1. Introduction
Global climate models (GCMs) typically have horizontal resolutions of 100–300 km (Jalota et al. 2018; Hannah 2015). However, to retrieve practical information on climate impacts, model outputs must be available at higher spatial resolution. To achieve this, different downscaling techniques have been developed. These can be grouped into two main categories.
Statistical downscaling (SD) techniques develop statistical relationships between large-scale GCM outputs and high-resolution observational data to later apply the statistical model on future simulations. These methods rely on the assumption that the statistical relationship developed between the large-scale predictors and high-resolution predictands holds true in future scenarios (Wilby et al. 2004). This assumption is also the main disadvantage of SD methods, as it is virtually impossible to verify. Furthermore, physical relationships between weather variables linked to mesoscale dynamics or synoptic conditions are lost in these methods. However, SD methods remain useful in providing site-specific data for impact assessments due to their efficient run times and low computational demand (Lange 2020; Nourani et al. 2019; Sachindra et al. 2018; Vanderkelen et al. 2018).
Dynamic downscaling refers to the use of regional climate models (RCMs), which employ the atmosphere simulated by GCMs as boundary conditions to physically model regional climate systems (Xu et al. 2019). Consequently, unlike SD methods, RCMs are more robust in representing changing conditions such as mesoscale flow, large-scale circulation and land–atmosphere interactions (Van de Walle et al. 2020, 2021; Thiery et al. 2015, 2016; Souverijns et al. 2016). Recent studies have enabled RCMs to produce outputs as fine as 1 km (Brousse et al. 2020). The ability of RCMs to capture local climate dynamics and factors that control them, such as topography, also makes them computationally demanding (Zhang et al. 2020). Furthermore, RCMs often do not sample uncertainty associated with parent GCMs, scenario, and modes of natural variability well. Therefore, they often require statistical postprocessing once their outputs are generated (Sangelantoni et al. 2019).
One of the factors influencing the climate is the effect of land–atmosphere interactions stemming from land-use and land-cover change (LULCC), which can be categorized as biogeophysical or biogeochemical (Lawrence et al. 2016). Biogeochemical effects result mainly from changes in carbon and nitrogen fluxes. Carbon emissions from LULCC are estimated to represent 25% of cumulative historic emissions between 1750 and 2011, and 12.5% in the years 1990–2010 (Ciais et al. 2014). The biogeophysical effects of LULCC represent perturbations in water and energy cycles, thereby changing radiative forcing and the surface energy balance. These changes can be caused by processes that adjust sensible and latent heat fluxes, such as changes in albedo, evaporative fraction, and surface roughness (Davin and de Noblet-Ducoudré 2010). Depending on the region, the biogeophysical impact of LULCC can mask or amplify climate change impacts from greenhouse gas emissions (Davin and de Noblet-Ducoudré 2010; Thiery et al. 2020). However, the representation of LULCC climate impacts in GCMs is strongly inconsistent among models (de Noblet-Ducoudré et al. 2012; Lejeune et al. 2020). This was also evident in studies that showed the response of near-surface temperature to idealized deforestation or re/af-forestation (Boysen et al. 2020; Davin et al. 2020). The results showed considerable variation among models with some showing cooling in all regions and others showing a combination of cooling and warming in different regions. The difference among models came from model representation of land-use changes and corresponding land–atmosphere interactions such as energy flux decomposition that affect sensitivity of temperature to land-cover changes (Boysen et al. 2020).
To incorporate the effects of LULCC on climate, GCMs typically use tile-level setups for different land-cover types within each GCM grid cell in their land model components. These tiles represent fractional areas of different land-cover types that each interact with the atmosphere individually (Lawrence et al. 2016). Although there is no interaction between tiles of the same grid cell, tile-level outputs in some contexts give a stronger signal than gridcell averages, especially in studies involving surface temperature and temperature extremes (Thiery et al. 2017; Schultz et al. 2016; Malyshev et al. 2015; Fischer et al. 2012; Hirsch et al. 2018). However, until phase 6 of the Coupled Model Intercomparison Project (CMIP6), model intercomparison projects did not store tile-level output, which inhibited exploitation of this information.
Despite the established importance of land cover, its application in downscaling techniques is often absent. The Coordinated Regional Climate Downscaling Experiment (CORDEX) is an extension of CMIP, where RCMs contribute to a continuous generation of dynamic downscaling simulations under a shared framework (Gutowski et al. 2016). The most recent CORDEX experimental framework for dynamic downscaling only recommends participating RCMs use their default land-cover maps with no standardized dataset specified (CORDEX 2021). A pilot study named Land Use and Climate Across Scales (LUCAS) under the CORDEX initiative is currently investigating the biogeophysical effects of intense land-cover changes and sensitivity of RCMs to local land-cover changes (Rechid et al. 2017; Davin et al. 2020). However, it remains the case that land-cover change is not addressed well in either dynamic or statistical downscaling methods.
Here, we develop and test a new downscaling method for surface variables that builds on tile-level GCM outputs, hereinafter referred to as land-cover tile downscaling (LTD). LTD incorporates land-use responses to atmospheric forcing and their effect on surface fluxes and land state variables by estimating higher-resolution data based on tile-level outputs. It subsequently incorporates topographical data by applying an elevation-based correction to the output. As its basis lies in redistributing the simulations of parent GCMs, the new method therefore produces high-resolution data that are more physically grounded than the purely statistical downscaling methods. The only statistical assumption made in LTD is the relationship between temperature and elevation that is used to apply correction on the downscaled data. The method is also less expensive than RCMs as it consists of simple algorithms that do not involve solving numerical equations.
Since LTD is based on land-cover tiles, it is only expected to work well with a limited number of variables that are controlled by local land-cover characteristics. Furthermore, its performance strongly relies on the suitability of the tile-level information to the variable being downscaled. Studies have shown that variables such as surface/near-surface temperature and surface energy fluxes are regionally controlled by land-cover characteristics (Davin and Noblet-Ducoudré 2010; Zhao et al. 2019). Therefore, in this study, the method is applied on near-surface air temperature and surface temperature.
2. Data and methods
a. Datasets
In this research, only the historical climate is considered for the downscaling approach, since the downscaled data must be evaluated against observational data. The CMIP6 historical scenario spans 1850–2014 with natural and anthropogenic forcing data applied based on observations (Eyring et al. 2016). Models participating in the CMIP6 historical experiment are validated against observed data. Currently, there are only two models that provide tile-level air temperature outputs for the historical scenario with all necessary parameters (Table 1).
Models that provide land-use tile near-surface air temperature under CMIP6-endorsed MIPs’ Historical scenario. Tile numbers 0, 1, 2, and 3 represent primary and secondary land, pasture, crops, and urban tiles, respectively.
The models that provide at least three of the requested tiles (except pasture) and within the time frame of the available observation data are the Community Earth System Model, version 2 (CESM2), and U.K. Earth System Model, version 1, low resolution (UKESM1-0-LL; herein referred to as UKESM1) (Table 1). The LL in UKESM1-0-LL describes the resolution of the model, in this case, low resolution. Model evaluations of CESM2 near-surface temperature show that observations generally fall inside the ensemble spread; however, the ensemble mean shows a warm bias in the years 2000–14 (Danabasoglu et al. 2020). On the other hand, UKESM1 is characterized by a cold bias, especially in the Northern Hemisphere, after 1950 where observational data lie outside the ensemble spread (Sellar et al. 2019).
The land component of CESM2, the Community Land Model (CLM), uses glacier, lake, urban, vegetated, and crop land units with the vegetated and crop classes further divided based on different plant functional types (PFT) and crop functional types (CFT) (Lawrence et al. 2019). It is possible to define up to 16 PFTs for vegetated land unit while 16 active CFTs and 46 inactive CFTs (parameterized using the active CFTs) are available under crop land unit (Lawrence et al. 2019). Similarly, the land component of UKESM1, Joint U.K. Land Environment Simulator (JULES), uses 5 PFTs along with urban, inland water, bare soil, and land-ice tiles (Clark et al. 2011).
The tile-level data are provided as layers of fields on each grid cell with no specific information on where each land-cover type is located within the grid. Each tile contains a single value in each grid that represents the fraction of area covered by the corresponding land-cover type. Therefore, the average of all tiles weighted by their corresponding fractions gives the grid average data. Although GCMs make use of several land-cover tiles in implementation, CMIP6 requests participating models to provide outputs for four general tiles, namely, primary and secondary land, pastureland, crops, and urban tiles (Lawrence et al. 2016). Primary land includes areas that have not been affected by human activity, while secondary land is affected by humans but with potential to regrow (Lawrence et al. 2016). In combination, primary and secondary land contain forests, grasslands, and bare ground. Pastureland includes managed pasture and rangelands.
For each model, the tile output of a single realization is downscaled and compared with the gridcell average output of the same realization. Hence, the realization of natural variability in the individual ensemble member is not expected to affect this comparison. However, the overall agreement of the downscaled data with observed data can be affected based on which realization is being considered.
For surface temperature, similarly, a single realization of tile-level data that incorporates the same three tiles (except pasture) is downscaled. However, only CESM2 is used because the two other models that provide the variable at the tile level each have two missing land-cover tiles (Table 2).
1) Observational data
A global dataset of near-surface air temperature is used to evaluate the downscaling technique (Hooker et al. 2018). The dataset was produced by combining station data from the Global Historical Climatology Network–Monthly (GHCN-M) and satellite data from the Moderate Resolution Imaging Spectroradiometer (MODIS) aboard the Aqua satellite. This was done using a statistical model to regress and combine the two data sources to provide monthly 2-m air temperature from 2003 to 2016 at a spatial resolution of 0.05°.
For surface temperature, the data used as observational reference are the MODIS land surface temperature data (LST) (downloaded from https://lpdaac.usgs.gov/products/mod11c3v006/). The MODIS data are available from 1-km spatial resolution to 0.05° at daily and monthly time scales. In this study, the monthly data at a spatial resolution of 0.05° aboard the Terra satellite (MOD11C3) are used to homogenize the resolution with the near-surface air temperature data.
2) Land-cover map
The European Space Agency (ESA), under the Climate Change Initiative (CCI), provides high-resolution land-cover maps. The maps were generated using satellite data from the Advanced Very High-Resolution Radiometer (AVHRR) and Project for On-Board Autonomy–Vegetation (PROBA-V) and provide yearly land-cover maps at 300-m resolution (ESA 2017). In this analysis, maps from 2003–14 are used to be consistent with the available observation and simulation data.
3) Elevation map
The Global Multiresolution Terrain Elevation Data 2010 (GMTED2010) dataset provides global coverage of elevation at 250-, 500-, and 1000-m resolutions over all land areas by combining data from the Shuttle Radar Topography Mission (SRTM) with other data sources to complement the coverage and voids in SRTM (Danielson and Gesch 2011). In this study, the lowest-resolution dataset at 1 km is used since the resolution at which the downscaling is performed is coarser.
b. Methods
In this section, the implementation of each step involved in LTD is explained in detail and in accordance with the accompanying flowchart (Fig. 1). The method is implemented using five steps that are seen numbered in the figure and described in the text that follows.
Methods procedure flowchart: tas* is the air temperature of a grid cell, and tasLut* is the air temperature of a land-use tile; Δh is the difference between subgrid and grid elevation (see step 4).
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-21-0265.1
1) Step 1: Reclassifying land-cover data
ESA-CCI land-cover maps have a relatively large number of land-cover classes (37) in comparison with the CMIP6-requested tile-level outputs that provides only four classes: primary and secondary land, pasture, crops, and urban. Therefore, the ESA-CCI maps need to be reclassified to correspond to the four tile classes. Since ESA-CCI does not distinguish between managed pastures and rangelands from other land-cover types, the number of target classes is reduced to three (primary and secondary land, crops, and urban), and the pasture tile output from the GCMs is omitted form the analysis. Consequently, the ESA-CCI land-cover classes are reclassified to the three new tiles by binning land-cover types such as different categories of trees and bare lands under primary and secondary land, and different crop types under the general class of crops (as illustrated in Table 3). Accordingly, different types of crops (rainfed, irrigated, or mosaic) were categorized under crops, urban land was taken as the only urban tile and all the remaining land-cover types including different types of trees, shrublands, and bare ground are classified under primary and secondary land following the specification described in (Lawrence et al. 2016).
ESA-CCI land classes and corresponding reclassified classes. NaN values are attributes used to represent no data values (ESA 2017).
2) Step 2: Regridding
We perform all computations and downscaling evaluation at the scale of the temperature observations (0.05°). To this end, we regrid the data to interpolate or aggregate it to the target resolutions. In normal cases, regridding would add some information since it involves spatial interpolation within the grid cells. This is mainly based on the assumption of homogenous spatial gradients within each grid. As the extent of regridding gets higher, such as in this study (regridding from >1° to 0.05°), the assumption becomes less and less valid as fine-scale processes become more important but are not captured by the interpolation technique. Therefore, we expect little additional information from this step, and the regridded data would carry most of the biases of the raw data. This step is only performed to bring all datasets to the same resolution before performing subsequent operations.
(i) GCM output
The GCM grid-scale temperature output is regridded to the observation grid using bilinear interpolation. This method is selected since bilinear interpolation fits the smooth, spatially correlated behavior of the variable (Shea 2014).
(ii) Land cover
Land-cover maps are coarsened to the observation resolution by calculating land area fractions in each target grid cell from the original contained grid cells. This method stores the fraction of each land-cover type inside the new grid scale as layers of data. For instance, if 50% of the target grid cell is primary land, 20% cropland, and 30% urban, then these data are recorded by creating coordinates for each land-cover type and saving the calculated land area fraction of each class in the corresponding coordinate.
(iii) Elevation
Elevation maps are regridded in two ways to create elevation data corresponding to the GCM grid scale and the observation grid scale. The resulting data corresponding to the GCM grid scale are hereinafter referred to as grid-scale elevation, whereas those corresponding to observation grid scale are referred to as subgrid-scale elevation.
As coarsening of elevation data requires that all pixels that are considered reside inside the target grid, conservative regridding is preferred over bilinear interpolation, which only considers a few neighboring pixels (Jones 1998). Conservative regridding is performed by calculating weights based on area fraction overlap between the source and target grids. The level of accuracy of conservative remapping can be enhanced by employing a second-order regridding accounting for gradients of the variable (Jones 1999). For this reason, the remapping on elevation data is performed using second-order conservative remapping.
3) Step 3: Mapping tile-level GCM output to land cover
4) Step 4: Elevation correction
The subgrid-scale and grid-scale elevation datasets produced in step 2 are used to correct the variables mapped in step 3 by computing local lapse rates. Recent research in using local lapse rates to downscale GCM outputs has proven skillful when compared with observational data (Praskievicz 2018).
Here, a simplified form of the local lapse rate downscaling method is used to correct the output of the mapping performed in step 3. The original method performs a singular value decomposition regression between the variable in question (minimum/maximum temperature or precipitation) and location parameters (longitude, latitude, and elevation) within a defined window for each pixel (Praskievicz 2018). To maximize the fitness of the trend, the method first iterates different window sizes and arrives at an optimum size for each variable.
The lapse rate produced through linear regression is then used to generate a correction factor. The grid-scale elevation corresponds to an average elevation that represents the original GCM grid cell. Therefore, the difference between the grid-scale elevation and the subgrid-scale elevation is taken to represent the deviation of each point inside a GCM grid cell from the average elevation of the GCM grid cell itself.
5) Step 5: Evaluation
There are several statistical metrics that are used to evaluate downscaling methods. In this project, the root-mean-squared error (RMSE) and Kling–Gupta efficiency (KGE) are used to compare the downscaled GCM data with the gridded observations. The metrics are generated based on the entire time series of the available observation data (2003–14). This is done with seasonal time series and zonal means to assess the downscaling performance across different scales.
3. Results
We aim to compare the results of the mapped data with the standard gridcell averages generated by the models. However, since the gridcell average and tile data are different outputs, we first reconstruct gridcell averages from the tile data and compare them with the gridcell averages generated by the models. The reconstruction was performed by taking weighted averages of the tile data where, for a given grid cell, the weights are the fraction of space allotted to each tile (fracLut). Here, grid cells where the sum of fracLut does not add up to at least 0.95 are neglected. The data from UKESM1 become sparse after grid cells with small fracLut have been removed. This may be because UKESM1 does not include bare soil in its land-cover tile definition (E. Robertson 2022, personal communication). This in part explains the disagreement since most of the areas with missing data are dominated by bare lands such as the Sahara, and large parts of Australia and Asia (Fig. 2).
Mean differences between reconstructed and acquired grid average air temperature datasets based on (a) CESM2 and (b) UKESM1 output.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-21-0265.1
The difference between the reconstructed and the standard gridcell average air temperature data are provided in Fig. 2. The results indicate that the bias between the two datasets is not negligible. This deviation could come from the use of only four tiles that force the models to aggregate their original land-use classification. Especially, in UKESM1, the tile-level data do not encompass all land-cover classes used in the model configurations, which is evident in the sum of fracLut values being less than 1 in several parts of the world. Surface temperature results from CESM2 show bias differences consistent with air temperature data (Fig. B1 of appendix B). We will not assess these differences in this study since our main goal is not testing the validity of tile-level data. However, these results are still important for future work analyzing the implementation of tile-level outputs in GCMs.
Since the main aim of this article is to assess the added value of land-cover mapping, we use the reconstructed grid averages as the control data (CTL). The data after the mapping step (referred to as MAP) and the final downscaled data after both mapping and correction are applied (referred to as MAP_COR) are compared with CTL. We also analyze the effect of elevation correction by comparing CTL with the data after only elevation correction is applied (referred to as COR).
The RMSE and KGE metrics of CTL from CESM2 air temperature are provided in Fig. 3. CESM2 has lower RMSE (better skill) in the equatorial region, but RMSE increases with increasing latitude. The opposite is true in KGE where equatorial regions show lower values (less skill) than higher-latitude regions. This indicates that the model reduces bias in equatorial regions but does not improve variability and correlation with the observed data. This is also true in surface temperature (Fig. B3 of appendix B) and grid average outputs of UKESM1 (Fig. B2 of appendix B). However, RMSE scores for surface temperature show higher bias in most parts of the globe when compared with RMSE for air temperature (Fig. B3 of appendix B). It is important to note that these results are based on a single realization of each model. Therefore, the results should not be used to compare the performance of the models with each other because internal variability can play a role in how well they perform. The results should only be viewed as a reference for each model before LTD is applied.
(a) RMSE and (b) KGE scores for air temperature data based on CESM2 output.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-21-0265.1
The performance of LTD applied on CESM2 varies spatially and between the two variables considered. Differences (MAP-CTL) in both RMSE and KGE are more pronounced in maps from surface temperature than air temperature (Fig. 4). Results from air temperature show improvement in certain parts of the globe with only a few areas showing decline in skill (Figs. 4a,c). However, results from surface temperature are more mixed where areas showing improvement and decline are almost comparable in size, especially in RMSE (Figs. 4b,d). In both cases, RMSE metrics show more widespread and stronger signal than KGE. Although there is only limited data left in UKESM1, the results show consistent increase in skill in both metrics except for very few areas that show decline in KGE (Fig. 5).
Difference in (a),(b) RMSE and (c),(d) KGE between MAP and CTL (left) air temperature and (right) surface temperature data based on CESM2.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-21-0265.1
Difference in (a) RMSE and (b) KGE between MAP and CTL air temperature data based on UKESM1.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-21-0265.1
In CESM2, there is a consistent decline of skill in both air and surface temperature RMSE scores (Figs. 4a,b). The regions displaying this decline are consistent with areas that changed into crops and pastureland since preindustrial periods (Fig. 1; de Noblet-Ducoudré et al. 2012). Therefore, the temperature signal of land-use changes may not appear in the corresponding tile depending on the model’s sensitivity to LULCC. This is supported by a CMIP5 study that showed the temperature response to deforestation has strong signal in some models while it barely appears in others (Winckler et al. 2019).
The effect of elevation correction (COR-CTL) on RMSE scores is strongly concentrated in highland areas in CESM2 (Figs. 6a,b), for example in the Andes, the Rocky Mountains, the Himalayas, and the Ethiopian highlands (Fig. B4 of appendix B). This is to be expected as spatial temperature patterns are controlled to a higher extent by elevation than by land cover in mountainous regions (Praskievicz 2018). The results in KGE show improvement to a limited extent in different parts of the world (Figs. 6c,d) except in South America, specifically the Amazon region, where stark increase and decrease in KGE are observed in air and surface temperature results, respectively. Results from UKESM1 show similar results, where data are available, with the increase in RMSE concentrated in the Andes and the anomalous decline in air temperature KGE in the Amazon is also observed (Figs. 7a,b). The Amazon region, although oddly specific in both models for its distinct KGE results, does not contain any unique elevation characteristics (Fig. B4 of appendix B).
Difference in (a),(b) RMSE and (c),(d) KGE between COR and CTL (left) air temperature and (right) surface temperature data based on CESM2.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-21-0265.1
Difference in (a) RMSE and (b) KGE between COR and CTL air temperature data based on UKESM1.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-21-0265.1
The overall change in metrics in CESM2 shows that both RMSE and KGE scores improved as a result of mapping (Fig. 8). However, the boxplots show little improvement due to elevation correction. KGE scores show little change in median values, but improvements can be seen in the lower quantiles. In UKESM1, results show changes similar to CESM2 although the improvements in both metrics are more pronounced than results from CESM2 (Fig. 9). The median RMSE values of MAP_COR shows a reduction from CTL by 1.4 in CESM2 and by 11.6 in UKESM1. Although there is significant improvement in both metrics in UKESM1, it is important to reiterate the fact that this is based on very limited areas in the globe since areas of incomplete data were removed earlier. Results from surface temperature (from CESM2) show similar behavior as air temperature results. However, the improvement in RMSE is more pronounced with the median RMSE decreasing by a value of 12 (Fig. 10). Interestingly, although the global plots indicated mixed results, surface temperature shows stronger improvement in RMSE than air temperature data.
(a) RMSE score and (b) KGE score distribution of CTL, MAP, and MAP_COR of air temperature data based on CESM2. The boxplots show the distribution of RMSE or KGE aggregated from all regions (center line: median; box limits: upper and lower quartiles; green triangle: mean) with whiskers extending to the last value located within a distance of 1.5 times the interquartile range from the 25th and 75th quantile, respectively. Outliers are not shown. Median values are shown at the top corner of each distribution.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-21-0265.1
As in Fig. 8, but based on UKESM1.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-21-0265.1
As in Fig. 8, but for surface temperature data.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-21-0265.1
The change in RMSE due to LTD shows improvements in pixels dominated by primary and secondary land and crops in both models (Figs. 11 and 12). KGE scores also show changes consistent with the RMSE scores but are limited to lower quantile regions in CESM2. Areas dominated by urban land show decline in skill in upper quantiles for CESM2, although the median values show improvements and the overall distribution in MAP_COR show improvements (Fig. 11). Urban tiles in UKESM1 show a decline of skill in RMSE and a minimal change in KGE (Fig. 12). Results from surface temperature data are similar, however, here all land-cover tiles show consistent improvement in both metrics (Fig. 13). The decline of skill in urban tiles for air temperature, especially in UKESM1, could arise from possible mis-mapping of tiles in LTD; however, it is highly unlikely that this could happen only for urban tiles, and specifically for air temperature in UKESM1 only. If we eliminate the mapping process as a source of error, it implies that the grid average data better agree with observations than urban tile data. This can only happen if the tile data of urban land in UKESM1 contain substantial biases from observations that are masked when averaged together with other tiles. Seasonal analysis of LTD shows very similar trends to the overall metrics presented here indicating the method’s independence on seasonal variations (see appendix A).
(top) RMSE score and (bottom) KGE score distribution of CTL, MAP, and MAP_COR of air temperature data based on CESM2 for pixels dominated by (left) primary and secondary land, (center) crops, and (right) urban. All boxplots follow the format described in Fig. 8.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-21-0265.1
As in Fig. 8, but based on UKESM1.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-21-0265.1
As in Fig. 8, but for surface temperature.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-21-0265.1
4. Discussion
a. Limitations in land-cover mapping
Limitations in GCM land-cover classes and external land-cover datasets used to develop LTD likely contribute to its mixed performance. First, tile-level outputs are provided for four tiles as per the standard LUMIP requirement and in this method, we have further reduced that to three, excluding pasture (primary and secondary land, crop, and urban), owing to the available land-cover types in the ESA-CCI land-cover map. Accordingly, the large variety of land-cover classes provided by the ESA-CCI land-cover map is reclassified under three broad classes of primary and secondary land, crop, and urban land. This leaves out important land-cover types that dictate land–atmosphere interactions, leading to possible errors. The requested tiles themselves also merge land-cover types that deviate strongly among each other. This, for instance, is evident in primary and secondary land tile that includes both trees and bare soil under the same classification although their impact on near-surface or surface temperature and their interaction with the atmosphere in general deviates strongly.
The annual time resolution of land-cover maps in the GCMs as well as ESA-CCI precludes seasonal variations in land cover like multicropping or slash-and-burn agriculture, which, in reality, may affect the observations we evaluate against. CLM uses the Land Use Harmonized Dataset 2 (LUH2), version 2f, annual dataset (Danabasoglu et al. 2020). Similarly, UKESM1 uses the harmonized History Database of the Global Environment (HYDE) annual dataset (Burton et al. 2019). This means that intra-annual variations in land cover are not accounted for both in the generation of the tile-level outputs and the mapping method implemented in this study—even though the models do account for seasonal variations in phenology within a given land-cover type. This can further contribute to realized errors as seasonal variations are not accounted for.
b. Limitations in elevation-based correction
The effect of the elevation correction is manifested especially over complex terrain through lowered biases while its effect on correlation and variability is not as strong. There are two simplifications in the elevation correction that could be contributing to the mixed signals observed in KGE. The first is that the window size is predefined to 1 unit of the GCM grid scale. The original method proposed by Praskievicz (2018) uses a moving window around each pixel with sizes defined based on an iterative approach to maximize the accuracy of the estimate. In the simplified method used here, the window size is not optimized, thus, areas that in reality do not affect the pixel could be considered to affect it, thereby potentially skewing the estimated lapse rate.
Additionally, the lapse rate is estimated for each window as a common value for all pixels residing in it. Therefore, the deviation in observed temperature between two windows that both affect a pixel (such as a pixel located near the border between two windows) that is only assumed to be affected by one of them could minimize the efficiency of the elevation correction step. Consequently, the resulting error would not manifest in the bias as the overall signal of an elevation correction would be in the right direction but variability and correlation over time would be affected, which could explain the mixed signals in KGE results. It is also worth noting that the elevation data were initially regridded from a very high resolution, which would also contribute to some of the observed errors. Additionally, other topographic features such as slopes that are known to affect temperature lapse rate (Tang and Fang 2006) were not considered in the estimation of local lapse rate here and may further contribute to the performance of the elevation correction.
c. Performance of LTD
LTD does not provide a consistent increase of skill through the land-cover mapping procedure. Although there is an overall improvement in the skill in both models for air temperature, the response of surface temperature is more mixed with certain regions showing decline of skill. The difference in the response of air temperature and surface temperature could be due to surface temperature being more sensitive to land-cover types. Surface temperatures are warmer than air temperatures for bare ground (which is relatively warmer than other land-cover types due to the dominance of sensible heat) and are colder in fully vegetated areas (which is relatively colder due to the dominance of latent heat) (Good et al. 2017). The broader range of surface temperature over different land-cover types implies that it is more sensitive to land-cover type and perturbations than air temperature. Several CMIP5 models showed that surface temperature responds more strongly to deforestation than air temperature in most midlatitude regions (Winckler et al. 2019). Therefore, the broad classification of land-cover types into only three tiles can have a stronger negative impact on surface temperature than it does on air temperature.
The effect of LTD is relatively more pronounced in RMSE and is a higher percentage of the original RMSE of CTL when compared with KGE, which shows very little change relative to CTL. Therefore, LTD could have the potential for reducing bias in GCM data, specifically for air temperature data where LTD showed clear improvements. This would especially be beneficial once the quality of tile-level data is assured.
5. Conclusions
In this study, an alternative downscaling method that makes use of tile-level outputs provided by phase 6 of the Coupled Model Intercomparison Project (CMIP6) global climate models (GCMs) is proposed. The method, named land-cover tile downscaling (LTD), maps tile-level variables to their respective land-cover classes in land-cover maps to provide downscaled data at 0.05° resolution. LTD is a relatively simple method that accounts for the impact of land cover on local climate. LTD minimizes the need for assumed statistical relationships that are the main disadvantages of statistical downscaling methods while it remains computationally inexpensive, unlike dynamical downscaling methods.
Here, the added value of LTD is analyzed by applying it to near surface air temperature and surface temperature and comparing the outputs with grid-averaged data, yielding the following conclusions:
-
The performance of LTD is strongly dependent on the variable considered; with air temperature showing improvement of skill in both models while land surface temperature shows mixed response.
-
The effect of LTD is more pronounced in RMSE results than in KGE implying its potential to reduce bias in GCM data.
-
Skill improvements due to LTD are concentrated in areas dominated by primary, secondary, and crop lands, while urban tiles show decline.
Overall, although there are clear signs of improved performance, especially in reducing bias, the observed signals are not consistent across different variables and tiles. The method tends to work better for air temperature data and tiles other than urban land. Given the fact that the method applies a consistent framework for all tiles and all variables, the results call for further research into the integration of land-cover changes in GCMs and the implementation of tile-level outputs.
This study was limited to only two models because other models in CMIP6 had incomplete data due to missing tiles and years. There is also a significant deviation between gridcell representations reconstructed from tile data and standard gridcell averages in both CESM2 and UKESM1. Although, this is to be expected in some areas where lakes, rivers, and glaciers play a significant role (since they are not included in the tile classification), the deviation was widespread in several parts of the world. This further underlines the need for a thorough evaluation of the tile-level datasets themselves to ensure their quality and remove inherent bias.
Based on what is observed in this study, we make the following recommendation for future work:
-
Modelers should give priority to quality controlling tile-level outputs with regard to their availability and accuracy. The accuracy of tile data can be assessed by creating aggregated observational data with the same format as the GCM outputs to assess both the data and relative contribution of each tile.
-
In models where seasonally varying schemes are implemented, a possible improvement could be achieved by including seasonal land-cover data and better representing land-cover types. The sensitivity of models to the biogeophysical effects of different land-cover types can also be investigated to ensure the signal of different land-cover types are better captured.
-
It would be useful for models to provide a short description of how their tiles are classified. Although the LUMIP protocol defines what should be included in each tile, models may interpret the classifications differently, which would affect their outputs.
Acknowledgments.
This study was supported by the LAMACLIMA project, part of AXIS, an ERA-NET initiated by JPI Climate, and funded by BELSPO (BE, Grant B2/181/P1) with cofunding by the European Union (Grant 776608). The computational resources and services used in this work were provided by the VSC (Flemish Supercomputer Center), funded by the Research Foundation—Flanders (FWO) and the Flemish Government—department EWI. We also express our deepest gratitude to Professor Edouard Davin from the University of Bern for his insightful comments on the study performed here. His broad expertise in the subject matter has contributed to the success of our analysis and interpretation of the results yielded. We also extend our appreciation to Eddy Robertson from the Met Office for providing us with helpful discussion on some uncertainties we were having. His insights were very useful for interpreting key results in the study.
Data availability statement.
All data products used to produce the results of this article are discussed in section 2. For reproducibility of the results, all scripts that were used in this study are provided in a GitHub repository (https://github.com/VUB-HYDR/2022_Admasu_etal_JAMC.git). Each script is tied to a specific step in the method and is described in Table 4.
Description of scripts provided in the GitHub repository.
APPENDIX A
Seasonal Analysis
As highlighted in the main text, tile-level outputs from GCMs are based on annual land-cover data that do not account for seasonal variation in land-cover type. Following this, LTD was also performed by mapping the GCM outputs to corresponding annual land-cover maps. Therefore, here we investigate if the performance of LTD varies significantly from season to season. This is done to investigate if the annual land-cover maps used have better or worse agreement with different seasons. If so, the performance of LTD may be relatively better in some seasons than others.
The results are presented here for air temperature in CESM2 (Fig. A1), surface temperature in CESM2 (Fig. A2), and air temperature in UKESM1 (Fig. A3). As can be observed, the effect of LTD in each season is mostly similar, with little variation among seasons for both models and variables. Therefore, it is fair to conclude that the method is not dependent on seasons.
(left) RMSE score and (right) KGE score distribution of CTL, MAP, and MAP_COR of air temperature data based on CESM2 in (a),(b) December–February; (c),(d) June–August; (e),(f) March–May; and (g),(h) September–November. All boxplots follow the format described in Fig. 8.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-21-0265.1
As in Fig. A1, but for surface temperature.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-21-0265.1
As in Fig. A1, but for UKESM1 results.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-21-0265.1
APPENDIX B
Additional Figures
Figure B1 shows that surface temperature results from CESM2 show bias differences that are consistent with the air temperature data in Fig. 2. Figure B2 shows that grid average outputs of UKESM1 support the findings of Fig. 3 that both models reduce bias in equatorial regions but do not improve variability and correlation with the observed data. Figure B3 illustrates that surface temperature results of CESM2 are consistent with air temperature results shown in Fig. 3 but show higher bias (RMSE) in most parts of the globe. Figure B4, which contains a global elevation map that is based on the regridded GMTED2010 dataset (Danielson and Gesch 2011), is used to support the analysis of the results shown in Figs. 6 and 7.
Mean bias between reconstructed and acquired grid average surface temperature datasets based on CESM2 output.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-21-0265.1
(a) RMSE and (b) KGE scores for air temperature data based on UKESM1 output.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-21-0265.1
(a) RMSE and (b) KGE scores for surface temperature data based on CESM2 output.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-21-0265.1
Global elevation map based on the regridded GMTED2010 dataset (Danielson and Gesch 2011).
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-21-0265.1
REFERENCES
Boysen, L. R., and Coauthors, 2020: Global climate response to idealized deforestation in CMIP6 models. Biogeosciences, 17, 5615–5638, https://doi.org/10.5194/bg-17-5615-2020.
Brousse, O., H. Wouters, M. Demuzere, W. Thiery, J. V. d. Walle, and N. P. M. v. Lipzig, 2020: The local climate impact of an African city during clear-sky conditions—Implications of the recent urbanization in Kampala (Uganda). Int. J. Climatol., 40, 4586–4608, https://doi.org/10.1002/joc.6477.
Burton, C., and Coauthors, 2019: Representation of fire, land-use change and vegetation dynamics in the Joint UK Land Environment Simulator vn4.9 (JULES). Geosci. Model Dev., 12, 179–193, https://doi.org/10.5194/gmd-12-179-2019.
Ciais, P., and Coauthors, 2014: Carbon and other biogeochemical cycles. Climate Change 2013: The Physical Science Basis, T. F Stocker et al., Eds., Cambridge University Press, 465–570.
Clark, D. B., and Coauthors, 2011: The Joint UK Land Environment Simulator (JULES), model description—Part 2: Carbon fluxes and vegetation dynamics. Geosci. Model Dev., 4, 701–722, https://doi.org/10.5194/gmd-4-701-2011.
CORDEX, 2021: CORDEX experiment design for dynamical downscaling of CMIP6. CORDEX Doc., 8 pp., https://cordex.org/wp-content/uploads/2021/05/CORDEX-CMIP6_exp_design_RCM.pdf.
Danabasoglu, G., and Coauthors, 2020: The Community Earth System Model Version 2 (CESM2). J. Adv. Model. Earth Syst., 12, e2019MS001916, https://doi.org/10.1029/2019MS001916.
Danielson, J. J., and D. B. Gesch, 2011: Global multi-resolution terrain elevation data 2010 (GMTED2010). USGS Open-File Rep. 2011-1073, 34 pp., https://pubs.usgs.gov/of/2011/1073/pdf/of2011-1073.pdf.
Davin, E. L., and N. de Noblet-Ducoudré, 2010: Climatic impact of global-scale deforestation: Radiative versus nonradiative processes. J. Climate, 23, 97–112, https://doi.org/10.1175/2009JCLI3102.1.
Davin, E. L., and Coauthors, 2020: Biogeophysical impacts of forestation in Europe: First results from the LUCAS (Land Use and Climate Across Scales) regional climate model intercomparison. Earth Syst. Dyn., 11, 183–200, https://doi.org/10.5194/esd-11-183-2020.
de Noblet-Ducoudré, N., and Coauthors, 2012: Determining robust impacts of land-use-induced land cover changes on surface climate over North America and Eurasia: Results from the first set of LUCID experiments. J. Climate, 25, 3261–3281, https://doi.org/10.1175/JCLI-D-11-00338.1.
ESA, 2017: Land cover CCI product user guide version 2. Université catholique de Louvain (UCL)-Geomatics Rep. CCI-LC-PUGV2, 105 pp., maps.elie.ucl.ac.be/CCI/viewer/download/ESACCI-LC-Ph2-PUGv2_2.0.pdf.
Eyring, V., S. Bony, G. A. Meehl, C. A. Senior, B. Stevens, R. J. Stouffer, and K. E. Taylor, 2016: Overview of the Coupled Model Intercomparison Project phase 6 (CMIP6) experimental design and organization. Geosci. Model Dev., 9, 1937–1958, https://doi.org/10.5194/gmd-9-1937-2016.
Fischer, E. M., K. W. Oleson, and D. M. Lawrence, 2012: Contrasting urban and rural heat stress responses to climate change. Geophys. Res. Lett., 39, L03705, https://doi.org/10.1029/2011GL050576.
Good, E. J., D. J. Ghent, C. E. Bulgin, and J. J. Remedios, 2017: A spatiotemporal analysis of the relationship between near-surface air temperature and satellite land surface temperatures using 17 years of data from the ATSR series. J. Geophys. Res. Atmos., 122, 9185–9210, https://doi.org/10.1002/2017JD026880.
Gupta, H. V., H. Kling, K. K. Yilmaz, and G. F. Martinez, 2009: Decomposition of the mean squared error and NSE performance criteria: Implications for improving hydrological modelling. J. Hydrol., 377, 80–91, https://doi.org/10.1016/j.jhydrol.2009.08.003.
Gutowski, W. J., Jr., and Coauthors, 2016: WCRP COordinated Regional Downscaling EXperiment (CORDEX): A diagnostic MIP for CMIP6. Geosci. Model Dev., 9, 4087–4095, https://doi.org/10.5194/gmd-9-4087-2016.
Hannah, L., 2015: The climate system and climate change. Climate Change Biology, 2nd ed. Academic Press, 13–53, https://doi.org/10.1016/B978-0-12-420218-4.00002-0.
Hirsch, A. L., R. Prestele, E. L. Davin, S. I. Seneviratne, W. Thiery, and P. H. Verburg, 2018: Modelled biophysical impacts of conservation agriculture on local climates. Global Change Biol., 24, 4758–4774, https://doi.org/10.1111/gcb.14362.
Hooker, J., G. Duveiller, and A. Cescatti, 2018: A global dataset of air temperature derived from satellite remote sensing and weather stations. Sci. Data, 5, 180246, https://doi.org/10.1038/sdata.2018.246.
Jalota, S. K., B. B. Vashisht, S. Sharma, and S. Kaur, 2018: Climate change projections. Understanding Climate Change Impacts on Crop Productivity and Water Balance, Academic Press, 55–86, https://doi.org/10.1016/B978-0-12-809520-1.00002-1.
Jones, P. W., 1998: A user’s guide for SCRIP: A spherical coordinate remapping and interpolation package, version 1.4. Los Alamos National Laboratory Doc., 27 pp.
Jones, P. W., 1999: First- and second-order conservative remapping schemes for grids in spherical coordinates. Mon. Wea. Rev., 127, 2204–2210, https://doi.org/10.1175/1520-0493(1999)127<2204:FASOCR>2.0.CO;2.
Kling, H., M. Fuchs, and M. Paulin, 2012: Runoff conditions in the upper Danube basin under an ensemble of climate change scenarios. J. Hydrol., 424–425, 264–277, https://doi.org/10.1016/j.jhydrol.2012.01.011.
Lange, S., 2020: ISIMIP3b bias adjustment fact sheet. ISIMIP, https://www.isimip.org/.
Lawrence, D. M., and Coauthors, 2016: The Land Use Model Intercomparison Project (LUMIP) contribution to CMIP6: Rationale and experimental design. Geosci. Model Dev., 9, 2973–2998, https://doi.org/10.5194/gmd-9-2973-2016.
Lawrence, D. M., and Coauthors, 2019: The Community Land Model version 5: Description of new features, benchmarking, and impact of forcing uncertainty. J. Adv. Model. Earth Syst., 11, 4245–4287, https://doi.org/10.1029/2018MS001583.
Lejeune, Q., E. L. Davin, G. Duveiller, B. Crezee, R. Meier, A. Cescatti, and S. I. Seneviratne, 2020: Biases in the albedo sensitivity to deforestation in CMIP5 models and their impacts on the associated historical radiative forcing. Earth Syst. Dyn., 11, 1209–1232, https://doi.org/10.5194/esd-11-1209-2020.
Malyshev, S., E. Shevliakova, R. J. Stouffer, and S. W. Pacala, 2015: Contrasting local versus regional effects of land-use-change-induced heterogeneity on historical climate: Analysis with the GFDL Earth System Model. J. Climate, 28, 5448–5469, https://doi.org/10.1175/JCLI-D-14-00586.1.
Nourani, V., Z. Razzaghzadeh, A. H. Baghanam, and A. Molajou, 2019: ANN-based statistical downscaling of climatic parameters using decision tree predictor screening method. Theor. Appl. Climatol., 137, 1729–1746, https://doi.org/10.1007/s00704-018-2686-z.
Praskievicz, S., 2018: Downscaling climate-model output in mountainous terrain using local topographic lapse rates for hydrologic modeling of climate-change impacts. Phys. Geogr., 39, 99–117, https://doi.org/10.1080/02723646.2017.1378555.
Rechid, D., E. Davin, N. de Noblet-Ducoudré, and E. Katragkou, 2017: CORDEX Flagship pilot study “LUCAS—Land use & climate across scales”—A new initiative on coordinated regional land use change and climate experiments for Europe. Geophysical Research Abstracts, Vol. 19, Abstract EGU2017-13172, https://meetingorganizer.copernicus.org/EGU2017/EGU2017-13172.pdf.
Sachindra, D. A., K. Ahmed, M. M. Rashid, S. Shahid, and B. J. C. Perera, 2018: Statistical downscaling of precipitation using machine learning techniques. Atmos. Res., 212, 240–258, https://doi.org/10.1016/j.atmosres.2018.05.022.
Sangelantoni, L., A. Russo, and F. Gennaretti, 2019: Impact of bias correction and downscaling through quantile mapping on simulated climate change signal: A case study over central Italy. Theor. Appl. Climatol., 135, 725–740, https://doi.org/10.1007/s00704-018-2406-8.
Schultz, N. M., X. Lee, P. J. Lawrence, D. M. Lawrence, and L. Zhao, 2016: Assessing the use of subgrid land model output to study impacts of land cover change. J. Geophys. Res. Atmos., 121, 6133–6147, https://doi.org/10.1002/2016JD025094.
Sellar, A. A., and Coauthors, 2019: UKESM1: Description and evaluation of the U.K. Earth System Model. J. Adv. Model. Earth Syst., 11, 4513–4558, https://doi.org/10.1029/2019MS001739.
Shea, D., 2014: Regridding overview. NCAR Climate Data Guide, https://climatedataguide.ucar.edu/climate-data-tools-and-analysis/regridding-overview.
Souverijns, N., W. Thiery, M. Demuzere, and N. P. M. V. Lipzig, 2016: Drivers of future changes in East African precipitation. Environ. Res. Lett., 11, 114011, https://doi.org/10.1088/1748-9326/11/11/114011.
Tang, Z., and J. Fang, 2006: Temperature variation along the northern and southern slopes of Mt. Taibai, China. Agric. For. Meteor., 139, 200–207, https://doi.org/10.1016/j.agrformet.2006.07.001.
Thiery, W., E. L. Davin, H.-J. Panitz, M. Demuzere, S. Lhermitte, and N. Lipzig, 2015: The impact of the African Great Lakes on the regional climate. J. Climate, 28, 4061–4085, https://doi.org/10.1175/JCLI-D-14-00565.1.
Thiery, W., E. L. Davin, S. I. Seneviratne, K. Bedka, S. Lhermitte, and N. P. M. van Lipzig, 2016: Hazardous thunderstorm intensification over Lake Victoria. Nat. Commun., 7, 12786, https://doi.org/10.1038/ncomms12786.
Thiery, W., E. L. Davin, D. M. Lawrence, A. L. Hirsch, M. Hauser, and S. I. Seneviratne, 2017: Present-day irrigation mitigates heat extremes. J. Geophys. Res. Atmos., 122, 1403–1422, https://doi.org/10.1002/2016JD025740.
Thiery, W., and Coauthors, 2020: Warming of hot extremes alleviated by expanding irrigation. Nat. Commun., 11, 290, https://doi.org/10.1038/s41467-019-14075-4.
Vanderkelen, I., N. P. M. van Lipzig, and W. Thiery, 2018: Modelling the water balance of Lake Victoria (East Africa)—Part 2: Future projections. Hydrol. Earth Syst. Sci., 22, 5527–5549, https://doi.org/10.5194/hess-22-5527-2018.
Van de Walle, J., W. Thiery, O. Brousse, N. Souverijns, M. Demuzere, and N. P. M. van Lipzig, 2020: A convection-permitting model for the Lake Victoria Basin: Evaluation and insight into the mesoscale versus synoptic atmospheric dynamics. Climate Dyn., 54, 1779–1799, https://doi.org/10.1007/s00382-019-05088-2.
Van de Walle, J., W. Thiery, R. Brogli, O. Martius, J. Zscheischler, and N. P. M. van Lipzig, 2021: Future intensification of precipitation and wind gust associated thunderstorms over Lake Victoria. Wea. Climate Extremes, 34, 100391, https://doi.org/10.1016/j.wace.2021.100391.
Wilby, R. L., S. P. Charles, E. Zorita, B. Timbal, P. Whetton, and L. O. Mearns, 2004: Guidelines for use of climate scenarios developed from statistical downscaling methods. IPCC Data Distribution Centre Doc., 27 pp., https://www.ipcc-data.org/guidelines/dgm_no2_v1_09_2004.pdf.
Winckler, J., and Coauthors, 2019: Different response of surface temperature and air temperature to deforestation in climate models. Earth Syst. Dyn., 10, 473–484, https://doi.org/10.5194/esd-10-473-2019.
Xu, Z., Y. Han, and Z. Yang, 2019: Dynamical downscaling of regional climate: A review of methods and limitations. Sci. China Earth Sci., 62, 365–375, https://doi.org/10.1007/s11430-018-9261-5.
Zhang, L., Y. Xu, C. Meng, X. Li, H. Liu, and C. Wang, 2020: Comparison of statistical and dynamic downscaling techniques in generating high-resolution temperatures in China from CMIP5 GCMs. J. Appl. Meteor. Climatol., 59, 207–235, https://doi.org/10.1175/JAMC-D-19-0048.1.
Zhao, G., J. Dong, Y. Cui, J. Liu, J. Zhai, T. He, Y. Zhou, and X. Xiao, 2019: Evapotranspiration-dominated biogeophysical warming effect of urbanization in the Beijing-Tianjin-Hebei region, China. Climate Dyn., 52, 1231–1245, https://doi.org/10.1007/s00382-018-4189-0.
