Present and future Köppen-Geiger climate classification maps at 1-km resolution

We present new global maps of the Köppen-Geiger climate classification at an unprecedented 1-km resolution for the present-day (1980–2016) and for projected future conditions (2071–2100) under climate change. The present-day map is derived from an ensemble of four high-resolution, topographically-corrected climatic maps. The future map is derived from an ensemble of 32 climate model projections (scenario RCP8.5), by superimposing the projected climate change anomaly on the baseline high-resolution climatic maps. For both time periods we calculate confidence levels from the ensemble spread, providing valuable indications of the reliability of the classifications. The new maps exhibit a higher classification accuracy and substantially more detail than previous maps, particularly in regions with sharp spatial or elevation gradients. We anticipate the new maps will be useful for numerous applications, including species and vegetation distribution modeling. The new maps including the associated confidence maps are freely available via www.gloh2o.org/koppen.

We present new global maps of the Köppen-Geiger climate classification at an unprecedented 1-km resolution for the present-day  and for projected future conditions (2071-2100) under climate change. The present-day map is derived from an ensemble of four high-resolution, topographicallycorrected climatic maps. The future map is derived from an ensemble of 32 climate model projections (scenario RCP8.5), by superimposing the projected climate change anomaly on the baseline high-resolution climatic maps. For both time periods we calculate confidence levels from the ensemble spread, providing valuable indications of the reliability of the classifications. The new maps exhibit a higher classification accuracy and substantially more detail than previous maps, particularly in regions with sharp spatial or elevation gradients. We anticipate the new maps will be useful for numerous applications, including species and vegetation distribution modeling. The new maps including the associated confidence maps are freely available via www.gloh2o.org/koppen.

Köppen-Geiger climate classification
We follow the Köppen-Geiger climate classification as described in Peel et al. 21 , which was also used by Kriticos et al. 22 (Table 2). This classification is identical to that presented by Köppen in 1936 1 with three differences. First, temperate (C) and cold (D) climates are distinguished using a 0°C threshold instead of a 3°C threshold, following the suggestion of Russell 30 . Second, the arid (B) sub-climates W (desert) and S (steppe) were identified depending on whether 70% of precipitation occurred in summer or winter. Third, the sub-climates s (dry summer) and w (dry winter) within the C and D climates were made mutually exclusive by assigning s when more precipitation falls in winter than in summer and assigning w otherwise. Note that the tropical (A), temperate (C), cold (D), and polar (E) climates are mutually exclusive but may intersect with the arid (B) class. To account for this, climate type B was given precedence over the other classes.

Climate data
The present Köppen-Geiger classification map was derived from three climatic datasets for air temperature (WorldClim V1 and V2, and CHELSA V1.2) and four climatic datasets for precipitation (WorldClim V1 and V2, CHELSA V1.2, and CHPclim V1; Table 1). All datasets have a 0.0083°resolution with the exception of CHPclim V1.2, which has a 0.05°resolution. For consistency CHPclim V1.2 was downscaled to 0.0083°using bilinear interpolation. The future Köppen-Geiger classification was produced using monthly historical and future air temperature and precipitation data from the CMIP5 archive 29 . For the future scenario, we used Representative Concentration Pathway 8.5 (RCP8.5 31 ). All climate models with data during the 1980-2016 and 2071-2100 periods were used. Data for 1980-2016 was derived by concatenating historical runs (which end in 2005) and future runs (which begin in 2006). For each model, we only considered a single initialization ensemble. In total 32 models had sufficient data and hence were used for deriving the future map (Table 1).

Present-day Köppen-Geiger map
The present-day Köppen-Geiger map (Fig. 1a) was derived from an ensemble of high-resolution climatic datasets (Table 1)   precipitation using GPCC FDR V7, for both the 1980-2016 period and the temporal span of the climatic dataset. Next, for each month we calculated climate change offsets (for temperature) or factors (for precipitation) between the two periods, and resampled these offsets or factors from 0.5°to 0.0083°r esolution using bilinear interpolation, and adjusted the climatic maps by addition (for temperature) or multiplication (for precipitation). For each adjusted temperature and precipitation climatic dataset combination, we derived a Köppen-Geiger map at 0.0083°resolution. From this ensemble of 4 × 3 = 12 maps we derived a final map by selecting, for each grid-cell, the most common class (Fig. 1a). A corresponding confidence map was derived by dividing the frequency of occurrence of the most common class by the ensemble size and converting these fractions to percentages (Fig. 2a). For example, if Csa is the most common class for a particular grid-cell, and it has been assigned eight times out of 12, the resulting confidence level is 100 8 12 ¼ 66:6%. This confidence level should be interpreted as the degree of trust we place in our final present-day classification. Confidence levels are generally lower in the vicinity of borders between climate zones, in particular at high latitudes where the climatic data show more uncertainty.

Future Köppen-Geiger map
The future Köppen-Geiger map (Fig. 1b) was derived by the so-called "anomaly method" 32 based on an ensemble of climate projections from the 32 CMIP5 models (Table 1). First, observed monthly present-day reference temperature and precipitation climatologies (0.0083°resolution) were derived, by simple averaging of the ensemble of temporally-homogenized, high-resolution climatic maps. Then, for each climate model and each month, we subsequently calculated climate change offsets (for temperature) or factors (for precipitation) between 1980-2016 and 2071-2100 and resampled these offsets or factors from the native model resolution to 0.0083°using bilinear interpolation (Fig. 3). Finally, future high-resolution climatic temperature and precipitation maps were derived from the present-day, observed reference climatologies by addition of the offsets (for temperature) or multiplication by the factors (for precipitation). We want to emphasize that the change factors are never excessively high (i.e., >5; Fig. 3), because (i) model simulations tend to overestimate the precipitation frequency 33 (resulting in the near-absence of areas with close to zero precipitation), and (ii) over the majority of arid regions the future projections tend toward drying rather than wetting 34 (resulting in factors o 1).
For each climate model, we derived a future Köppen-Geiger map at 0.0083°resolution from the downscaled future temperature and precipitation data. From this ensemble of 32 maps we derived a final map by selecting, for each grid-cell, the most common class (Fig. 1b). A corresponding confidence map was derived by dividing the frequency of occurrence of the most common class by the ensemble size and T = air temperature; P = precipitation. a The following climate models were used (initialization ensemble between parentheses):  converting these fractions to percentages (Fig. 2b). For example, if Cfa is the most common class for particular grid-cell, and it has been assigned 24 times out of 32, the corresponding confidence level is 100 24 32 ¼ 75:0%. This confidence level should be interpreted as the degree of trust we have in our final future classification based on the uncertainties in climate change projections. Thus, uncertainties are larger than for the present-day map. In particular, they are larger at high latitudes because of the greater model spread in projected warming in those regions.

Code availability
The new Köppen-Geiger classifications have been produced using MathWorks MATLAB version R2017a. The function used to classify the temperature and precipitation data according to the criteria listed in   21 . a Variable definitions: MAT = mean annual air temperature (°C); T cold = the air temperature of the coldest month (°C); T hot = the air temperature of the warmest month (°C); T mon10 = the number of months with air temperature >10°C (unitless); MAP = mean annual precipitation (mm y −1 ); P dry = precipitation in the driest month (mm month −1 ); P sdry = precipitation in the driest month in summer (mm month −1 ); P wdry = precipitation in the driest month in winter (mm month −1 ); P swet = precipitation in the wettest month in summer (mm month −1 ); P wwet = precipitation in the wettest month in winter (mm month −1 ); P threshold = 2 × MAT if >70% of precipitation falls in winter, P threshold = 2 × MAT+28 if >70% of precipitation falls in summer, otherwise P threshold = 2 × MAT+14. Summer (winter) is the six-month period that is warmer (colder) between April-September and October-March.

Data Records
The present and future Köppen-Geiger classification maps and the corresponding confidence maps are freely available for download at (Data Citation 1) and www.gloh2o.org/koppen. The maps are stored in GeoTIFF format as unsigned 8-bit integers. We also provide a legend file (legend.txt) linking the numeric values in the maps to the Köppen-Geiger climate symbols and providing the color scheme used for displaying the maps in the current study (adapted from Peel et al. 21 ). The maps are referenced to the World Geodetic Reference System 1984 (WGS 84) ellipsoid and made available at three resolutions (0.0083°, 0.083°, and 0.5°; approximately 1 km, 10 km, and 50 km at the equator, respectively). The classifications are upscaled from 0.0083°to 0.083°and 0.5°using majority resampling and the confidence levels using bilinear averaging. Table 3 presents the file naming convention. The maps can be visualized and analyzed using most Geographic Information Systems (GIS) software (e.g., QGIS, ArcGIS, and GRASS).

Technical Validation
We validated the new high-resolution present-day Köppen-Geiger classification (Fig. 1a), and previous maps from Kottek et al. 20  The newly derived high-resolution present-day Köppen-Geiger classification (Fig. 1a) exhibited a classification accuracy of 80.0%, while the maps of Kottek et al. 20 , Peel et al. 21 , and Kriticos et al. 22 exhibited classification accuracies of 66.1, 70.9, and 73.4%, respectively. These results confirm that the new map is more accurate, which is primarily due to its high (1 km) resolution and use of an ensemble of topographically-corrected climatic datasets. The map of Kottek et al. 20 showed the lowest classification accuracy, due to its low (0.5°) resolution. The map of Peel et al. 21 also performed less well, due to a lack of topographic corrections and the use of a relatively small number of stations.
We also tested the usefulness of the confidence map associated with the new present-day classification (Fig. 2a) using station observations. We obtained a mean confidence level of 92.6% for the correctly classified stations (n = 17,667) and 77.4% for the misclassified stations (n = 4411). The mean confidence level was thus substantially lower for the misclassified stations, confirming that the confidence map provides a useful indication of the classification accuracy. Figures 4 and 5 show historic Köppen-Geiger classification maps from all three previous studies and our present-day map for the Alps (Europe) and the central Rocky Mountains (North America), respectively, illustrating the enhanced detail in our map. The other maps sometimes fail to depict important topographic features; the map of Peel et al. 21 , for example, does not represent the Apennine mountains (Italy), due to a lack of topographic corrections (Fig. 4e). The new map (Figs. 4a and 5a) also exhibits better agreement with a Landsat-based forest cover map 36 (30-m resolution; Figs. 4c and 5c). The spatial extent of the polar (E) climate, for example, corresponds closely with treelines in the forest cover maps. Additionally, the new present-day and future Köppen-Geiger maps ( Fig. 4a and g, respectively) agree well with equivalent high-resolution maps derived for the Alps 8 (their Figs. 1 and 2, respectively).

Usage Notes
The future Köppen-Geiger classification (Fig. 1b) should be viewed as providing insights into potential spatial changes in regional climatic zones under climate change. However, caution should be exerted not to equate those changes directly with changes in actual biomes. First, vegetation changes by 2100 may lag the change in climate zones. Secondly, factors not accounted for in the Köppen-Geiger classification, such as higher atmospheric CO 2 levels, may alter the relationship between climate classes and vegetation. It is thus advised to interpret the future Köppen-Geiger classification first and foremost from a 'climatic conditions' perspective.
The rationale for using the anomaly method to build future maps using climate model projections, instead of directly computing present and future maps from model outputs, is that superimposing future modeled anomalies onto the observed climate removes mean biases from climate model outputs. This is a widely used method in climate change impact assessments 32 . However, an unavoidable limitation of this approach is that because of model spatial biases, modeled climate change anomalies may not be fully geographically consistent with the baseline observed climatology to which they are added 37 (e.g., if the climate of one region in a given model is spatially shifted relative to reality).  Another irredeemable limitation is that because of their coarser resolution (typically 1-2°), climate model outputs do not resolve future climate change at the same scale as our baseline climatology. Thus, in cases where there could be significant heterogeneities in precipitation change and/or warming below the model resolution (e.g., along coastlines and/or in regions with strong land-cover differences and/or   20 (1951-2000); (e) Peel et al. 21 (1916-1992); and (f) Kriticos et al. 22 (1960-1990 elevation gradients), future changes at the 1-km scale might be under-or over-estimated, because only the model-scale mean anomalies are used to compute future changes. High-elevation mountainous regions are a prime example of this because they are expected to experience considerably more warming than adjacent valleys 38 .  20 (1951-2000); (e) Peel et al. 21 (1916-1992); and (f) Kriticos et al. 22 (1960-1990