The Climate Data for Adaptation and Vulnerability Assessments and the Spatial Interactions Downscaling Method

The Climate Data for Adaptation and Vulnerability Assessments and the Spatial Interactions Downscaling Method | Scientific Data

Abstract

This study presents the spatial interactions downscaling (SPID) method and introduces the climate data for adaptation and vulnerability assessments (ClimAVA) dataset. SPID employs random forest models to capture the relationship between spatial patterns at global circulation model (GCM) resolution and fine-resolution pixel values. In summary, a random forest model is trained for each fine spatial resolution pixel of the reference data as the predictand, and nine pixels from the spatially resampled (coarser) version of the reference data at the GCM’s resolutions as predictors. Models are then utilized to downscale the bias-corrected GCM data. The ClimAVA-SW dataset offers a high-resolution (4 km), bias-corrected, downscaled future climate projection derived from seventeen CMIP6 GCMs. It includes three variables (daily precipitation, minimum and maximum temperature) for three shared socioeconomic pathways (SSP245, SSP370, SSP585) across the U.S. Southwest region. The ClimAVA dataset sets itself apart with the SPID method’s capacity to provide remarkable climate realism, high physical plausibility of change, and excellent representation of extreme events while maintaining user-friendliness and requiring relatively low computational resources.

Background & Summary

The repercussions of global climate change on the environment, economy, and social frameworks, once considered theoretical projections just a few decades ago, have materialized into reality 1. Independent of future emission choices, the world is committed to serious economic and environmental impacts 2, compelling us to adapt to already rapidly changing climate conditions. Adapting to climate impacts requires careful justification and planning, which should be grounded in a thorough understanding of these impacts’ expected scale and geographic patterns 3. A key component in this effort is leveraging future climate projections generated by GCMs to inform decision-making and planning processes. However, a common challenge in local assessments of climate change impacts is the inherent spatial coarseness of outputs from GCMs (50–300 km), rendering them unsuitable for direct application in local impact models 4.

Local to regional climate change impact, adaptation, and vulnerability assessments necessitate the utilization of data at higher spatial resolutions 5, 6, 7. Transforming coarse data to the desired high-resolution format is known as downscaling. Downscaling is crucial for converting the projected large-scale climate from GCMs into scales that properly represent the specific local areas of interest and their needs 8, 9. Two main categories of downscaling methods exist: statistical downscaling and dynamical downscaling, also known as regional climate models (RCMs) 10. Dynamical downscaling involves the utilization of an RCM for a specific region of interest nested within a GCM 11, 12. However, dynamical downscaling comes with drawbacks such as significant computational requirements, substantial data storage needs—particularly when multiple experiment realizations are necessary—and the propagation of biases from GCMs to RCMs 13, 14. On the other hand, statistical downscaling is considered more efficient and straightforward. These methods directly incorporate the relationship between observed data and GCM outputs 5.

Statistical methods, which involve establishing statistical relationships between coarse-scale (GCM-scale) and fine-scale climate from observed data and subsequently applying these relationships to future projections, have seen more application compared to dynamical model downscaling in studies addressing hydrologic impacts of climate change in the western United States 8, 9, 13, 15, 16. The most common statistical downscaling methods used in the literature include the delta method 17 and its variations bias correction spatial disaggregation (BCSD) 13, and delta hybrid 18; the localized constructed analogs (LOCA) 19, multivariate adaptive constructed analogs (MACA) 20 and deterministic methods 21.

Delta-based methods apply the shift (difference/delta) from GCM historical and future projections to observed data. It produces spatially and temporally consistent datasets with high climate realism, as precipitation and temperature events are equal to observed events plus the shift from the GCM projections. However, its main disadvantage is that the sequence of events on the future projections is the same as observed in the past - plus the shift (delta). This issue is known as low physical plausibility of change, and it limits the usability of these datasets for several studies; for example, studies that account for the number of dry days, length of the growing season, and the number of days over or under a temperature or precipitation threshold will likely present virtually the same result from the past (observed) climatology and not from the GCM projections 22. BCSD acknowledges that GCMs have biases, so it applies a bias correction step before spatially disaggregating climate data (similar to the delta method). The delta hybrid method took BCSD one step further, using a quantile delta mapping scheme to map monthly shifts from the GCMs to selected months of observed data 18. Quantile delta mapping adds more “variability” to the downscaled dataset; however, it suffers from disadvantages similar to those of the delta method as it keeps the sequence of daily events from the observed/past climatology from the selected months.

Deterministic downscaling methods use models trained on predictands (fine-scale climate data of interest) and predictors (coarse-scale climate data) and apply those models to downscale coarse GCMs. Under the umbrella of deterministic methods, the most used are regression models, support vector machines (SVM), and artificial neural networks (ANN) 21. Predictors can be any climatic variable present in both observed data and data generated by the GCM, such as different wind components, pressure, vorticity, etc. The advantage of these methods is that they keep the signal and sequence of events from the GCM, increasing the physical plausibility of change of the downscaled projections. Other statistical methods focus on using Constructed Analogs 19, 20. The main disadvantages of statistical methods are (1) the need for matching predictands and predictors from observed and GCM data, although some methods can extrapolate values, (2) most methods are limited by the data provided during training, (3) data manipulation is more complex than delta-based methods, these methods are (4) more computationally demanding when compared to delta-based methods, and (5) often present low climate realism, especially for daily and monthly precipitation.

In summary, the application of an ideal downscaling method should result in a high spatial resolution dataset that is bias-corrected, offers high climate realism, maintains physical plausibility of change from GCMs, maintains changes of signal from the GCMs, provides an accurate representation of extreme events, is easy to apply, and is not overly computationally demanding. Efforts to develop such an ideal method should be pursued to facilitate the democratization of the use of future climate data in local to regional climate change impact, adaptation, and vulnerability assessments.

This paper aims to introduce the spatial interactions downscaling (SPID) method and the climate data for adaptation and vulnerability assessments (ClimAVA) dataset, providing a high-resolution (4 km) bias-corrected, downscaled projection of future climate based on seventeen CMIP6 GCMs. The dataset includes three variables (pr, tasmin, tasmax) and three scenarios (SSP245, SSP370, SSP585) for the entire U.S. Southwest region. The ClimAVA dataset sets itself apart with the SPID method’s capacity to provide remarkable climate realism, high physical plausibility of change, and excellent representation of extreme events while maintaining user-friendliness and requiring relatively low computational resources.

Methods

Data acquisition

Gridded datasets of daily precipitation (pr), minimum surface air temperature (tasmin), and maximum surface air temperature (tasmax) at 4 km spatial resolution, covering the period from 1981 to 2020, were acquired from the parameter-elevation relationships on independent slopes model (PRISM 4 K) project ( https://prism.oregonstate.edu/) to be utilized for both bias correction and the training of downscaling models. The PRISM gridding method integrates geographically and elevation-weighted regression, station weighting based on topography, distance to the coast, and atmospheric variables 23. This dataset was chosen as our reference data for its temporal coverage from 1881 to the present, well-suited 4 km spatial resolution for our objectives, demonstrated reliability in representing precipitation patterns in complex terrains 24, favorable performance compared to alternative datasets 25, and its status as one of the most commonly used spatial climate datasets in the United States 26. Hereafter, PRISM data will be referred to as reference data.

Coupled model intercomparison project phase 6 (CMIP6) historical and future climate projection data 27, generated through GCMs for pr, tasmax, and tasmin, were obtained from the ESGF MetaGrid ( https://aims2.llnl.gov/search). The selection of models adhered to specific criteria: (1) daily temporal resolution; (2) inclusion of all three shared socioeconomic pathways (SSP) of interest – SSP245, SSP370, and SSP585 within the same model and realization; (3) availability of data for all variables of interest – pr, tasmax, and tasmin; and (4) nominal resolution falling between 100 and 250 km. Due to data storage limitations, only the first realization of each model was acquired. These criteria led to a total of 17 models (Table 1).

Area of interest and data preparation

The ClimAVA-SW 28 dataset encompasses the geopolitical boundaries of the six states in the United States Southwest: California, Nevada, Arizona, New Mexico, Utah, and Colorado. To facilitate hydrology modeling efforts, the complete area of watersheds that run into these states has also been incorporated into the spatial coverage of the dataset (Fig. 1).

Schematic diagram of the SPID approach, which involves the following steps: (1) downloading reference and (2) CMIP6 (GCMs) data, (3) resampling reference data to match the native resolutions of GCMs, (4) bias correcting GCM data using (a) Quantile Mapping (QM) for pr and (b) Equidistant Quantile Matching (EDCDFm) for tasmax and tasmin, (5) training random forest models for each pixel based on original (fine resolution) and resampled (coarse resolution) pixels from the reference data, and finally (6) downscaling the bias-corrected data from step 4 using the random forest models trained in step 5.

Before conducting any analysis, all datasets were subsetted to the designated area of interest. Among the 17 chosen GCMs, there are 15 different grid spacings (spatial resolutions) (Table 1). Grid space varies from 0.703125° × 0703125° (EC-Earth3 models) to 1.875° × 2.5° (NorESM2-LM). Our aim is to demonstrate the performance of SPID in downscaling from various resolutions without the need to regrid the GCMs. This means that all analyses are conducted on the models’ native grid, thereby avoiding any loss of information or distortion caused by regridding the data. To enable bias correction and training of the downscaling method, the observed dataset was resampled to match each of the 15 different grid spacings corresponding to the grid spacing of each GCM. A conservative resampling method was applied to adjust the reference data to match the GCM resolutions. Essentially, the resampled pixel represents the average value of all the pixels from the reference data that fell within each pixel of the GCM. This procedure was consistently applied to every GCM under consideration.

Bias correction

Bias correction was accomplished by quantile mapping (QM) for pr, and equidistant quantile matching (EDCDFm) for tasmax and tasmim (Fig. 1). Quantile mapping 17, 29 corrects the bias in a model value by adjusting it to the observed value (here represented by the resampled PRISM data) at the quantile within the model’s historical distribution. The EDCDFm method 30 corrects the bias in model value by adding the difference (∆) between the model future and historical value to the observed values based on the quantile from the model value (CDFs are constructed in 30-year windows). When bias-correcting historical data, EDCDFm behaves as QM (Fig. 1). Different methods were applied to different variables due to the concern that QM may alter the magnitude and even direction of mean changes projected from the original GCM 31, 32, 33, 34, especially for temperature data. Table 2 depicts the distribution for all 17 models of the difference in mean change from historical (1985–2014) and future (1971–2100) at SSP585 before and after bias correction. Differences were negligible up to 99 percentiles, indicating these methods successfully maintained the mean change (in magnitude and direction) from the original GCM after bias correction.

Downscaling

The SPID method assumes that climatic spatial patterns observed in the resampled (coarser) reference data are analogous to spatial climate patterns in GCMs. If those relationships hold true, they can be used to downscale GCM data. SPID employs a random forest model 35 to translate the relationship between spatial patterns at coarse resolution and fine-resolution pixel values. A random forest algorithm grows numerous decision trees on bootstrapped subsamples of the data. It combines the predictions of all these trees to forecast the out-of-bag data, which were not utilized in training the trees, thus quantifying prediction error. This algorithm is a fast, supervised classification method that enhances classification performance through the amalgamation of multiple decision trees and introduces stochasticity in the modeling process. Consequently, it can effectively handle correlated and interacting predictor variables 36. A random forest model is trained for each pixel, with the finer pixel of the observed data as the predictand and nine pixels from the spatially resampled (coarser) version of the reference data (at the GCMs spatial resolution) as predictors.

In practice, each fine-resolution pixel is associated with the corresponding coarse pixel (resampled observed data), which contains the finer pixel within its spatial boundaries. The model selects the eight surrounding coarse pixels, totaling nine predictors per pixel (see Fig. 1). All available daily values are used to train the model. Subsequently, during downscaling, the random forest model translates spatial patterns from coarse-resolution values (from GCM) to finer-resolution values on a daily basis. This process is repeated for each fine pixel, i.e., there is a random forest model for each pixel. The number of predictors may be less than nine for pixels located along the coast or the U.S. border. This limitation arises due to spatial coverage constraints within the reference dataset.

Using only one coarse-resolution pixel—the central one—as a predictor significantly diminishes the effectiveness of the downscaling process. On the other hand, employing more than nine predictors (e.g., 25) did not substantially enhance downscaling performance and led to substantially increased computational demands.

For the validation process, models were trained using 30 years of data (1981–2010) and validated using 10 years (2011–2020). Models intended for the downscaling application to create the ClimAVA-SW 28 dataset are trained using all 40 years of available data. The effectiveness of models is influenced by the “events” they learned during training. Consequently, they perform best with a larger training dataset as it provides more climatic spatial patterns to learn from over an extended time series. This limitation becomes particularly significant with the intensification of extreme precipitation and heat/cold waves 37, 38 and their implications for human society 39, 40 and the environment 41. For instance, extreme precipitation events not observed in the training datasets might be poorly represented by the downscaling process, as random forest models struggle to extrapolate information beyond its training range. More details regarding SPID’s capabilities in downscaling extreme events are discussed in the validation section.

SPID computational demand

The SPID method has been shown to be computationally efficient. To enhance speed while utilizing available resources (AMD 7713 P nodes with 64 cores and 256 GB memory), especially memory RAM, we train and export the models for each pixel. In our setup, training each model takes less than a second for pr and less than 2 seconds for tasmax and tasmin, per core. Parallelization significantly expedites the process. With all models trained, downscaling the entire Southwest region of the United States for five years’ worth of daily data requires only 11 minutes for pr and 31 minutes for tasmax and tasmin, utilizing 20 cores and 80GB of memory per job. Another advantage of SPID is its straightforward coding, which, when coupled with its relatively low computational requirements, positions it as a valuable tool for democratizing access to high-quality downscaled future climate projections; especially in areas where high-performance computing resources are not readily available.

Data Records

ClimAVA-SW 28 provides daily climatic data at ~4 km spatial resolution for 17 GCMs (Table 1), three different climatic variables (pr, tasmax, and tasmin), and three different shared socioeconomic pathways (SSP245, SSP370, and SSP585). Historical runs span from January 1, 1981, to December 31, 2014. Future scenarios span from January 1, 2015, to December 31, 2100. All years have 365 days (leap years are not included). Data is provided as NetCDF4 files. The ClimAVA-SW 28 dataset encompasses the geopolitical boundaries of the six states in the southwestern United States: California, Nevada, Arizona, New Mexico, Utah, and Colorado, and watersheds that run into these states (Fig. 1).

ClimAVA-SW 28 can be found at the World Data Center for Climate (WDCC) ( https://doi.org/10.26050/WDCC/ClimAVA-SW).

Technical Validation

As discussed, grid spacing (spatial resolution) varies considerably among the 17 selected GCMs. To demonstrate SPID downscaling capabilities, validation results are shown from downscaling done for two different grid spacings, representing the finest (0.703125° × 0.703125° from EC-Earth3) and coarsest (1.875° × 2.5° from NorESM2-LM) resolutions, hereafter referred to as FR and CR, respectively.

In this section, the coefficient of determination (R2), root mean squared error (RMSE), and mean prediction error (MPE) are employed to assess the comparison between SPID downscaled results and reference data on a daily time scale; the analysis of 10-year annual averages offers insights into the climate realism and spatial variability of SPID downscaled outcomes; the examination of temporal variability in annual data at specific locations in northern Utah (Cache Valley and Bear River Range) and California (Sacramento and High Sierras), provides insights into how SPID maintains temporal consistency in its downscaled results and; the distribution of daily event magnitudes in percentiles is further explored to evaluate SPID’s capability to downscale extreme events. Additionally, the matching of dry-wet days for precipitation is demonstrated.

Precipitation

SPID shows satisfactory results when downscaling both from FR and CR; however, downscaling from FR yields better results than those from CR (Fig. 2). The mean R2 for FR was 0.88 against 0.72 from CR. Downscaling from both resolutions yields higher R2 values in mountain complex regions (for ex., the Intermountain West and High Sieras) than in flatter regions (Western Colorado and Western New Mexico). I.e., SPID yields more accurate results in complex climatic regions where precipitation events are influenced by orographic processes compared to flatter regions where meteorology is strongly linked to synoptic signatures. As expected, RMSE is slightly lower for FR (1.08 mm) than CR (1.65 mm). Higher RMSE values are in higher precipitation regions, such as the High Sierras and the Pacific Coast Range. Meaning the errors are small when compared as a ratio of total daily precipitation, and flatter regions of Western Colorado and Western New Mexico. Mean MPEs were −0.04 mm and −0.05 for FR and CR, respectively. However, spatial and histogram distribution of MPE reveals how SPID performs best when downscaling from finer resolutions.

Daily precipitation (pr) validation for downscaling from the finest (FR) and coarsest (CR) spatial resolutions. R2 = coefficient of determination, RMSE = Root Mean Square Error, MPE = Mean Prediction Error.

Evaluation of the 10-year annual average precipitation demonstrates how SPID maintains climate realism and spatial consistency in downscaled data (Fig. 3). FR and CR show mean negative biases of −15.9 and −18.2 mm, respectively. Further analysis of the bias distribution indicates a generally dispersed pattern throughout the region of interest. However, larger disparities are notable in high-precipitation areas, where errors are relatively small when considered as a proportion of total precipitation. It is worth noting that the scattered errors observed may stem from spatial/temporal inhomogeneities in the reference dataset. Such inconsistencies could arise, for example, from temporal inconsistencies in station data, where station data are used when available but disappear from the station network used in the interpolation process on days (months, years) with missing data 25. Further analysis of the bias distribution in seasonal precipitation (Supplementary Figure 1) reveals a generally dispersed pattern similar to that observed in the annual data, indicating that errors are distributed similarly across different seasons.

Ten-year annual average precipitation from the reference data (Obs.) and their difference to SPID downscaled data from finest (FR) and coarsest (CR) spatial resolutions.

A comprehensive analysis of multiple locations was conducted to assess the potential impact of the biases encountered in the 10-year annual average analysis on climatic trends in downscaled data. For brevity, an analysis comparing annual data from the reference with downscaled data at four representative locations characterized by complex terrain is presented (Fig. 4). Results indicate that the overall bias observed in the 10-year annual average precipitation is distributed over time and does not influence trends. Downscaled annual precipitation values closely align with the reference data across the majority of locations. However, a few exceptions are notable, such as Naomi Peak in Northern Utah (Fig.

Annual precipitation values from reference data (Obs.) and SPID downscaled data from finest (FR) and coarsest (CR) spatial resolutions for Cache Valley (CV) and Naomy Peak (NP) in northern Utah and Sacramento (Sac) and Mount Price (MP) in California.

GCMs, due in part to their coarse spatial resolution, often exhibit many days with minimal precipitation, commonly referred to as “drizzle precipitation” 42, 43, 44. Given that drizzle precipitation is often unrealistic and can adversely affect subsequent analyses, it’s crucial to ensure that downscaled products do not include artificial drizzle precipitation. For validation purposes, days with precipitation amounts below < 0.1 mm/day were categorized as dry days 32, 45, after which the correspondence between dry and wet days in downscaled data and the reference dataset was computed (Fig. 5). Downscaling from FR and CR resulted in mean correspondences of 93.0% and 91.4%, respectively. Lower correspondence percentages were noted in the mountainous and topographically complex regions of the Intermountain West, particularly in Colorado. However, the percentages rarely dipped below 75%, even in areas with lower correspondence.

Correspondence between precipitation/non-precipitation days in reference data and SPID downscaled data from finest (FR) and coarsest (CR) spatial resolutions. Non-precipitation days were considered as days with less than 0.1 mm of precipitation.

Climate change has intensified the frequency and severity of extreme precipitation events, especially in dry areas like the Southwestern U.S. 37. Given the substantial impacts of these events and their anticipated increase with global warming, accurately representing them in terms of spatial distribution and magnitude through downscaling is crucial. SPID has exhibited an exceptional ability to downscale these extreme events (Fig. 6). For daily precipitation events at the 95th percentile, downscaling from both FR and CR yielded results that approach perfection, with mean differences of −0.03 mm and −0.12 mm, respectively, and values closely approximating zero across the region of interest. At the 99th percentile, mean differences were recorded as −0.99 mm and −2.25 mm for FR and CR, respectively. Once again, higher errors were detected in high-precipitation areas like the High Sierras and the Pacific Coast Range, as well as in the flatter regions of Western Colorado and Western New Mexico. As with previous analyses, downscaling from finer spatial resolutions yields better results than coarser resolutions, as observed in mean errors and the spatial distributions of errors.

Difference between the 95th and 99th percentile of daily precipitation events from the reference data (Obs) and SPID downscaled data from the finest (FR) and coarsest (CR) spatial resolutions.

It is worth noting that for daily extreme precipitation events, most differences between the reference and downscaled data were negative, indicating a tendency to underpredict extreme events. This phenomenon can be attributed to the nature of the downscaling method employed in SPID, which relies on Random Forest algorithms. Notably, only events present in the training dataset can be accurately reproduced during the downscaling process, as random forest models have limited capacity to extrapolate from values in the training data set. Therefore, it is hypothesized that the larger and more diverse the training dataset, the better the downscaling results should be, particularly when capturing the uniqueness of extreme events.

Surface air temperature

As observed for pr, SPID shows satisfactory results when downscaling tasmax and tasmin from both FR and CR resolutions and downscaling from finer resolution yields better results than from coarse resolution (Figs. 7, 8). Average values of R2 for both variables (tasmax and tasmin) and resolutions range from 0.97 to 1, indicating a strong correlation between the downscaled and the reference data. Average RMSE is slightly lower for FR (0.56 °C) than CR (0.98 °C) for tasmax and tasmin (FR = 0.80 °C and CR = 1.25 °C). RMSEs are homogeneously distributed in space, with no apparent error clusters associated with landscape features or regions. The average MPE values are 0.04 °C (FR) and 0.01 °C (CR) for tasmax and 0.02 °C (FR) and 0.07 °C (CR) for tasmin. Although mean MPE is virtually zero for both variables and resolutions analyses of the spatial distribution and pixel count (histogram) indicate again how downscaling from FR yields better results than when downscaling from CR.

Daily maximum air temperature (tasmax) validation for downscaling from the finest (FR) and coarsest (CR) spatial resolutions. R2 = coefficient of determination, RMSE = Root Mean Square Error, MPE = Mean Prediction Error.

Daily minimum air temperature (tasmin) validation for downscaling from the finest (FR) and coarsest (CR) spatial resolutions. R2 = coefficient of determination, RMSE = Root Mean Square Error, MPE = Mean Prediction Error.

Evaluation of the 10-year annual average tasmax and tasmin demonstrates how SPID maintains climate realism and spatial consistency in downscaled data (Fig. 9). The average difference between the reference and downscaled data for both tasmax and tasmin for FR and CR is virtually zero, implying little to no biases in the tasmax and tasmin data downscaled by SPID. However, a closer look into spatial distributions of errors indicates that pockets of errors, in both directions, are spatially scattered and with no recognizable pattern and, as discussed for pr, could be partially due to spatial/temporal inhomogeneities in the PRISM data used as reference. Further analysis of the bias distribution in seasonal tasmax and tasmin (Supplementary Figures 2, 3) reveals a generally dispersed pattern similar to that observed in the annual data, indicating that errors are distributed similarly across different seasons.

Ten-year annual average of daily maximum (tasmax) and minimum (tasmin) surface air temperature from the reference data (Obs) and their difference to SPID downscaled data from the finest (FR) and coarsest (CR) spatial resolutions.

As for pr, a comprehensive analysis of multiple locations was conducted to assess the potential impact of the biases encountered in the 10-year annual average analysis on climatic trends in downscaled data (Fig. 10). Results indicate that the overall bias observed in the 10-year annual average of tasmin and tasmax is also distributed over time and does not influence trends. Downscaled annual tasmax values perfectly match the reference data in all analyzed locations. SPID tasmin downscaled data has been shown to follow trends in annual data closely but not as well as for tasmax. An unexpected pattern found in the data is that SPID yielded better results in peaks (see Naomy peak in northern Utah and Mt. Price in the High Sierras) than in the vales (see Cache Valley and Sacramento), highlighting the capacity of SPID in resolving climate in challenging landscapes.

Annual average of daily maximum (tasmax) and minimum (tasmin) values from reference data (Obs.) and SPID downscaled data from the finest (FR) and coarsest (CR) spatial resolutions for Cache Valley (CV) and Naomy Peak (NP) in northern Utah, and Sacramento (Sac) and Mount Price (MP) in California.

Climate change has intensified the frequency and severity of extreme temperature events 46. Given the substantial impacts of these events, accurately representing them in terms of spatial distribution and magnitude through downscaling is crucial. SPID has demonstrated exceptional ability in downscaling extreme temperature events (Figs. 11, 12). The tasmax data downscaled from FR has yielded accurate results across various percentiles, with an average difference between reference and SPID data consistently below 0.1 °C for the 5th to 99th percentiles. Notably, the average disparity remains minimal even for the 1st percentile, which presents a slightly less impressive outcome, at only 0.16 °C across the region of interest. As anticipated, downscaling from CR has resulted in slightly inferior outcomes; however, the average difference between reference and downscaled data remains below 0.21 °C from the 5th to 99th percentile.

Difference between 1st, 5th, 50th, 95th, and 99th percentile of daily maximum surface air temperature (tasmax) from the reference data (Obs) and SPID downscaled data from the finest (FR) and coarsest (CR) spatial resolutions.

Difference between 1st, 5th, 50th, 95th, and 99th percentile of daily minimum surface air temperature (tasmin) from the reference data (Obs) and SPID downscaled data from the finest (FR) and coarsest (CR) spatial resolutions.

The downscaling of tasmin extreme temperatures has yielded remarkable outcomes. From FR, across the 5th to 95th percentiles, the results approach perfection, with an average mean error ranging from 0.12 °C to −0.08 °C. The average errors at the 1st and 99th percentiles stand at 0.21 °C and −0.24 °C, respectively. While results from downscaling CR data are satisfactory, they are less impressive, particularly for the 1st and 99th percentiles, where average errors of 0.47 °C and −0.47 °C were observed, respectively. As anticipated, downscaling for the first percentile tends to be overestimated, while those for the 99th percentile tend to be underestimated.

Downscaled GCM data

As described before, SPID employs a model per pixel and downscales daily data. This implies that each downscaled value is processed independently of its neighboring pixels or the preceding and succeeding days’ values. Consequently, the downscaling procedure is both temporally and spatially independent. One might wonder if climate realism is maintained in such an approach. Figure 13 illustrates the spatial distribution of climatic patterns for a single day, a month, and a year of pr and tasmax downscaled from the EC-Earth3 model. Precipitation patterns (on this wet day) conform to expectations, exhibiting smooth transitions that depict precipitation over mountainous terrain, flat landscapes, and regions with zero precipitation. Precipitation patterns also conform to expectations in monthly and annual time frames. The same conformity is observed for temperature, where transitions from colder to hotter areas are smooth and realistic. In addition, the overall patterns of climate change in space, time, and magnitude mirror changes from the original GCMs (Figs. 14, 15). These are similar to changes observed in other downscaled datasets, such as the LOCA2 47( https://crt-climate-explorer.nemac.org/).

Spatial climatic patterns for a day, a month, and a year of pr and tasmax downscaled from the EC-Earth3 model.

Change from the historical period (1985–2014) to the end of the century (2071–2100) for annual average of precipitation (pr), daily maximum (tasmax), and minimum (tasmin) air temperature. The analysis encompasses the average change from all 17 selected models under the Shared Socioeconomic Pathway SSP585.

Climate projections for annual precipitation (pr) and the annual average of daily maximum (tasmax) and minimum (tasmax) air temperatures under three Shared Socioeconomic Pathways (SSP245, SSP370, and SSP585) in Logan, Utah. The solid lines depict the average projections derived from the seventeen models, while the shaded areas signify the range between the smallest and largest projected values per year among these models.

Figure 16 illustrates the differences in extreme events from ClimAVA-SW 28 and LOCA2 47 for the EC-Earth3 GCM (r1i1p1f1) under the SSP585 scenario by the end of the century (1970–2100). At the 95th percentile, extreme precipitation events show virtually identical results between LOCA2

Differences in daily precipitation (pr) and maximum (tasmax) and minimum (tasmax) air temperature events from ClimAVA-SW 28 and LOCA2 47 for the EC-Earth3 GCM (r1i1p1f1) under the SSP585 scenario by the end of the century (1970–2100) at different percentiles. The analysis is done at the LOCA2 28.

As ClimAVA and LOCA2 47 use the same bias correction method for temperatures, the observed differences could mainly stem from two causes: the downscaling method itself or the reference datasets used. To determine if the differences arise from the downscaling methods, the methods should be compared using the same dataset and bias correction scheme. However, we hypothesize that the differences primarily stem from the reference datasets. ClimAVA uses PRISM data from 1981 to 2020, while LOCA2 48, an adaptation from Livneh et al. 49, covering the period from 1915 to 2018. As previously stated, the larger the reference dataset, the better the representation of extreme events, as it provides more variability to sample from. However, it is crucial to understand how the reference data can affect downscaling extreme events from both LOCA and SPID methods.

Attention has been given to comparing precipitation datasets 24 and developing extreme-preserving long-term gridded daily precipitation datasets 48. We hypothesize that PRISM and Pierce et al. 37, and the most significant events might be present in the 1981–2020 period covered by both datasets, explaining why extreme precipitation events look similar in both downscaled products. However, a thorough examination of extreme temperatures in gridded datasets is warranted, particularly for the first half of the 20th century when station densities were significantly lower 49. This issue is especially pertinent in mountainous areas, where larger differences were found. These studies will be essential to clarify if non-climatic extreme temperature events are present in the reference datasets and how these features can influence the results of extreme events in downscaled products.

As noted, downscaling methods like SPID and LOCA have difficulty downscaling values that exceed those present in their reference data. Therefore, care should be taken when analyzing the magnitude of future extreme record events. However, examining how frequently extreme events present in the reference data might occur in the future under different scenarios is highly recommended.

Code availability

All code necessary to reproduce the results is available at https://github.com/ClimAVA/ClimAVA-Project.

1.
Jay, A. K. et al. Fifth National Climate Assessment. Fifth National Climate Assessment.
https://doi.org/10.7930/NCA5.2023(U.S. Global Change Research Program, Washington, DC, USA, 2023).
2.
Kotz, M., Levermann, A. & Wenz, L. The economic commitment of climate change. Nature 628, 551–557 (2024).
3.
Kousky, C. Informing climate adaptation: A review of the economic costs of natural disasters. Energy Econ. 46, 576–592 (2014).
4.
Maurer, E. P., Hidalgo, H. G., Das, T., Dettinger, M. D. & Cayan, D. R. The utility of daily large-scale climate data in the assessment of climate change impacts on daily streamflow in California. Hydrol. Earth Syst. Sci. 14, 1125–1138 (2010).
5.
Fowler, H. J., Blenkinsop, S. & Tebaldi, C. Linking climate change modelling to impacts studies: Recent advances in downscaling techniques for hydrological modelling. International Journal of Climatology 27, 1547–1578,
https://doi.org/10.1002/joc.1556(2007).
6.
Harris, R. M. B. et al. Climate projections for ecologists. Wiley Interdisciplinary Reviews: Climate Change 5, 621–637,
https://doi.org/10.1002/wcc.291(2014).
7.
Ekström, M., Grose, M., Heady, C., Turner, S. & Teng, J. The method of producing climate change datasets impacts the resulting policy guidance and chance of mal-adaptation. Clim. Serv. 4, 13–29 (2016).
8.
Christensen, N. S. & Lettenmaier, D. P. A multimodel ensemble approach to assessment of climate change impacts on the hydrology and water resources of the Colorado River Basin. Hydrol. Earth Syst. Sci. 11, 1417–1434 (2007).
9.
Alder, J. R. & Hostetler, S. W. The Dependence of Hydroclimate Projections in Snow-Dominated Regions of the Western United States on the Choice of Statistically Downscaled Climate Data. Water Resour. Res. 55, 2279–2300 (2019).
10.
Keller, A. A., Garner, K. L., Rao, N., Knipping, E. & Thomas, J. Downscaling approaches of climate change projections for watershed modeling: Review of theoretical and practical considerations. PLOS Water 1, e0000046 (2022).
11.
Eum, H. II, Gupta, A. & Dibike, Y. Effects of univariate and multivariate statistical downscaling methods on climatic and hydrologic indicators for Alberta, Canada. J. Hydrol. 588, 125065 (2020).
12.
Attique, R., Rientjes, T. & Booij, M. Comparison between statistical and dynamical downscaling of rainfall over the Gwadar-Ormara basin. Pakistan. Meteorol. Appl. 30, 1–19 (2023).
13.
Wood, A. W., Leung, L. R., Sridhar, V. & Lettenmaier, D. P. Hydrologic implications of dynamical and statistical approaches to downscaling climate model outputs. Clim. Change 62, 189–216 (2004).
14.
Lanzante, J. R., Dixon, K. W., Nath, M. J., Whitlock, C. E. & Adams-Smith, D. Some pitfalls in statistical downscaling of future climate. Bull. Am. Meteorol. Soc. 99, 791–803 (2018).
15.
Maurer, E. P. Uncertainty in hydrologic impacts of climate change in the Sierra Nevada, California, under two emissions scenarios. Clim. Change 82, 309–325 (2007).
16.
Payne, J. T., Wood, A. W., Hamlet, A. F., Palmer, R. N. & Lettenmaier, D. P. Mitigating the effects of climate change on the water resources of the Columbia River Basin. Clim. Change 62, 233–256 (2004).
17.
Wood, A. W., Maurer, E. P., Kumar, A. & Lettenmaier, D. P. Long-range experimental hydrologic forecasting for the eastern United States. J. Geophys. Res. Atmos. 107, ACL 6-1–ACL 6-15 (2002).
18.
Byun, K. & Hamlet, A. F. Projected changes in future climate over the Midwest and Great Lakes region using downscaled CMIP5 ensembles. Int. J. Climatol. 38, e531–e553 (2018).
19.
Pierce, D. W., Cayan, D. R. & Thrasher, B. L. Statistical downscaling using localized constructed analogs (LOCA). J. Hydrometeorol. 15, 2558–2585 (2014).
20.
Abatzoglou, J. T. & Brown, T. J. A comparison of statistical downscaling methods suited for wildfire applications. Int. J. Climatol. 32, 772–780 (2012).
21.
Hernanz, A. et al. Evaluation of statistical downscaling methods for climate change projections over Spain: Present conditions with perfect predictors. Int. J. Climatol. 42, 762–776 (2022).
22.
Ekström, M., Grose, M. R. & Whetton, P. H. An appraisal of downscaling methods used in climate change research. Wiley Interdisciplinary Reviews: Climate Change 6, 301–319,
https://doi.org/10.1002/wcc.339(2015).
23.
Daly, C. et al. Physiographically sensitive mapping of climatological temperature and precipitation across the conterminous United States. Int. J. Climatol. 28, 2031–2064 (2008).
24.
Henn, B., Newman, A. J., Livneh, B., Daly, C. & Lundquist, J. D. An assessment of differences in gridded precipitation datasets in complex terrain. J. Hydrol. 556, 1205–1219 (2018).
25.
Behnke, R. et al. Evaluation of downscaled, gridded climate data for the conterminous United States. Ecol. Appl. 26, 1338–1351 (2016).
26.
Daly, C. & Bryant, K. The PRISM Climate and Weather System - An Introduction. PRISM Climate Group
http://www.prism.oregonstate.edu/documents/PRISM_history_jun2013.pdf(20….
27.
Eyring, V. et al. Overview of the Coupled Model Intercomparison Project Phase 6 (CMIP6) experimental design and organization. Geosci. Model Dev. 9, 1937–1958 (2016).
28.
de Lima Moraes, Andre Geraldo; Khoshnood Motlagh, S. Climate data for adaptation and vulnerability assessments — southwest. at
https://doi.org/10.26050/WDCC/ClimAVA-SW(2024).
29.
Panofsky, H. W. & Brier, G. W. Some Applications of Statistics to Meteorology. The Pennsylvania State University Press (Philadelphia, 1953).
30.
Li, H., Sheffield, J. & Wood, E. F. Bias correction of monthly precipitation and temperature fields from Intergovernmental Panel on Climate Change AR4 models using equidistant quantile matching. J. Geophys. Res. Atmos. 115 (2010).
31.
Hagemann, S. et al. Impact of a statistical bias correction on the projected hydrological changes obtained from three GCMs and two hydrology models. J. Hydrometeorol. 12, 556–578 (2011).
32.
Pierce, D. W. et al. Probabilistic estimates of future changes in California temperature and precipitation using statistical and dynamical downscaling. Clim. Dyn. 40, 839–856 (2013).
33.
Maurer, E. P. & Pierce, D. W. Bias correction can modify climate model simulated precipitation changes without adverse effect on the ensemble mean. Hydrol. Earth Syst. Sci. 18, 915–925 (2014).
34.
Pierce, D. W., Cayan, D. R., Maurer, E. P., Abatzoglou, J. T. & Hegewisch, K. C. Improved bias correction techniques for hydrological simulations of climate change. J. Hydrometeorol. 16, 2421–2442 (2015).
35.
Liaw, A. & Wiener, M. Classification and Regression by randomForest. R News 2, 18–22 (2002).
36.
Ziegler, A. & König, I. R. Mining data with random forests: Current options for real-world applications. Wiley Interdiscip. Rev. Data Min. Knowl. Discov. 4, 55–63 (2014).
37.
Donat, M. G., Lowry, A. L., Alexander, L. V., O’Gorman, P. A. & Maher, N. More extreme precipitation in the world’s dry and wet regions. Nat. Clim. Chang. 6, 508–513 (2016).
38.
Habeeb, D., Vargo, J. & Stone, B. Rising heat wave trends in large US cities. Nat. Hazards 76, 1651–1665 (2015).
39.
Barnett, A. G., Hajat, S., Gasparrini, A. & Rocklöv, J. Cold and heat waves in the United States. Environ. Res. 112, 218–224 (2012).
40.
Hodgkins, G. A., Dudley, R. W., Archfield, S. A. & Renard, B. Effects of climate, regulation, and urbanization on historical flood trends in the United States. J. Hydrol. 573, 697–709 (2019).
41.
Stillman, J. H. Heat waves, the new normal: Summertime temperature extremes will impact animals, ecosystems, and human communities. Physiology 34, 86–100,
https://doi.org/10.1152/physiol.00040.2018(2019).
42.
Chen, C. T. & Knutson, T. On the verification and comparison of extreme rainfall indices from climate models. J. Clim. 21, 1605–1621 (2008).
43.
Wehner, M. F., Smith, R. L., Bala, G. & Duffy, P. The effect of horizontal resolution on simulation of very extreme US precipitation events in a global atmosphere model. Clim. Dyn. 34, 241–247 (2010).
44.
Polade, S. D., Pierce, D. W., Cayan, D. R., Gershunov, A. & Dettinger, M. D. The key role of dry days in changing regional climate and precipitation regimes. Sci. Rep. 4 (2014).
45.
Caldwell, P., Chin, H. N. S., Bader, D. C. & Bala, G. Evaluation of a WRF dynamical downscaling simulation over California. Clim. Change 95, 499–521 (2009).
46.
Intergovernmental Panel on Climate Change. Water Cycle Changes. Climate Change 2021 – The Physical Science Basis.
https://doi.org/10.1017/9781009157896.010(2023).
47.
Pierce, D. W., Cayan, D. R., Feldman, D. R. & Risser, M. D. Future Increases in North American Extreme Precipitation in CMIP6 Downscaled with LOCA. J. Hydrometeorol. 24, 951–975 (2023).
48.
Pierce, D. W. et al. An Extreme-Preserving Long-Term Gridded Daily Precipitation Dataset for the Conterminous United States. J. Hydrometeorol. 22, 1883–1895 (2021).
49.
Livneh, B. et al. A spatially comprehensive, hydrometeorological data set for Mexico, the U.S., and Southern Canada 1950–2013. Sci. Data 2, 1–12 (2015).

Acknowledgements

This research was supported by the Utah Agricultural Experiment Station, Utah State University, and approved as journal paper number 9779, the Utah Climate Center, and the Center for High-Performance Computing at the University of Utah.

Ethics declarations

Competing interests

The authors declare no competing interests.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary information

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

About this article

Cite this article

de Lima Moraes, A.G., Khoshnood Motlagh, S. The Climate Data for Adaptation and Vulnerability Assessments and the Spatial Interactions Downscaling Method. Sci Data 11, 1157 (2024). https://doi.org/10.1038/s41597-024-03995-6

-
Received:
-
Accepted:
-
Published:
-
DOI: https://doi.org/10.1038/s41597-024-03995-6