arctic pressure drops. This setup leads to varying winter weather: southern Europe experiences warmer and drier conditions, northern Europe sees warmer and wetter weather, and Greenland faces colder, drier winters .

Zone 2, corresponding to Italy, is located in the northern-central Mediterranean basin and stretches from latitude 36° N to latitude 47° N, making for contrasting climatic conditions from one region to another. The Italian orography is also complex and contrasted due to the presence of the Alps to the north and the Apennines, which stretch along the entire peninsula, acting as a barrier to air masses from continental Europe. Both mountain chains influence the pathway of weather fronts and interact with dominant winds, thereby exposing various regions of Italy to distinct circulation patterns . The Alps act as a barrier, limiting the influx of cold air masses from central Europe into the Po Valley and northern Italy. The Apennines diminish the vulnerability of western areas to cold easterly winds that originate from the Balkans, while the Adriatic Sea, situated on the upwind side, is more susceptible to strong winds and heavy rainfall . The Italian Peninsula, surrounded by the Mediterranean Sea to the west, south and east, has its climate strongly mitigated by the presence of the water body, which represents an important source of heat and moisture . The simultaneous impact of multiple geographic elements (e.g., elevation, distance from the sea, presence of particular coastal currents, exposure to dominant winds, etc.) shapes the presence of distinct climatic zones, each associated with particular prominence of weather types.

Over the last decades, several methods have been developed to reconstruct climate from pollen data see review by. As these methods are based on different ecological concepts and mathematical algorithms, results can be strongly method-dependent . In this frame, multi-method approaches have been developed to increase the reliability of the results . These methods were initially developed to calibrate the relationship between modern pollen data (soils, mosses) and current climate parameters. Multi-method approaches include (1) assemblages approach, based on the principle of dissimilarity between fossil and modern assemblages (Modern Analogue Technique); (2) transfer functions, based on linear or non-linear regressions (Weighted Averaging Partial-Least Squares regression) between pollen taxa and climate parameters, and (3) recent machine learning techniques with regression trees (Random Forest and Boosted Regression Trees) to quantify climate parameters.

The Modern Analogue Technique MAT; is widely used to reconstruct past climates due to its straightforward application, efficacy, and sensitivity. This technique relies on evaluating the dissimilarity between each fossil and modern pollen assemblages and selecting the closest modern samples (known as analogues). The WA-PLS method, introduced by , operates as a transfer function that assumes an unimodal relationship between the proportions of pollen and climatic conditions. It suggests that the abundance of a species is intrinsically linked to its environmental tolerance. WA-PLS calculates the climatic optimum for a species based on calibration data by determining the average climatic conditions in which the species is found, weighted according to its abundance .

The two additional methodologies, namely Random Forest (RF) and Boosted Regression Trees (BRT), have emerged more recently and are grounded in Machine Learning principles . These techniques utilise regression trees to systematically partition pollen data through successive divisions based on the abundance observed in the pollen spectrum. The Random Forest approach relies on the estimation and aggregation of multiple regression trees, with each tree being derived from a collection of pollen samples using a bootstrapping technique . In contrast, Boosted Regression Trees differ from RF in their treatment of the modern dataset: while RF assigns equal selection probabilities to all samples, BRT increases the likelihood of selecting under-represented samples from the previous tree. This technique, known as boosting, enhances the model's predictive accuracy for elements that are less accurately predicted . Here, the final output of the BRT method is derived from averaging the results of 15 independent executions of the BRT algorithm, reflecting the variability inherent in regression tree signals.

For each method, the reliability of the results was estimated by bootstrap cross-validation by calculating the values of the correlation coefficient between the variables (R2) and those of the root mean square error criterion (RMSE).

The relationships between climate variables were assessed with a scatter-plot matrix to provide an overview of the distributions and correlations of the variables, and a pair-plot was drawn.

To assess the relationships between modern pollen spectra and the climatic variables, an ordination technique was used on each modern pollen dataset. Major pollen taxa (those present in a least 15 samples and with a maximum ≥ 3 %) of each dataset were square-root-transformed to stabilise variances and optimise the signal-to-noise ratio . A detrended correspondence analysis DCA; was applied to each modern pollen dataset to determine whether a linear-based analysis or unimodal-based analysis was more appropriate based on gradient length as the criterion. The DCA results showed that, for the three modern datasets, the gradient length was over 3.0 standard deviation, suggesting that unimodal-based methods should be used in further analyses on the modern pollen datasets .

A canonical correspondence analysis (CCA) was carried out to detect the influence of climate variables on each modern pollen dataset, and variance inflation factors (VIFs) were also used to determine the correlations between environmental variables. VIF values > 10 indicate co-linearity with other variables, which may lead to unreliable results . To eliminate the correlation between the variables, we selected the temperature and precipitation variables with a large explained variance and eliminated those with a VIF > 10.

Climate reconstructions based on different methods and reliability tests (R2 criterion and RMSE) were performed with the packages rioja version 1.0.7, randomForest version 4.7-1.2 and dismo version 1.3.16. All analyses were performed on R Studio version 4.4.1, using the ggplot2 package version 3.5.1 for graph creation.

To facilitate the comparison of climate signals between each record and the identification of potential patterns of changing climate trends between records, a normalisation between -1 and 1 (rescaling min-max) is applied to the smoothing of the mean of selected models for each record. Data normalisation is based on the following equation Eq. (): 1x′=a+(x-min⁡(x))(max⁡(x)-min⁡(x))×(b-a)

With x non-smoothed initial mean values, x′ normalised smoothed values, a and b minimal and maximal normalisation values, in this case, respectively -1 and 1.

Because normalisation is based on the overall mean of the record, any interval influenced by human activity is deliberately excluded from the standardisation process. This precaution prevents the representation from being skewed toward artificially drier or warmer values. Determining the boundaries of human impact relies either on author notes from the original fossil record publications or on detecting a significant abundance of anthropogenic indicator taxa such as: Olea, Juglans, Castanea, Cerealia-type, Plantago lanceolata, and Rumex-type.

Composite curves are constructed for the central Mediterranean area and will be compared to other palaeoclimate syntheses of the Mediterranean region . For the composite curves, non-normalised reconstruction results were used, without the exclusion of periods where the human activity is discernible. Reconstructed palaeoclimate values of every climate parameter were averaged in a 300 year bin, the maximal resolution of all records being 306 years. For each record, the first bin was centred on 0 years BP, and the following bins were centred on a 300 year increment throughout the record. Then the binned values of each record were averaged to produce a regional climate signal for the central Mediterranean. Finally, the regional climate signal values were transformed into anomalies. The anomaly calculation is based on the reconstructed modern average value. The estimation of confidence intervals for each composite, encompassing the 5th and 95th percentiles, was achieved through bootstrap resampling at the site level, utilising 1000 iterations.

The pair-plots of the ten climatic variables for all three modern pollen datasets (Fig. ) indicate that mean annual temperature (MAAT) and precipitation (MAP) are highly correlated (Pearson correlation coefficient > 0.7) with the seasonal parameters (Taut, Tspr, Twin, Tsum and Paut, Pspr, Pwin, Psum). Among seasonal parameters, spring (Tspr and Pspr) and autumn (Taut and Paut) parameters are the ones which are also highly correlated between them. This co-linearity is supported by the canonical correspondence analyses (CCAs) of the pollen assemblages and the climate variables for the three modern datasets (Fig. ). For each modern pollen dataset, the first CCA, with every climate parameter, indicated that the variance inflation factor (VIF) values of annual (TANN, PANN), spring (Tspr and Pspr) and autumn (Taut and Paut) parameters are greater than 10 (Fig. a–c). After deleting those parameters, the remaining four climate variables (Tsum, Twin, Psum and Pwin) have VIF values lower than 10 and, therefore, can be used in the final CCA to investigate their influence on modern pollen datasets (Fig. d–f).

Results of model performances, estimated by bootstrap cross-validation, are summarised in Fig.  with the correlation coefficient (R2, Fig. a) and the root mean square error criterion (RMSE, Fig. b1 and b2). For all climate parameters and all modern datasets, best R2 and RMSE values, i.e., respectively highest and lowest values, are obtained for MAT and BRT methods. Conversely, the WA-PLS and RF methods show lower R2 and RMSE values. Therefore, we will retain here the MAT and BRT methods for the interpretation and discussion of our climate reconstructions.

The MAT and BRT were applied to each fossil pollen record to reconstruct the summer and winter climate parameters (Figs.  and ). To compare more easily the different sequences, the results obtained with both methods were averaged for each record. This averaged signal was then normalised (rescaling min-max method, Eq. ) according to the mean value of the smoothed climatic signal for the entire record.

We discuss here the chronology range, quality and resolution for each record, which are particularly important as they may induce some biases in the interpretation of the climatic reconstructions. For this study, specific criteria were applied to select records with a fairly good chronology, i.e., a continuous temporal range between 10 000 and 5000 yrcal.BP and an average temporal resolution of less than 350 yrsample-1. However, it should be pointed out that the quality of the chronologies is not equal between each fossil pollen record, which may initially affect the accuracy of the reconstructions, but also the interpretation that will be made of them and can be considered as an unavoidable limitation in our reconstructions.

Another chronological concern is linked to the partitioning of Holocene climatic trends. Certain climatic periods that are emblematic of the Holocene in Europe, such as the HTM, are not always identified at the same time and/or with the same intensity, for reasons that may depend on chronology, proxy sensitivity (marine vs terrestrial) and/or local response to climatic change e.g., elevation,. This is why the time interval definition is not always a simple or obvious choice. Divisions specific to the Mediterranean region have been proposed, notably by who proposed a division into three periods with (1) a lower humid Holocene (11 500–7000 yrcal.BP), (2) a transition phase (7000–5000 yrcal.BP) and (3) an upper Holocene (5500 yrcal.BP–present) characterised by aridification.

In contrast to the aforementioned study and to better discuss the climate changes of the central Mediterranean area in terms of spatio-temporal patterns, we choose to divide the Holocene into four periods: (1) 12 000–8000 yrcal.BP corresponding to the Lateglacial and early Holocene, (2) 8000–5000 yrcal.BP during which the HTM should be observable if recorded, (3) 5000–3000 yrcal.BP associated with post-HTM conditions and (4) 3000 yrcal.BP–present during which the human impact on vegetation is recorded in most of fossil pollen records.

Among the uncertainties that can impact our palaeoclimatic reconstructions, it seems essential to address the limitations of the pollen proxy when used in palaeoclimatic quantification. One of the first uncertainties that can be stated is the effect of the agglomeration of several archive types (e.g., lakes and peat bogs) with different local recording conditions (e.g., sediment accumulation rate, edaphic conditions, hydroseral succession). Variability in spatial representativeness can affect the reconstructed palaeoclimatic signal, e.g., a lake associated with a small catchment will record a more local signal than a lake associated with a large catchment, which would better represent regional vegetation . A second factor, a migrational lag, can impact our reconstructions based on pollen data, as vegetation can take time to return to equilibrium with its environment after a major and rapid climate change . The impact of the migrational lag is generally considered for the Lateglacial–early Holocene transition, which corresponds to the last rapid climate shift of the last 12 000 years BP, when most of the postglacial vegetation began to take hold and before climate changes became less strong and less rapid. These uncertainties could impact the reconstructions of the Lateglacial and early Holocene periods of our study. Finally, it seems essential to address the impact of human presence and its interaction with vegetation when pollen data are used to reconstruct palaeoclimates. Several studies have already shown that human activity, through land clearance for grazing or cultivation, can lead to an over-representation of non-arboreal pollen, and thus influence palaeoclimate reconstructions in which vegetation changes do not entirely correspond to climatic variations . In the central Mediterranean, studies showed that the main cause of vegetation changes before 4000 years BP was climatic variations, but after, from the late Holocene onward, vegetation changes seemed to be attributed to human activity and climatic influences altogether . For our study, the modification of vegetation by humans is a point to be taken into account for palaeoclimate interpretations since the presence of anthropogenic taxa (i.e., Olea, Juglans, Castanea, Cerealia-type, Plantago lanceolata and Rumex-type) is observed in several records of zones 1 and 2 from around 3000 years BP onward.

Studies have highlighted the importance of site effects on the response to climate change, such as complex topography and elevation, which can act as important perturbation factors to large-scale atmospheric flows . The study by , focused on the Alps, showed that differences in the pollen assemblages of high (≥ 1000 m) and low (< 1000 m) elevation sites play a significant role in the quality of climate reconstructions based on pollen. In our study, elevation also seems to play a role, as the HTM observed in the south of France and the Italian Alps is best recorded at high elevations north of 43° N. However, this interpretation may be questioned because (1) most pollen located north of 43° N is located at high elevations (19 out of 28 records) and (2) high elevation vegetation in the Alps may be mostly controlled by summer temperature while low elevation vegetation in southern Italy may be more controlled by water availability and/or precipitation. A greater number of high-elevation records south of 43° N would enable us to refine our climate interpretation, particularly in the south of France, where continuous pollen records are lacking.

The pioneering study by first highlighted a north–south palaeohydrological contrast in the central Mediterranean during the Holocene. In addition, they identified a latitudinal “tipping point” around 40° N , splitting and distinguishing two distinct and opposite patterns on either side of latitude 40° N. In contrast to the study of , which is based on 6 records, with only one record located between latitude 40 and 45° N, our synthesis is based on 38 records, 22 of which are located along the Italian Peninsula, giving us a unique opportunity to see whether the same north–south division also appears in our results at a wider scale. Here, a latitudinal division on either side of 43° N in Italy throughout the Holocene is evidenced in summer while winter climate conditions appear to have been more spatially homogeneous and rather be marked by a stronger temporal dynamic (Figs. , , , and ). Our study corroborates the findings of but suggests that the “tipping point” takes place at higher latitudes, around 43° N, and only impacts summer climate conditions (Figs.  and ). Through a multi-proxy study, also highlighted a north–south climate division in Italy during the Lateglacial period, proposing a threshold around latitude 42° N. Our study aligns with this pattern and suggests that the location of the latitudinal threshold dividing the Italian climate may have varied slightly through time and space.

During the early Holocene period (12 000–8000 yrcal.BP), the south of 43° N is characterised by relatively cold and wet summer conditions, regardless of elevation, while the north of 43° N is rather associated with hot and dry ones, particularly at high elevation (Figs.  and ), which is in line with the observations made by at this period. Winter conditions were, however, rather cold and dry throughout the early Holocene period (Figs.  and ), in contrast to what was suggested by the aforementioned study. The mid Holocene period (8000–5000 yrcal.BP) was associated with relatively warm and humid winter conditions in the main part of Italy. Warmer and drier summers are reconstructed south of 43° N, in contrast to wetter summers recorded north of 43° N. This pattern for the mid Holocene differs from the study, which reported humid winters with dry summers north of 40° N and humid winters and summers south of 40° N during the mid Holocene. From the mid-to-late Holocene transition onward (5000 yrcal.BP to modern-day), winter and summer climate trends south of 43° N are globally similar to the mid Holocene ones, while colder climate conditions following the HTM are depicted in summer from the sites located north of 43° N at high elevation (Fig. a).

In the Mediterranean, the seasonal and spatial variability of precipitation is directly linked to the global atmospheric circulation . In their study, highlighted the combined effects of the blocking of the North Atlantic anticyclone linked to variations in summer insolation and the influence of ice sheets and the forcing of freshwater melt in the north Atlantic ocean, both of which have an impact on large-scale circulation processes such as the NAO. In addition to global atmospheric circulation, other more local factors, such as orography, latitude and oceanic and/or continental influences, are directly linked to the seasonal and spatial variability of precipitation. The intricate topography of central Italy, combined with its central location within the Mediterranean basin, leads to precipitation patterns that arise from a variety of meteorological processes and influences, including flows from the south–west, west, north–east, and south–east . This complexity complicates the relationship between precipitation variability and alterations in large-scale atmospheric circulation, rendering the connections among precipitation fluxes, orography, and the spatial distribution of rainfall more intricate than in other regions characterised by simpler precipitation mechanisms , such as those predominantly influenced by westerly winds, as seen in Southern California .

The north–south climate contrast does not appear to be the only specific climate dynamic in the Mediterranean basin. have shown in a synthesis on the mid Holocene climate transition that both proxy data and model outputs suggested a west–east division in the Mediterranean climate history. Specifically, western Mediterranean early Holocene changes in precipitation were significantly smaller in magnitude and spatially less coherent than the eastern ones, and the mid Holocene (around 6000–3000 yrcal.BP) recorded the rainfall maximum before a decline to present-day values. Studies suggest that these contrasting hydroclimate regimes indicate a complicated interaction between different atmospheric systems, e.g., the main phases of NAO-like circulation , the East Atlantic pattern and the size and position of the North African anticyclone, expressed differently in various sectors of the Mediterranean . This longitudinal component of the Mediterranean climate is also found in our study, which contrast the western regions of southern France with the northern Italian regions (Figs. b and b), and placing the central zone of southern France in a zone where the climatic signal is much more contrasted, suggesting the presence of a “buffer zone” where the different atmospheric dynamics confront each other. However, this hypothesis merits a more detailed regional study to understand the complexity of this transition zone, located in the southern part of the Massif Central region where relatively high elevation are encountered (> 600 m) and several climatic influences come together, i.e., Mediterranean influence from the south; the influence of the Atlantic Ocean from the west due to Atlantic air masses arriving from the country's west coast, which are not prevented by any geographical barriers obstacles in the Aquitaine Basin; and a mountainous regime from the north.

Our climate reconstruction, based on 38 records located in the central Mediterranean, highlights reliable spatio-temporal patterns at a regional scale despite an expected variability across the records that could reflect differences in latitude, elevation or in situ characteristics. We thus reconstruct homogeneous winter conditions in contrast to summers, which were characterised by a marked north–south reversal throughout the Holocene (Figs.  and ).