Background & Summary

Weather and climate influence a wide range of natural processes, with significant implications for ecological phenomena. For instance, they shape geomorphological processes, such as erosion and soil formation1 and influence biodiversity by affecting population dynamics, species distribution, community assembly, and ecosystem productivity2. Therefore, further study of climate and climate change is crucial to improving our understanding of both the climate itself and the natural processes it affects3. This is also of practical importance due to the connections between climate and key areas such as water resources, food supply, health, and the economy4. In recent decades, the establishment and coordination of extensive meteorological networks have enabled the generation of global meteorological and climate data. When combined with atmospheric models, interpolation algorithms, and data assimilation methods, these data have facilitated the creation of comprehensive global meteorological and climatological datasets5. These datasets are invaluable in many fields, including climatology, geography, agronomy, architecture, urban planning, and economics6,7. In ecology and biogeography, they enhance our understanding of macroecological patterns8, enable the study of how biological entities (individuals, populations, communities, or biomes) interact with climate9, and help predict the impacts of climate change on biodiversity10.

Understanding how past environmental changes affected biodiversity, and how it responded, is essential for comprehending these dynamics11. This is especially relevant when studying periods of change that parallel the current rate of climate change, as it allows us to anticipate future effects and design appropriate management and conservation strategies12. Most studies in this field are based on spatial paleoclimate reconstructions derived from General Circulation Models (GCMs)13. These reconstructions are typically validated with independent paleoclimate data or proxies, such as fossil records or stable isotopes. These reconstructions have been instrumental in advancing our understanding of the climate system14, past environmental changes15, and the interactions between these processes16, as well as in improving future projections17.

Several projects have aimed to estimate climate data for the past, present, and future to improve predictions. For instance, the Paleo Model Intercomparison Project (PMIP) seeks to enhance climate models and their projections by analyzing and validating GCMs for different periods, such as the Last Glacial Maximum (LGM; ~21 ka BP) and the mid-Holocene (~6 ka BP). Although these projections were primarily designed for model evaluation, they have also been used in paleoecological studies through downscaling techniques to provide fine-scale paleoclimate reconstructions18. Similarly, the TraCE-21ka project used a coupled atmosphere-ocean GCM (Community Climate System Model version 3) to estimate global monthly climate data since the LGM, providing a foundation for other projects aiming to increase spatial resolution (e.g., PaleoView, CHELSA-TraCE21k) or the number of variables available (e.g., CHELSA-TraCE21k). Both PMIP and TraCE-21ka have been combined with modern climate datasets (e.g., WorldClim, CHELSA-TraCE21k) and future projections (e.g., CMIP) to create standardized climate datasets spanning the past, present, and future.

However, all these datasets have limitations. For example, most of them (e.g., PMIP3, TraCE-21ka, PaleoView) are available only at coarse spatial resolutions (Table 1), which can be problematic for regional or local studies where meso- and microclimates are critical. Other projects, such as WorldClim and EcoClimate, have high spatial resolution (~1 km) but limited temporal coverage, focusing on specific periods (e.g., LGM, mid-Holocene, 1970–2000, and 2100). This restricts their usefulness for studying continuous historical biogeographic patterns. CHELSA-TraCE21k addresses this limitation by providing data at 100-year intervals, although it does not retain the monthly temporal resolution of the original TraCE-21ka dataset. Regional analyses have produced finer resolution versions of these datasets, but they are often region-specific, require advanced programming skills, and are not always open access (but see19). The complexity of downscaling techniques can also deter non-experts from creating datasets at the desired resolution for specific regions.

Table 1 Description of several climate products in raster format at large/global scale publicly available, including their spatial extent, their finest spatial resolution, their temporal coverage and resolution, their data origin, and data source, and whether they are derived products.
Full size table

Most downscaling efforts in paleoecology use the delta approach, also known as the standard change-factor method, due to its simplicity and speed20. However, this method relies on assumptions, such as orography not affecting changes in climate variables equally. More advanced methods, such as mechanistic models, simulate physical atmospheric dynamics but are computationally intensive and require extensive data21. A compromise is the perfect prognosis approach, which uses observational data for both predictors (independent variables) and predictands (dependent variables) during calibration22. In this approach, observational data for predictands often comes from reanalysis projects that readjust model predictions through the integration of daily observations. Large-scale circulation variables, well represented by global models, are typically chosen as predictors. This approach assumes that the relationship between micro- and macroclimate is constant over time, making it more robust than the delta approach20. While perfect prognosis demands greater computational resources, it is more efficient than mechanistic downscaling in terms of cost and complexity.

The Mediterranean region, located between Africa, Asia, and Europe, is of great significance in biogeography due to its unique combination of diverse climatic conditions, varied topography, and long evolutionary history, which have contributed to its recognition as a biodiversity hotspot23. All these factors, and others like its role as a glacial refugia or long impact by human activities24, contribute to high levels of species richness, including a large number of endemic species25. The Western Mediterranean is particularly important, with a high rate of endemism26, and expected to be highly vulnerable to the impacts of climate change, such as increasing warming and reduced precipitation27. However, there is a lack of standardized, high-quality climate reconstructions for this region, especially at fine spatial and temporal scales, which are needed for studying past environmental changes and predicting future impacts.

This manuscript aims to develop a toolkit for applying advanced downscaling techniques to coarse-resolution climate datasets within the widely used R programming framework, specifically in ecology and paleoecology. This toolkit, in the form of two R-packages, will then be used to downscale climate data for the Western Mediterranean at the finest possible spatial and temporal resolution. While initially focused on the Western Mediterranean, the toolkit is designed to be adaptable to other regions globally. We aim to provide open access to a cutting-edge climate dataset for those interested in this region, and to promote the reuse of both the dataset and the toolkit to facilitate the creation of similar high-resolution climate products for other regions and scientific applications.

Methods

To achieve the proposed objectives, two R packages were developed: dsclim (https://dnietolugilde.com/dsclim/index.html) and dsclimtools (https://dnietolugilde.com/dsclimtools/index.html). The former carries out the downscaling process, while the latter manages the resulting dataset and, hence, will be most useful for users interested only in working with our dataset. The dsclim package was used to select the optimal configuration of the downscaling algorithm by performing a cross-validation across multiple parameter spaces. It was subsequently used to downscale monthly climate reconstructions from the LGM (22 ka BP; 1000 years before the end of the LGM) to future projections (up to 2100 cal years), using data from the TraCE-21ka project, the Coupled Model Intercomparison Project 5 (CMIP5), and the UERRA reanalysis project. This process yielded a high-resolution dataset for the Western Mediterranean region, both spatially and temporally (DS-TraCE-21ka)28. The dsclimtools package was subsequently used to analyze the downscaled dataset, assessing various aspects of climate change in the region since the LGM until the end of the 21st century, including spatio-temporal patterns. Additionally, our downscaled data (DS-TraCE-21ka data)28 were compared to the external CHELSA-TraCE21k dataset. Both packages are openly available in their respective GitHub repositories and documented through websites. These sites offer instructional guides on downscaling climate data and loading the dataset into R, with potential applications.

Package development

The developed packages integrate existing R packages (e.g., ncdf4, loader, downscaleR, transformer, etc.), rather than creating an entirely new framework. The dsclim package supports loading data from diverse sources, standardizing these data, and executing the downscaling process. On the other hand, the dsclimtools package simplifies the management and usage of downscaled data (DS-TraCE-21ka)28, allowing for filtering and combining datasets. By adhering to existing standard formats in R for spatio-temporal data (e.g., stars or terra packages), dsclimtools also facilitates statistical calculations (e.g., mean absolute deviation) and the derivation of bioclimatic variables using other R packages (e.g., dismo).

Source climate data

Climate data were retrieved from three sources to allow for an overlap period between past, present, and future periods at coarse resolution, as well as present data (i.e., historical period) at fine resolution. The database for the historical period was chosen for its superior spatial resolution. In all cases, identical variables were obtained across the three periods: maximum temperature, minimum temperature, average temperature, precipitation, wind speed, atmospheric pressure, relative air humidity, and cloud cover.

Paleoclimate data at coarse resolution were obtained from the TraCE-21ka paleoclimatic simulation29. This experiment provides a monthly reconstruction of climate since 22 ka BP until 1990, utilising the standard T31_gx3 grid with a spatial resolution of approximately 3.75° × 3.75°. Further detailed information regarding the dataset, including the meaning of file names, variable names, and units, can be found in the following websites: https://www.cgd.ucar.edu/projects/trace and https://www2.cgd.ucar.edu/ccr/strandwg/TraCE. Further information can also be found on the documentation website for the model and the developing laboratory: http://www.cesm.ucar.edu/projects/community-projects/LENS/data-sets.html and http://www.cesm.ucar.edu/models/atm-cam/docs/cam2.0/UsersGuide/UG-45.html.

The reanalysis climate data for the historical period (1961–1990) were obtained from the UERRA project30. The project provides reanalysis data for Europe and North Africa from two different systems (UERRA-HARMONIE and MESCAN-SURFEX) that produce data in the Lambert Conformal Conic Grid at a temporal resolution of six hours and two spatial resolutions: The UERRA-HARMONIE system provides data at a resolution of 11 km × 11 km, while the MESCAN-SURFEX system offers data at a resolution of 5.5 km × 5.5 km. The data can be obtained from the Climate Data Store on the Copernicus website (https://cds.climate.copernicus.eu/datasets/reanalysis-uerra-europe-single-levels). Further technical details may be found in the aforementioned Climate Data Store. The resolution of the precipitation variable (in the MESCAN-SURFEX) was aligned with that of UERRA-HARMONIE by calculating the mean value of the four pixels at 5.5 km, which were equivalent to each of the pixels at 11 km in the rest of variables. We opted in favor of UERRA against other data sources because of its superior spatial resolution (e.g. CRU-TS and ERA5), being a reanalysis rather than an interpolation (e.g. E-OBS), and/or being specially designed for the European and Mediterranean context rather than at a global scale (e.g. ERA5-Land).

The future climate data were obtained from the CMIP5 multi-model ensemble31, which are accessible from a variety of sources, including https://pcmdi.llnl.gov/mips/cmip5/ and the Climate Data Store. The CMIP5 data were obtained for two distinct time periods. In terms of the historical period, this corresponds to the interval between 1850 and 2005 as defined by the CMIP5 project. In contrast, projections for future scenarios were made for the 2006–2100 interval. Despite the availability of data from multiple GCMs within the CMIP5 datasets, only three models (i.e., CESM1-CAM5, CSIRO-Mk3-6-0 and IPSL-CM5A-MR; see Table 2) possessed all the selected variables for all four representative pathways (RCPs; RCP2.6, RCP4.5, RCP6.0 and RCP8.5). Consequently, the variables were downloaded from these three GCMs and four RCPs.

Table 2 CMIP5 models and the institutions that provided model output used in this study.
Full size table

Downscaling

A perfect prognosis downscaling process was carried out using the climate4R framework32 for R33. The core of this statistical downscaling process is the construction of a transfer function between the predictand variables (i.e., fine resolution data) and a linear combination of the predictor variables (i.e., data at coarse resolution) for each pixel at fine resolution22. Accordingly, the model is calibrated using historical data at two distinct spatial resolutions, with a GCM estimated for each pixel at fine resolution (see Fig. 1 for a graphical illustration). Once the coefficients of the regression models have been calculated, the software uses past or future data at coarse resolution as input and applies the regression models to them, thereby obtaining past or future data at fine resolution. This process is repeated for each variable that is to be downscaled.

Fig. 1
figure 1

Diagrammatic workflow for the perfect prognosis downscaling approach. The simple regressions displayed in the scatter plot represent an example of those used in the actual downscaling approach, since they only include a predictor variable while the actual downscaling was performed with multiple regressions including eight predictor variables combined in different ways (spatial, local and global predictors). Number of points in scatterplot is 360 (30 years × 12 months) for each spatial point (blue or orange).

Full size image

The transfer function was built using Generalized Linear Models (GLMs) with a Gaussian link family for all variables except precipitation and wind speed, which were fitted according to a Gamma link family. Prior to the formal downscaling, a number of configurations of the downscaling algorithm were tested and evaluated. To this end, a cross-validation was conducted, whereby the historical period (1961–1990) was split into five-year intervals. In each cross-validation exercise, the models were fitted using 25 years of data, with the remaining five years used for model evaluation. Consequently, six models were fitted for each configuration. In particular, different configurations were tested in order to define the predictor variables (see Table 3). These included the use of ‘spatial predictors’ (i.e. the predictor variables were transformed and reduced using the main axis of a Principal Component Analysis that explain more than 70% of the original variation); the use of ‘global predictors’ (i.e. the application of values from all pixels in the study area at coarse resolution as predictors); and the use of ‘local predictors’ (i.e. the application of only values from a specific number of closest pixels as predictors). All tests were conducted using the full set of eight variables at the original, coarse resolution. The downscaled data from each configuration were evaluated by calculating the bias for each dataset, comparing the downscaled values (i.e., the estimated fine-resolution values for each pixel at a monthly time resolution) against the observed data from the UERRA dataset. In particular, three summary metrics of error were calculated for each pixel, representing the mean, standard deviation, and skewness of bias, respectively. The optimal configuration for downscaling the entire period was identified as M1.sp, in which spatial predictors were employed, and PCA axes explaining up to 70% of the variance were retained.

Table 3 Focal variable is the variable that is being downscaled.
Full size table

The resulting downscaled dataset was then compared with an external source of data, namely the CHELSA-TraCE21k dataset. As illustrated in Table 1, the CHELSA-TraCE21k project provides climate data at a spatial resolution of 30 arc sec from 21 ka to the present at one-hundred-year intervals. For purposes of comparison, these data were spatially aggregated to align with the resolution of our own dataset. A comparison between CHELSA-TraCE21k and our own dataset is of interest in order to identify any procedural bias or error present in any of the datasets. The information was extracted from the two datasets at two different pixels of the downscaled data (high spatial resolution), which were part of the same pixel of the TraCE-21ka dataset (coarse spatial resolution) and corresponded with two different orographic conditions (i.e. valley bottom and mountain range).

Data Records

Our dataset is stored on Zenodo28, organized within the following folder structure. There is one folder for each climate change scenario (i.e., RCP 2.6, RCP 4.5, RCP 6.0, and RCP 8.5). Within each RCP folder, there are three subfolders corresponding to the GCMs used (i.e., CESM1-CAM5, CSIRO-Mk3-6-0, and IPSL-CM5A-MR), and each GCM folder has one subfolder per variable containing a NetCDF file for each year. On the other hand, the past data is fragmented per time and variable since it is heavier than the previously mentioned. All the folders have the same naming structure, the years range and the variable itself (e.g., −22000_-16000_BP_pr, precipitation data from −22000 to −16000 BP).

The downscaled dataset spans the Western Mediterranean, including Southern Europe and North Africa, covering the geographical area defined by the coordinates 11.04° W to 12.04° E and 27.96° S to 44.04° N. It provides monthly climate data since 22 ka BP up to the year 2100. The variables included in the dataset are: maximum temperature (°C), minimum temperature (°C), average temperature (°C), precipitation (mm), wind speed (m/s), relative air humidity (percent), and cloud cover (fraction). All variables are presented at a spatial resolution of 11 × 11 km, ensuring high-resolution data for regional and local-scale analyses.

Technical Validation

The cross-validation conducted for the downscaling of the historical period produced good results for all model configurations, showing consistently low mean bias across variables (Fig. 2). However, configurations without spatial predictors (M21, M24, M31, and M34) showed increased variability in both the standard deviation and skewness of the bias (Figures S1, S2 respectively). As a result, the simplest model (M1.sp) was selected to downscale all variables since the LGM until the year 2100. Furthermore, the spatial pattern of the bias from the simplest model (M1.sp) showed no artifacts at the borders of the TraCE-21ka pixels, suggesting that the bias from this downscaling configuration has a smoother spatial pattern. Although these evaluation metrics were favorable, it is important to acknowledge that the cross-validation relied on a relatively short historical period of only thirty years, while our downscaling approach aims to downscale thousands of years. Hence, this method’s main limitation is that uncertainties increase as the examined period deviates from the cross-validation period, requiring caution when interpreting the results and underscoring the need for external validation of the dataset.

Fig. 2
figure 2

Violin plots for mean bias in cross-validation of the downscaling of the historical period (1961–1990) along model configurations: M1.sp (uses spatial predictors but do not use global or local predictors), M21 (uses the focal variable as local predictors with 1 local pixel, but no spatial or global predictors), M21.sp (uses the focal variable as local predictor with 1 local pixel and spatial predictor, but no global predictors), M24 (uses the focal variable as local predictors with 4 local pixel, but no spatial or global predictors), M24.sp (uses the focal variable as local predictors with 4 local pixel and spatial predictors, but no global predictors), M31 (uses all variables as local predictors with 1 local pixel, but no spatial or global predictors), M34 (uses all variables as local variables with 4 local pixel, but no spatial or global predictors).

Full size image

To circumvent the limitation of our cross-validation, we compared our dataset with the original TraCE-21ka and the CHELSA-TraCE21k dataset, the last one relying on the same low resolution climate data (i.e. TraCE-21ka) but with different high resolution contemporary climate data and downscaling approach. When comparing the average temperature values from the original TraCE-21ka pixel with two downscaled pixels (both within the same TraCE-21ka original pixel) and the same two pixels from the CHELSA-TraCE21k data, two main conclusions emerged (Fig. 3). First, the higher resolution of both downscaled datasets enabled the differentiation between points with varying topographies; both downscaled datasets successfully captured the higher temperatures in the valley and the lower temperatures in the mountainous region. Second, both approaches agreed on the distinct dynamics observed between the two points, where mountain areas showed less variation over time compared to the valley. This alignment between the two datasets confirms that fine-resolution data can capture topographic effects on average temperature, providing valuable insights for biogeographical studies, especially at micro- or meso-scales. As the time series fluctuations align well (e.g., Bølling–Allerød, Younger Dryas), we can consider the downscaled dataset externally validated, eliminating the possibility of significant procedural errors. However, some differences were noted between our downscaled data and the CHELSA-TraCE21k data. In the case of the warmer region, our dataset shows temperatures 2-3 °C lower than CHELSA-TraCE21k, while in the colder region, our dataset is around 5 °C lower. These differences are already present in the original data used for the historical period in both downscaling procedures (CHELSA data for CHELSA-TraCE21k and UERRA data for our downscaling). Consequently, these discrepancies reflect limitations in the fine-resolution source data, rather than any bias introduced during downscaling. This consistency raises concerns about mixing downscaled datasets with different baselines. It also suggests that similar patterns could arise if we used a different data source for the high spatial resolution data in the historical period (calibrating phase; e.g. E-OBS or ERA5-Land), with uncertainties arising from the differences between data sources. Further venues of research could explore these differences to provide better assessments of uncertainties due to data sources.

Fig. 3
figure 3

Comparison of temperature and precipitation as reconstructed by -CHELSA-TraCE21K and DS-TraCE-21ka (our dataset of the downscaled TraCE-21ka). Black line is the temperature/precipitation as in the original TraCE-21ka for the coarse resolution pixel highlighted in Fig. 1. Orange and blue lines are the temperatures/precipitations in the orange and blue points, respectively, in Fig. 1 according to CHELSA-TraCE21k (dashed line) and DS-TraCE-21ka (continuous line). Arrows indicates the mean historical temperature/precipitation in the original data sources at fine resolution (UERRA for DS-TraCE-21ka and Present CHELSA for CHELSA-TraCE21k). Grey is representing two periods of rapid climate change in the series, cooling and warning, that are used to illustrate uses of the dataset.

Full size image

The precipitation comparison (Fig. 3), on the other hand, revealed some other discrepancies. Both datasets suggested lower precipitation in the valley than in the mountainous pixel, but the precipitation values in both datasets exceeded those of the original TraCE-21ka dataset at coarse resolution. This discrepancy could arise from fine-resolution pixels with exceptionally low precipitations within the corresponding coarse-resolution pixel. Additionally, CHELSA-TraCE21k exhibited greater precipitation variability than our dataset and TraCE-21ka, particularly in the mountain. However, past studies based on soil development have documented more stable precipitation regimes in floodplains34. Given the challenges of accurately estimating precipitation at global scale, this higher variability in the CHELSA-TraCE21k is not unexpected. It is noteworthy that the precipitation trend observed in the mountainous pixel conflicts with the general trend observed in the coarse resolution pixel. This is not entirely surprising since precipitation is often influenced by local processes, such as convection or rain shadow effects, which can cause variations at a finer resolution. In this case, downscaling at a regional scale could be an advantage to reconstruct paleoprecipitation regimes compared to global approaches (i.e., CHELSA-TraCE-21k), although we cannot confirm it. Unfortunately, literature on the topic doesn’t help to support the trends observed by the two approaches; discrepancies in precipitation trends at local scale during the last deglaciation have been reported to range from reduced precipitation to no change, or even highly spatially variable anomalies35. Given the higher variability in patterns and uncertainties in downscaling of precipitation, the choice of the high spatial resolution data for the historical period could have a greater impact than for temperature, which reinforces the need for further research exploring the uncertainties associated with data sources (e.g. ERA5-Land or E-OBS).

When incorporating future data alongside past data, both datasets demonstrated a coherent progression, as they were based on the same baseline (UERRA reanalysis, see Fig. 4). This smooth transition minimizes biases that might arise from merging different datasets. The consistency between historical reconstructions and future projections is critical for analyzing long-term climate trends. Notably, all future scenarios, regardless of emission levels, followed this trajectory, lending further validity to our methodological approach and allowing for a robust comparison of future climate projections with historical data.

Fig. 4
figure 4

Temperatures for the same two points of the study area for the calibration period (1960–1990) and the average future projections among three climate models (e.g. CESM1-CAM5, CSIRO-Mk3-6-0, and IPSL-CM5A-MR) under four different Representative Concentration Pathways (RCPs; rcp2.6, rcp4.5, rcp6.0, rcp8.5).

Full size image