El Niño La Niña cycle and recent trends in continental evaporation

Transcription

1 SUPPLEMENTARY INFORMATION DOI: /NCLIMATE2068 El Niño La Niña cycle and recent trends in continental evaporation Diego G. Miralles 1 *, Martinus J. van den Berg 2, John H. Gash 3,4, Robert M. Parinussa 3, Richard A. M. de Jeu 3, Hylke E. Beck 3, Thomas R. H. Holmes 5, Carlos Jiménez 6, Niko E. C. Verhoest 2, Wouter A. Dorigo 7, Adriaan J. Teuling 8 and A. Johannes Dolman 3 1 School of Geographical Sciences, University of Bristol, Bristol, United Kingdom 2 Laboratory of Hydrology and Water Management, Ghent University, Ghent, Belgium 3 Department of Earth Sciences, VU University, Amsterdam, The Netherlands 4 Centre for Ecology and Hydrology, Wallingford, United Kingdom 5 Hydrology and Remote Sensing Lab, USDA-ARS, Beltsville, MD, USA 6 Centre National de la Recherche Scientifique, Observatoire de Paris, Paris, France 7 Department of Geodesy and Geoinformation, Vienna University of Technology, Vienna, Austria 8 Hydrology and Quantitative Water Management Group, Wageningen University, Wageningen, The Netherlands NATURE CLIMATE CHANGE 1

2 This file presents a detailed description of the methodology (Sect. 1), datasets used as inputs (Sect. 2), experiment settings (Sect. 3) and validation of our estimates (Sect. 4). In addition it includes a more extensive discussion of some of our main findings described in the manuscript (Sect. 5). 1 Description of the methods GLEAM is a set of algorithms driven by observations that separately estimate all the components of land evaporation (or evapotranspiration, E): transpiration (E t ), interception loss (E i ), bare-soil evaporation (E b ), sublimation (E s ) and open-water evaporation (E w ) (see ref. 1,2). Additionally, GLEAM calculates root-zone soil moisture and evaporative stress conditions. The rationale of this process-based (yet simple) approach is to maximize the recovery of information about evaporation contained in the current satellite observations of climatic and environmental variables. Near-surface atmospheric humidity and windspeed are difficult to derive from satellite data, and current efforts are restricted to sensors like AIRS 3 with short temporal records of observation. Given these difficulties, the Priestley and Taylor (PT) approach 4 is preferred over the more complete Penman-Monteith (PM) equation 5. This choice relies on the rationale of keeping GLEAM as observation-based as possible, yet being able to derive multi-decadal records. The PT equation used in GLEAM calculates potential evaporation (E p ) based only on observations of surface net radiation (R n ) and surface air temperature (T). E p estimates for the fractions of vegetation and bare soil within each pixel (see below) are converted into E t and E b (respectively) using a multiplicative evaporative stress factor (S). The derivation of S is based on observations of vegetation optical depth (τ) a proxy for vegetation water content 6 and estimates of root-zone soil moisture (θ). Every land pixel in GLEAM is composed of three cover fractions 7 : tall canopy, short vegetation and bare soil. The root zone of the tall canopy fraction consists of three vertical soil layers: a first layer from 0 15 cm, a second from cm and a third from cm. The root zone for the fraction of short vegetation comprises the first two layers (0 15 cm and cm), while only the first 15 cm layer is considered for the bare soil fraction. θ is calculated at daily time steps based on a multilayer running water balance that describes the infiltration of precipitation (P) through the vertical soil profile, and in which the estimates of E from the previous time step are substracted. 2

3 Microwave surface soil moisture observations (θ obs ) can also be assimilated into the profile. To derive E i, Gash s analytical model of rainfall interception 8,9 is driven by observations of P while considering different vegetation characteristics (e.g. fraction of tall canopy per pixel, canopy storage properties, etc.) see ref. 10 for a complete description of the interception loss methodology and validation. The method has been recently used in an isotope study to determine the relative importance of E t in the terrestrial water cycle 11. For water bodies and for regions covered by ice and snow, E w and E s are calculated based on a PT equation run with look-up table parameter values for open water and ice 12 respectively. The main features of the approach are: (a) the consideration of soil moisture constraints acting on E (independently of whether the additional data assimilation of θ obs is included), (b) the detailed parameterization of E i, and (c) the extensive use of microwave observations (an asset under cloudy conditions). A schematic of the methodology is presented in Fig. S1. Figure S1 Schematic of GLEAM. Blue boxes represent the four modules. Observational driving data of snow water equivalents (D s ), net radiation (R n ), air temperature (T), surface soil moisture (θ obs ), vegetation optical depth (τ) and precipitation (P) are illustrated in red. The different evaporation (E) components are illustrated in green. E is estimated as the sum of its five components. Transpiration (E t ) and soil evaporation (E b ) are substracted from the soil profile in the next time step. GLEAM also uses ancillary data of land cover fractions 7, lightning frequency climatology 13 (to estimate mean rainfall rates 10 ) and soil properties 14. 3

4 Figure S2 Comparison of the current version of GLEAM to the version given by Miralles et al. 2. (a) Global annual time series of E for both versions. The SOI and the mean ± 1 standard deviation of the diagnostic datasets by ref. 15 are represented. Annual anomalies are calculated relative to the mean of the 11-year period to be consistent with Fig. 1. (b) Spatial variability of the average E for the period for current version, (c) Same but for the version in ref. 2. Inputs from e1 are used to run both versions. Here we use the current version of GLEAM algorithms and datasets. There have been three updates since ref. 2: (a) the previously potential and exponential-shaped stress curves have been linearized in order to reduce unnecessary complexity in the calculation of S; (b) τ is now also used to calculate S for tall vegetation, to better capture the effect of seasonal phenological changes in forests (before this was restricted to short vegetation); and (c) the data assimilation of soil moisture has been simplified. These first two modifications had minor effects on the estimates and are not further discussed. The original data assimilation in ref. 2 was based on a Kalman Filter design in which errors in the passive microwave θ obs by ref. 16 were calculated based on passive microwave observations of τ. In the current version, to incorporate the state-of-the-art combined active-passive microwave θ obs dataset by ref. 17, the assimilation has been simplified to follow a Newtonian Nudging (NN) design 18,19. NN proceeds by nudging the model predictions to θ obs, using a fixed factor based on an empirical estimate of the θ obs errors. For this study, only a single fixed factor was applied, as the use of 4

5 more factors did not improve assimilation as revealed by the validation exercise (see Sect. 4). The derivation of the errors of θ obs was undertaken by ref. 20 using triple collocation 21,22. Although a variety of data assimilation approaches were attempted including the Kalman Filter in various configurations no gains were observed for methods that allowed more flexibility in the representation of the model and measurement errors. As such, NN was chosen on the basis of its simplicity, its low computational cost and its effectiveness. After these three updates the current version of GLEAM remains very similar to the original one 2. This is demonstrated in Fig. S2, where both versions are executed with the same set of inputs (see inputs from experiment e1 in Sect. 2 and 3). Figure S2a compares the annual time series of global evaporation for the period in both versions. Annual dynamics remain very similar. Additionally, the visual comparison between Fig. S2b and c shows that the spatial distribution of the average evaporation is nearly identical for both versions. Estimates of E and θ by the original version of GLEAM have been validated for the year 2005 and evaluated at the global scale for the period in previous studies 1,2,10. These estimates have recently been inter-compared to other datasets 15 and applied to study land-atmosphere interactions 11,23. The errors of GLEAM have also been mapped in space using triple collocation 1. The performance relative to other methodologies has been investigated in recent activities like the European Space Agency (ESA) WACMOS-ET project or the NASA GEWEX Radiation Panel LandFlux initiative 24,25. Here we take the methodology one step further and we run it with multiple datasets as input (see Sect. 2 and 3) to gain insight into the sensitivity of our results to uncertainties in the input data. A detailed validation of the resulting E and θ estimates against in situ measurements is undertaken in Sect. 4. Comparison to other methods is also performed: Fig. 1a and Fig. S8 represent the annual time series of global E together with those reported by the machine learning algorithm by ref. 26, while Fig. 1a, S2a and S8 illustrate the mean ± 1 standard deviation of the annual E anomalies of the five diagnostic sets by ref. 15. These diagnostic datasets are all satellite-based E products: two PM approaches 27,28, the machine learning algorithm 26, the original version of GLEAM 1,2 and an atmospheric water balance 29. 5

6 2 Input data Rather than being oriented to produce a static product of E, GLEAM is intended to be a flexible tool able to run with different sets of its required input data. This flexibility allows the selection of the most suitable set of inputs based on the time period of a specific study, and the quality of these inputs over the spatial domain of interest. This applies mainly to P, R n, and T, but potentially also to θ obs, τ and snow water equivalents (D s ) when available. Currently, GLEAM runs at daily time steps exclusively, and therefore a minimum time resolution of a day is required for all its inputs. Because the goal of this study is to gain insight into the long-term trends of E and especially the potential causes of multi-year variability, we have selected datasets that spanned the entire satellite era as inputs to the methodology. Table S1 summarizes the datasets used here to run GLEAM. Note that alternative datasets are used in the case of P, R n and T to explore the errors induced by the choice of inputs and how these propagate into the estimates of E (see also Sect. 3). During the selection of datasets we have prioritized those of a more observational nature. The inputs of R n used to drive the methodology are obtained from ERA-Interim 30 and the NASA/GEWEX Surface Radiation Budget (SRB) 3.0 (ref. 31). ERA-Interim is the latest reanalysis by the European Centre for Medium-Range Weather Forecasts (ECMWF). The dataset is global and spans from 1979 to near-real time at a 0.75 latitude-longitude resolution. Compared to ERA-Interim, the R n from GEWEX SRB-3.0 is mainly based on satellite observations. However, as radiative fluxes at surface level cannot be directly measured from satellites, it still requires the combination of radiation algorithms with measured top-of-the-atmosphere radiances and meteorological inputs from satellites and atmospheric reanalysis. The spatial resolution of SRB-3.0 is 1.0 and it only spans the period For the input of T in GLEAM we also used ERA-Interim reanalysis data: the daily 2-metre air temperature calculated as the average of the eight 3-hour estimates over a daily cycle. However, more frequently we drive GLEAM with a merging of T data from the International Satellite Cloud Climatology Project (ISCCP) 32 and the Atmospheric Infrared Sounder (AIRS) onboard NASA s Aqua satellite. The ISCCP record spans from at 2.5 resolution; AIRS spans the period

7 2011 at 1.0 resolution. To blend both datasets we first use a 3-day moving average to gap-fill AIRS data and then we rescaled it to ISCCP via cdf matching at every pixel using the overlap period This blending methodology is similar to the ones applied to conform the θ obs and τ multi-satellite datasets used in our study 6,17 (see below). To extend the record of T back in time to 1980, we use data from ERA-Interim reanalyses for The blending is performed through the same procedure described above (i.e. cdf matching to the ISCCP record). Hence the resulting T merged dataset spans the period : (a) ERA-Interim cdf-matched to ISCCP for , (b) ISCCP for , and (c) AIRS cdf-matched to ISCCP for Table S1 Variables and datasets used as input to derive E in GLEAM. Despite the original spatial resolution listed, all datasets are re-gridded to The merged dataset of T consists of ERA-Interim for , ISCCP for and AIRS for (see Sect. 2). The merged dataset of D s consists of GlobSnow snow water equivalents for the Northern Hemisphere and NSIDC data for the Southern Hemisphere. Note that in the case of R n, P and T we use two alternative datasets (see Sect. 3). Variable Dataset Original resolution ( ) Period Net radiation (R n ) ERA-Interim GEWEX SRB Precipitation (P) Air temperature (T) CPC Unified 0.5 (0.25 in CONUS) GPCP-1DD v Merge of ISCCP with ERA-Interim and AIRS 1.5 / 2.5 / 1.0 (respectively) ERA-Interim Surface soil moisture (θ obs ) WACMOS-CCI Vegetation optical depth (τ) LPRM Snow depth (D s ) GlobSnow and NSIDC

8 The P input is critical for our study. We do not use reanalysis data, since purely observational datasets are available spanning the entire period of study. Two different products are used: (a) the Climate Prediction Center (CPC) Unified Gauge-Based Analysis of Global Daily Precipitation (CPC- Unified) 33, and (b) the daily Global Precipitation Climatology Project (GPCP-1DD) version 1.2 (ref. 34). CPC-Unified daily precipitation is based on gauge data as reported by different national and international agencies. The spatial resolution is 0.25 for the Continental US domain and 0.5 elsewhere. Data are available from 1979 to 2012; up to 2006 it is based on over 30,000 stations while more recent estimates are based on fewer rain gauges (~17,000). Alternatively, GPCP-1DD v1.2 is produced by merging precipitation from a range of microwave, infrared, and sounder satellite observations, yet it is also corrected using precipitation measured by gauges on the surface. The available record spans at 1.0 resolution. GPCP-1DD has arguably been the most widely used satellite-based P product over the last few years due to its overall quality and record length 35. The dataset of combined active-passive microwave surface soil moisture by ref. 17 is used as θ obs in the assimilation routine described in Sect. 1. This dataset blends information from the active European Remote Sensing scatterometers (ERS-1 and ERS-2) and the Advanced Scatterometer (ASCAT), with passive observations from the Scanning Multichannel Microwave Radiometer (SMMR), the Special Sensor Microwave Imager (SSM/I), the microwave imager from the Tropical Rainfall Measuring Mission (TRMM) and the Advanced Microwave Scanning Radiometer (AMSR- E). The dataset has been developed as part of the Water Cycle Multi-mission Observation Strategy (WACMOS) project of the European Space Agency (ESA), and has been further improved by ESA s Climate Change Initiative soil moisture project (see It currently spans from 1980 to 2011 and has a spatial resolution of 0.25 latitude-longitude. Note that this dataset is also used to evaluate global soil moisture trends and interannual variability in Fig. 2 and 3 the E estimates derived from the experiment run with no data assimilation of θ obs (e2, see next section) are used in those figures to keep E and θ obs independent through these comparisons. We also make use of τ by ref. 6 in GLEAM for the derivation of the evaporative stress (S). These data are based on the application of the dual-polarization Land Parameter Retrieval Model 8

9 (LPRM) 16 to passive microwave observations from SMMR, SSM/I, TRMM and AMSR-E. Finally, the input data of D s comes from ESA s GlobSnow Snow Water Equivalent (SWE) product version 1.0 (ref. 36). GlobSnow is based on data from SMMR, SSM/I and AMSR-E combined with ground-based weather station measurements. Since GlobSnow covers the Northern Hemisphere only, data from the National Snow and Ice Data Center (NSIDC) 37 are used in the snow-covered regions of the Southern Hemisphere. All datasets are re-gridded to a common 0.25 prior to their use in GLEAM via Shepard s Method of inverse distance weighted interpolation Experiment Settings Table S2 describes the five GLEAM experiments (e1 e5) that yield the five different sets of E illustrated in the main manuscript in Fig. 1a,c and 2a (and in multiple other figures within this Sup. Information). These experiments are based on combinations of the inputs presented in Table S1. The experiment referred to as e1 results in what we consider our GLEAM default product. It is the preferred experiment, representing a compromise between the extensive record length (32 years), and the use of observations over reanalysis inputs (the later limited to ERA-Interim R n ). The data products of θ obs, τ and D s are common to all simulations (see Table S1). Note that the length of the E datasets from each experiment in Table S2 varies according to the availability of the inputs listed in Table S1. The Clausius-Clapeyron expectation showed in Fig. 1c is calculated by driving GLEAM with a seasonal climatology of all the inputs of the experiment e1. The seasonal expectation of a variable for the day-of-the year, d, at a given pixel, is calculated as the average of the entire multi-year dataset for that day of the year (i.e. the average value of the variable in the 32 days d in ). Then a 31-day moving average is applied this is a common practice to calculate seasonal climatologies 22,39. In the case of T, however, the seasonal climatology is trended by adding an increment (every day) calculated based on the slope of the linear ( ) trend of T at the pixel. As GLEAM is based on the PT equation, which infers the Clausius-Clapeyron relation via the parameterization of Δ (the slope of the saturation water vapour against air temperature), it is suitable for calculating the Clausius- Clapeyron expected effect on E based on the observed long-term T change at a particular pixel. 9

10 Table S2 Description of the inputs used in each of the five different experiments. The experiment e1 yields the GLEAM reference product. The difference between the other experiments and e1 is marked bold fonts. R n P T Period e1 ERA-Interim CPC-Unified Merge of ISCCP with ERA-Interim and AIRS e2 ERA-Interim CPC-Unified No assimilation of surface soil moisture Merge of ISCCP with ERA-Interim and AIRS e3 GEWEX SRB 3.0 CPC-Unified Merge of ISCCP with ERA-Interim and AIRS e4 ERA-Interim GPCP-1DD v1.2 Merge of ISCCP with ERA-Interim and AIRS e5 ERA-Interim CPC-Unified ERA-Interim Validation We test the realism of GLEAM s estimates of E and θ by comparing them to ground measurements of evaporation from FLUXNET 40 and soil moisture from the International Soil Moisture Network 41 (ISMN), respectively. FLUXNET is a global network of micrometeorological towers that quantifies carbon, water vapour and energy fluxes using the eddy-covariance technique. As the method tends to miss the evaporation from wet canopies, we compare its measurements against the GLEAM E without the interception loss component. On the other hand, ISMN is a recent international cooperation to produce an organized database of soil moisture in situ measurements that comprises different national networks. Measurement techniques and quality vary across these networks, but data from different providers are harmonized, quality-checked, and converted into common volumetric soil moisture units. Here we use D s to mask out measurements taken under snow, as sensors are prone to fail in frozen soil conditions. While the longest records from FLUXNET and ISMN span at least two decades, stations with more than ten years of measurements are scarce. The only selection criterion for the use of FLUXNET and ISMN locations in this validation exercise is the requirement of a record length of at least one year. 10

11 In situ measurements have then been compared to GLEAM pixel estimates of E and θ from the five experiments, e1 e5. Since ISMN measurements may come from different depths in the soil, we compare against the first layer (i.e., 0 15 cm) or the second layer (i.e., cm) of model soil moisture depending on the depth of the sensors. FLUXNET stations sample a distance of approximately 100 to 2000 m upwind of the tower, while the ISMN measurements of soil moisture are essentially point measurements. Given the 0.25 resolution of GLEAM, station-to-pixel representativeness errors can be significant, especially in the case of θ. This scale contrast is aggravated by the large spatial variability of evaporation and soil moisture in nature. Nonetheless, one single 0.25 pixel may contain more than one single ground station. In those cases, the in situ measurements have been averaged prior to the comparisons. Consequently, the final validation study is based on 701 ISMN sensors and 163 FLUXNET stations, albeit only 100 and 126 pixel-to-location comparisons for θ and E, respectively. The total number of daily aggregates in the comparisons is 193,197 and 2,370,070 for E and θ, respectively. Note that this validation is substantially more extensive in terms of number of ground locations and length of the validation period than the one presented in ref. 2. All experiments lead to estimates of θ that compare favourably against daily ground measurements from ISMN: the average Pearson s correlation coefficients ( R ) of all the pixel-tostation comparisons ranges between for the different experiments. While the average correlation coefficient for all stations for the reference GLEAM product (simulation e1) is 0.71, the simulation e2, which uses the same inputs as e1 but does not assimilate any θ obs, displays only a slightly lower R = The main reason for this limited added skill by the data assimilation is the lower quality revealed by θ obs as compared to the GLEAM soil moisture without data assimilation (i.e. e2): when we limit the validation to days with satellite observations (θ obs ), the R of e2 against the ISMN measurements is 0.71, substantially larger than the R = 0.50 that the θ obs display in comparison against the in situ data. 11

12 Figure S3 Results of the validation of θ against ISMN measurements. Spatial distribution of the Pearson s correlation coefficients (R) for experiment e1. The lower panel shows R at each station for experiments e1 e5. In the maps, locations on a single 1 pixel have been combined to aid visualization. The symbol sizes are directly proportional to the length of in situ records (from one to fifteen years). Stations in bottom panel are represented individually and grouped per continent in the order released by ISMN based on geographical coordinates. Figure S3 illustrates the spatial variability of the Pearson s correlations between the daily GLEAM pixel estimates of θ and the ISMN soil moisture measurements for the reference GLEAM product (e1). The size of the symbols is directly proportional to the number of years in the comparison, ranging from one to fifteen. Additionally, the lower panel in Fig. S3 includes the correlations for the other four simulations (e2 e5) stations have been organized per continent and in no particular order. Consistently higher or lower values of R over particular regions, independently of the inputs used to drive GLEAM (i.e. e1 e5), can be explained by: (a) the structural errors in GLEAM, or (b) the quality and density of the in situ measurements differing from network to network. Overall, 12

13 the estimates of GLEAM θ agree better with the in situ data in Europe and Africa and worse in parts of central United States. This is consistent for all simulations. The validation also reveals a generally poorer performance of θ for the simulation e4, in which the input of precipitation is GPCP-1DD as opposed to the usual CPC product (Table S2). As expected, θ is less sensitive to the choice of R n and T. The same way, all experiments lead to estimates of E that compare favourably against daily FLUXNET measurements: the average Pearson s correlation coefficient ( R) of all the pixel-to-station comparisons of E ranges between for the different five experiments. Data assimilation of θ obs leads to a limited improvement in the estimates of E: R increases from 0.71 to 0.72 when θ obs are assimilated. This implies that the quality of the estimates of E when using the model soil moisture based on precipitation and evaporation only is similar to the quality of the E estimates when θ obs is also incorporated. This was expected given the marginal improvement in θ obtained after assimilation (see above). Note that in Fig. 1 and Fig. 2a of the main text, the emphasis is on experiment e1 (with data assimilation) this is to acknowledge that the data assimilation provides a small yet positive enhancement in the quality of the estimates. However, Fig. 2b,d,f and Fig. 3a,c,e are based on E from experiment e2 (no data assimilation), as the aim of those figures is to assess the relation between E and the SOI, and establish a link to θ obs ; thus estimates of E that are independent from θ obs are required in those figures. Nevertheless, our validation shows again that e1 and e2 are very similar on their estimates and their quality. Figure S4a represents the Pearson s correlations for the comparison of E against daily FLUXNET aggregates. The size of the symbols is directly proportional to the number of years in the comparison, ranging from one to ten. The quality of the in situ data is expected to be more homogeneous from network to network than in the case of ISMN, as all measurements are based on similar eddy-covariance instruments. Nevertheless, we do see some consistent spatial patterns that indicate a slightly better performance of GLEAM over Europe. While the choice of P and T inputs, and the use (or not) of θ obs, leads to estimates of E of similar quality (see lower panel in Fig. S4a), better results are obtained for the experiment e3, which uses R n from SRB 3.0 instead of ERA-Interim. However, this enhanced performance of e3 is restricted to North America only. 13

14 Figure S4 Validation against daily FLUXNET measurements of E. (a) Spatial distribution of the Pearson s correlation coefficients (R) for experiment e1; the lower panel shows R at each station for experiments e1 e5. (b) Spatial distribution of the bias (%) for experiment e1; the lower panel shows the bias at each station for experiments e1 e5. In the maps, stations falling on a single 1 pixel have been combined to aid visualization and the size of the symbols is directly proportional to the in situ record length (from one to ten years). In the time series, all the stations are represented individually and organized per continent in the order released by FLUXNET based on geographical coordinates. 14

15 Figure S4b illustrates the percentage bias of E (calculated as the difference between the estimated and measured total evaporation over the entire record at each FLUXNET station). We note, however, that the lack of closure in the energy balance of eddy-covariance towers complicates the use of FLUXNET data to identify the bias in E GLEAM estimates, and also the use of statistical diagnostics of model performance based on absolute differences (e.g. RMSE, Nash-Sutcliffe model efficiency coefficient, etc.). The reasons for the energy mis-closure in eddy-covariance stations remain largely unexplained, yet several hypotheses have been extensively studied Correction of the measured energy fluxes is possible, but the way to do it remains an open question 42. We did not apply any correction in our validation study and, therefore, care must be taken when interpreting the results in Fig. S4b. The size of the symbols in Fig. S4b is inversely proportional to the percentage of mismatch in the energy closure (ranging from 0 to >50%). Larger symbols, therefore, suggest more reliable estimates of the bias of GLEAM E. The bias of E against the measured fluxes at the stations is largely within the rage of ±20%, with generally positive values in North America. The all-stations average percentage bias is positive for the five experiments, albeit this is mainly due to the large positive bias at only four of the stations (see Fig. S5b). It is also worth noting that the eddy-covariance energy mis-closure is partly attributed to the underestimation of the latent heat fluxes 46. In regions with a marked seasonality in P, R n and T, correlations are likely to be larger (as long as the inputs used in GLEAM are able to reproduce these seasonal cycles correctly). We have therefore performed an analysis to evaluate the skill of the E products to capture daily anomalies. The seasonal expectations for each day of the year have been calculated and then removed from the raw series of E (for both GLEAM and FLUXNET). Seasonal expectations (or climatologies) are calculated as the average of the entire multi-year dataset for each day of the year and a 31-day window moving average 22,39, following the same approach as in the calculation of the Clausius-Clapeyron expectation described in Sect. 3. The deterioration of the results is shown in the form of a Taylor diagram 47 in Fig. S5a. Only the 56 stations with more than 5 years of data have been used due to the long timespan required to calculate seasonal expectations; this is instead of the usual 126 that are used in other analyses within this validation. Average correlations for all stations ( R) are lowered to values of ~0.4 15

16 or below, except for the experiment e3 that uses SRB R n, in which R = 0.6 even after the removal of the seasonal cycle from the E time series. Interestingly, while the standard deviation is overestimated by the raw GLEAM data by ~25%, the anomalies underestimate the variability of FLUXNET by ~10%. This points to excess in the amplitude of the seasonal cycle in GLEAM. Results generally confirm the slightly better performance of GLEAM when SRB is used instead of ERA-Interim R n, as indicated also in Fig. S4a. On the other hand, the improvement of the statistics when comparing monthly (instead of daily) aggregates is shown in Fig. S5b ( R = against in situ). Figure S5 Taylor diagrams of the FLUXNET validation. (a) Comparison of daily GLEAM E and daily GLEAM E anomalies, against the corresponding FLUXNET data. The origin of the arrows represents the statistics using (raw) daily data (i.e. including the seasonal cycle); the point of the arrows represents the results using daily anomalies (no seasonal cycle). Statistics represent the average of all the stations with more than five years of record (N = 56 point-to-pixel comparisons). (b) Comparison of daily GLEAM E and monthly GLEAM E, against the corresponding FLUXNET data. The origin of the arrows represents the statistics using daily data; the point of the arrows represents the results of the monthly aggregates. Statistics represent the average of all the stations (N = 126). Root Mean Squared Differences (RMSD) and standard deviations are normalized against the standard deviation of FLUXNET time series. The point denoted as FLUXNET represents the best possible statistics in the comparison. Experiments e1 e5 are illustrated independently to allow their comparison. 16

17 5 Extended discussion Uncertainty in our E estimates can arise from errors in the input data (see Table S1) and errors in GLEAM itself. Both sources of uncertainty are explored together in the validation exercise (Sect. 4). Additionally, the appropriateness of GLEAM as the main methodology in the study is confirmed in Fig. 1a by illustrating how E fits within the mean ± 1 standard deviation of the five diagnostic sets by ref. 15. Figures 1a,c and Fig. 2a, as well as the validation exercise, show the results of running GLEAM with different combinations of inputs. These figures point to a robust quality of the E estimates, rather independently from the specific range of datasets used as input. To gain more insight into this sensitivity of E to the choice of input data, Fig. S6 illustrates the global E during the 11-year overlap period of all the experiments ( ). Figure S6a shows the average annual E for e1, and Fig. S6b e represent the difference between e1 and e2, e3, e4, and e5 (respectively). Larger disagreement occurs when the inputs of P and R n are substituted; this is mainly due to the higher sensitivity of E to these inputs (not just in the model but also in nature), but it also highlights that there are disagreements between each dataset of P and R n (see also Fig. S7). In particular, the choice of SRB 3.0 R n, as opposed to ERA-Interim, leads to higher estimates of E over Amazonia (Fig. S6c). The choice of CPC-Unified P seems critical in central Africa (Fig. S6d), where the E becomes lower as a consequence of the underestimation of the CPC-Unified P in this region 33. The choice of T product, on the other hand, causes milder impacts on the E estimates (Fig. S6e). The assimilation of surface soil moisture tends to increase E in dry regions (see Fig. S6b): under dry soil conditions, E can hardly decrease after optimizing the soil moisture, even if θ obs is unbiased. As illustrated in Fig. 1c, we recognize a rising tendency in E in the Northern Hemisphere (30N 90N) that is slightly larger that the expectations from the Clausius-Clapeyron relation based on temperature trends only (see Sect. 3), and in agreement with the results by ref. 48. Figure S7a shows the annual time series of E as averaged over sections of 30 of latitude. Despite the potential impact of multi-annual climate variability as a consequence of the Artic, North Atlantic and North Pacific oscillations, temporal patterns in the Northern Hemisphere, and especially in the region 30N 90N, are rather stable pointing to a constant rise since the beginning of the 1980s. No apparent decline in E is 17

18 found that can be attributed to global dimming 49, nor more punctual decreases are found after the El Chinchon (1982) and Pinatubo (1991) eruptions. In fact, part of the interannual variability in the Northern Hemisphere seems to be also linked to ENSO. This is suggested by the similarity in the multi-year patterns in 30N 90N latitude bands to those in the 30N 60S bands, and the significant correlations found between the SOI and E in parts of the Northern Hemisphere (see Fig. 3a). But Fig. S7a also shows that different experiments yield similar interannual variability and multidecadal trends of E. This contrasts with the differences found in the interannual dynamics across the various sets of inputs as illustrated in Fig. S7b: while CPC-Unified and GPCP P agree well on their average temporal patterns, there are differences from product to product in the case of R n. This is especially so for 0 60N where R n from SRB 3.0 and ERA Interim differ substantially. The time series of ERA Interim R n show a lower interannual variability and no apparent trends at any of the latitudinal sections. Figure S6 Average annual E for the five experiments for their overlap period ( ). (a) Average E for experiment e1. Difference in average E between (b) e2 and e1, (c) e3 and e1, (d) e4 and e1, (e) e5 and e1. 18

19 Figure S7 Latitudinal averages of annual anomalies of E. (a) Anomalies of E for each of the five experiments (e1 e5), (b) Anomalies of the main datasets used as inputs (of precipitation, temperature and net radiation) in the five experiments. Anomalies are calculated relative to the corresponding mean and represented as the area-weighted average for each 30 latitudinal band. It is not surprising that GLEAM E shows no indication of an effect of dimming or brightening in the multidecadal trends, as the inputs of R n (both SRB 3.0 and ERA Interim) do not show signs of these phenomena either. Several other factors like the lack of consideration of trends in wind speed can also add uncertainty to our estimates. These simplifications can be responsible for potential disagreements with the interannual dynamics reported in previous studies using other methods; methods which of course have uncertainties of their own. As an example, Fig. S8 explores the differences between the interannual dynamics of e1 and those of the machine-learning algorithm used in ref. 26. The variability of global E differs substantially between both methods as it is already pointed in Fig. 1a in the manuscript. Particularly important for our findings, by the year 2000, when the decline noted by ref. 26 had already started, we find a peak of evaporation following an episode of La Niña. Therefore it is only after 2000 when evaporation decreases according to GLEAM as average conditions transition to El Niño. While not being a proof that the time series of GLEAM are more 19

20 realistic, we do note that the timing of the decline in GLEAM agrees better with the multi-model range by ref. 15. We note as well that specific peaks and valleys in the dynamics of e1 can be explained to a large extent by the interannual variability of T, P and R n (see Fig. S7b). It can be inferred, for instance, that for the year 2000 all the main inputs (and especially P) report high annual averages. We note as well that the dataset by ref. 26 has a low interannual variability (see Fig. S8). This has been highlighted several times in the past 50,51. Figure S8 Comparison to the time series revealed by Jung et al. 26. Time series of annual anomalies of global E for e1 and for the method by ref. 26. The SOI and the mean ± 1 standard deviation of the merging of diagnostic datasets by ref. 15 are also shown. Anomalies are calculated relative to the mean. Large part of the variability of continental evaporation responds to changes in the 30N 60S latitudinal band where ENSO effects are stronger, and where two-thirds of global E originate 1. To test the veracity of this statement, Fig. S9 illustrates the regions contributing the most to the variability of the global E. Figure S9a shows the similarities between the global monthly anomalies of E, the zonal (30N 60S) average, and the SOI. The map of the correlation coefficients between the time series of global-average monthly anomalies of E (see green line Fig. S9a) and the monthly anomalies of E at each pixel is represented in Fig. S9b. The high (significant) correlations found in eastern Australia and southern Africa the regions where we see the maximum effect of ENSO in E suggest that the strong variability of E in these regions makes them critical hot-spots controlling the global annual volume of water from continents into atmosphere. 20

21 Figure S9 Areas responsible for the variability of global average E. (a) E monthly anomalies as averaged for 30N 60S and for the entire globe. (b) Correlations between monthly anomalies of E at each pixel and the time series of global mean monthly anomalies of E. Results are based on experiment e1 and Dotting indicates statistical significance. A 3-month moving average is applied to the time series in a to smooth outliers. Results in Fig. 2 link the most prominent declines in E to periods of limited moisture supply. These periods coincide with El Niño episodes. During those periods the monthly SOI and the average (30N 60S) monthly anomalies of E show significant positive correlations ranging from for the different periods. Figure S10 represents these relationships during each period. Interestingly, monthly precipitation anomalies are also positively correlated to the SOI, but the monthly potential evaporation anomalies from GLEAM remain uncorrelated or slightly negatively (but not significantly) correlated. This points to restrictions in the moisture supply as the cause for the correlations between SOI and E. We also see that prolonged increases in E are, on the other hand, usually related to transitions to La Niña conditions. During La Niña, the availability of moisture, especially in Australia, is abnormally high. This is show in Fig. S11, that analyzes three prolonged periods of increase in monthly anomalies of E. Results from all three periods show agreement in highlighting Australia as a 21

22 key region defining the variability of global E (as pointed also by Fig. S9). There is less agreement in the sign and magnitude of the anomalies in southern Africa and especially in eastern South America. We draw the attention again to the correspondence between the moisture availability (indicated by independent satellite soil moisture observations 17 ) and the E anomalies (from e2). Figure S10 Correlations of E, potential evaporation (E p ) and P against the SOI for 30N 60S. For the three different periods of decline in E monthly anomalies illustrated in Fig. 2, the scatterplots represent the relation between the monthly SOI and E, E p and P anomalies. R is the Pearson s correlation coefficient. Each point in a scatter represents a different month. Periods are defined based on the timespan of the declines as revealed by the E series in Fig. 2a (and marked by the green boxes in that figure). E and E p come from experiment e1. P comes from the CPC-Unified product

23 Figure S11 Episodes of prolonged E increase and their relation to ENSO. (a) Zonal (30N 60S) average monthly anomalies of E and the SOI for Three prolonged periods of increase in E anomalies (i.e , and ) are highlighted in blue. Red and blue horizontal lines mark the 1 st and 3 rd quartile of the SOI (respectively). A 3-month moving average is applied to the time series of E and SOI to smooth outliers. (b) Trends of anomalies of E for the period. (c) Same but for soil moisture 17. (d) Trends of anomalies of E for the period. (e) Same but for soil moisture. (f) Trends of anomalies of E for the period. (e) Same but for soil moisture. Dotting represents statistical significant trends (p < 0.05). Grey shading represents missing data. Monthly anomalies are calculated relative to the multi-year mean for each month of the year (see Methods Summary). E data in b, d and f comes from the GLEAM experiment e2 that does not assimilate soil moisture observations; this is done to keep E independent from the soil moisture observations used in c, e and g. 23

24 Figure S12 Monthly anomalies of the different continental evaporation fluxes during ENSO episodes. (a) Time series of monthly anomalies of transpiration, bare soil evaporation and vegetation interception loss as averaged for the 30N 60S latitudinal band. They are plotted together with the monthly SOI during (b) Anomalies in transpiration during El Niño (considered as the months with SOI within the first quartile of the SOI density distribution). (c) Same but for La Niña (defined as the months with SOI within the forth quartile of the SOI density distribution). (d) Anomalies in soil evaporation during El Niño. (e) Same but for La Niña. (f) Anomalies in rainfall interception loss during El Niño. (g) Same but for La Niña. The period considered in the anomalies calculation is A 3-month moving average is applied to the time series shown in a to smooth outliers. Data corresponds to experiment e1. 24

25 Our results therefore suggest that a relationship exists between ENSO and the E dynamics in eastern Australia and southern Africa, and that it propagates to the global E averages (see Fig. S9). We show that this occurs because during El Niño the supply of water in eastern Australia and southern Africa is insufficient to meet the atmospheric demand for water vapour. To test this hypothesis we have analysed monthly anomalies of E, soil moisture and NDVI, and calculated their magnitude, trends and correlations with the SOI in Fig. 3. Figure S12 goes one step further and looks at the different components of E separately. Figure S12a shows the average zonal (30N 60S) transpiration, bare-soil evaporation and rainfall interception loss as calculated using the inputs from experiment e1. Figure S12 b g show the anomalies of these fluxes during El Niño and La Niña based on the record. Figure S12b reveals the importance of transpiration in vegetated regions where it is the largest contribution to E (ref. 11). Soil evaporation anomalies are also important in more arid regions, like the Australian deserts and the around Kalahari and Namibia. Despite not being water-limited, ENSO exerts a control over E in Amazonia and Indonesia (Fig. 3a,d). This is explained by negative anomalies in interception loss as a consequence of the reduced rainfall during El Niño (see Fig. S15b). In our calculations of correlations and anomalies against monthly SOI not just in Fig. 3 but also in other figures in this S.I., e.g., Fig. S12b g) a 1-month lag in the monthly SOI has been considered as the appropriate time to allow atmosphere and vegetation to adapt to the pressure changes in the Pacific. This 1-month lag was derived from the analyses in Fig. S13, where we apply different lag times to the monthly SOI and calculate the correlations with the (pixel) estimates of monthly anomalies of E (from e1). While for Australia the highest correlations are found when no lag is applied to the SOI, a 1-month lag increases the correlations in southern Africa. This points to a longer response period in Africa to the variations of pressure in the Pacific (as expected). The correlations are maxima in the East of South America when a longer lag-time of 3 5 months is applied (Fig. S13d f). The 1-month lag (Fig. S13b) seems the most appropriate for the analyses as it is both sufficiently short to capture the effects of ENSO changes in the variability of E in Australia, and sufficiently long to capture the effects in South Africa. Note however that the SOI is heavily auto-correlated (as also are the time series of monthly anomalies of E), and that therefore a shift of 1 month in the analysis only changes results marginally, as becomes apparent in Fig. S13. 25

26 In recent years, ENSO indices other than the SOI have become increasingly popular. Figure S14 support our main conclusions by reproducing similar results to the ones shown in the main text but using the Multivariate ENSO Index 52 (Fig. S14a c) and the Sea Surface Temperature (SST) Index (see (Fig. S14d f). Note that to ease the visual comparison with the SOI, both indices are presented as their negative value. Figure S13 Lagged (Pearson s) correlations between monthly anomalies of SOI and E for Different lags are introduced in the SOI to explore when the maximum sensitivity to the differences in atmospheric pressure occurs (i.e. the time taken by atmosphere and vegetation to respond to ENSO variations): (a) no lag, (b) lag of 1 month (use of SOI from 1 months before E), (c) 2 months (d) 3 months, (e) 4 months, (f) 5 months. E data comes from experiment e1. Dotting represents statistically significant correlations (p < 0.05). 26

27 Figure S14 Correlations with other ENSO indices. (a) Time series of monthly anomalies of E and the negative value of the MEI during for the 30N 60S latitude band. Correlation coefficients between monthly anomalies of the negative of the MEI 52 and (b) E, (c) soil moisture 17. (d) Time series of monthly anomalies of E and the negative of the SST ENSO index (i.e. average SST in 20N 20S minus the rest of the globe, see Correlation coefficients between monthly anomalies of the negative of the SST ENSO index and (e) E, (f) soil moisture. E data comes from experiment e2 to aid comparison to Fig. 2a and 3a,b. Dotting represents statistically significant (p < 0.05) correlations. Grey areas represent missing data. A 3-month moving average is applied to the time series in a and d to smooth outliers. 27

28 Figure S15 Anomalies in monthly ocean evaporation and land precipitation during ENSO episodes. (a) Time series of monthly anomalies of zonal (30N 60S) average ocean evaporation 53 and global land precipitation 33 plotted together with the monthly SOI and the anomalies in land evaporation (from e1) during (b) Global anomalies in ocean evaporation during El Niño (considered as the months with SOI within the first quartile of the SOI density distribution). (c) Same but for La Niña (defined as the months with SOI within the forth quartile of the SOI density distribution). (d) Anomalies in land precipitation during El Niño. (e) Same for La Niña. The period considered in the anomalies calculation is A 3-month moving average is applied to the time series in a to smooth outliers. The hypothesis that ENSO-related terrestrial water limitation controls land evaporation dynamics contains the tacit idea that changes in atmospheric pressure and SST in the Pacific cause changes in ocean evaporation and atmospheric circulation. These changes propagate to the terrestrial water cycle, via reduced supply of rainfall. We have seen that this occurs mainly during El Niño 28