Increases in Water Balance‐Derived Catchment Evapotranspiration in Germany During 1970s–2000s Turning Into Decreases Over the Last Two Decades, Despite Uncertainties

Understanding variations in catchment evapotranspiration (EC) is critical as it directly affects water availability for humans and ecosystems. Previous studies found increases in EC in Central Europe over recent decades, but fixed study periods may not fully reveal inter‐decadal hydroclimatological variability. We performed a multi‐temporal trend analysis of water balance‐derived EC for 461 German catchments and the period 1964–2019. We accounted for previously often neglected changes in storage and uncertainties in precipitation. EC generally increased throughout Germany during 1970s–2000s (>2 mm year−2), while it showed milder changes and decreases afterward. These variations were robust to uncertainties in precipitation (median relative uncertainty of 26%) and broadly coherent with sparse plot‐scale data. Variations in EC were related with variations in precipitation and radiation, with a potentially increasing influence of precipitation after 2000s. These findings provide a reference for synthesizing current knowledge on variations in EC and their uncertainties.


Introduction
Evapotranspiration (E) couples the water, energy, and carbon cycles.Since E is a major water flux, a proper understanding of current and potential future variations in E is critical to predict water availability for humans and ecosystems.Our knowledge of past variations in E across regions is still limited, mainly because of the difficulty in measuring E throughout the landscape and the interplay among multiple drivers, like changes in climate and land cover (Teuling et al., 2019).
Previous studies largely found increases in E in Central Europe over past decades (Duethmann & Blöschl, 2018;Hobeichi et al., 2021;Pan et al., 2020;Pluntke et al., 2023;Teuling et al., 2009;Ukkola & Prentice, 2013;Yang et al., 2023).Pluntke et al. (2023) reported increases in catchment E (E C , from the observed water balance) of 2.1 mm year 2 over 1969-2019 at the experimental Wernersbach catchment in Eastern Germany, and Duethmann and Blöschl (2018) reported average increases in E C of 2.9 mm year 2 across 156 Austrian catchments over 1977-2014.Most studies analyzed trends over fixed study periods, which may not fully reveal inter-decadal variability (Hannaford et al., 2013(Hannaford et al., , 2021;;Vicente-Serrano et al., 2021).For example, Duethmann and Blöschl (2018) noticed that increases in E C over Austrian catchments mostly occurred between 1980 and 1995, and Pluntke et al. (2023) did not detect significant increases in E from a flux tower in the Wernersbach catchment between 1997 and 2019.Whether increases in E extend to larger scales and more recent decades still remains open.
Flux towers and lysimeters provide direct E measurements at the plot-scale, but these are available for a limited number of sites and typically do not cover inter-decadal periods.Diagnostic products offer regional E estimates by relying on satellite data or by upscaling plot-scale data (Pan et al., 2020), but their representativeness at the catchment scale remains debated (Lehmann et al., 2022;Tan et al., 2022).One of the firmest observational approaches to assess long-term variations in E remains estimating catchment evapotranspiration (E C ) from the observed water balance: where P is precipitation, Q streamflow, and ΔS changes in the terrestrial water storage, S.
Yet, uncertainties in observed water balance components may involve considerable uncertainties in E C (Kampf et al., 2020).While Q can be measured at catchment scale and annual estimates are typically associated with relatively low uncertainties, P and the assumptions on ΔS can be considered the main sources of uncertainty for E C estimates.Uncertainties in P variations result from measurement errors, the upscaling of point measurements, and potential inhomogeneities over time.Whereas previous studies investigated the effect of different products on P trends (Duan et al., 2019;Gomis-Cebolla et al., 2023;Jiao et al., 2021), few works did so for E C trends (Ukkola & Prentice, 2013).Furthermore, most studies on E C trends assumed ΔS negligible over multi-year periods (Duethmann & Blöschl, 2018;Teuling et al., 2009;Ukkola & Prentice, 2013;Vadeboncoeur et al., 2018), even though ΔS can be relevant over 10-year periods for some catchments (Bruno et al., 2022;Han et al., 2020).The Gravity Recovery And Climate Experiment (GRACE) satellite mission provides estimates of S anomalies, but only at 1°resolution and since 2002.Groundwater is generally a main component of S; nevertheless data on groundwater levels (GWL) are usually limited and they require local knowledge on the specific yield of wells to gain S variations.For catchments where Q is mainly controlled by S, the portion of S connected to Q variations (dynamic S, S dyn , Staudinger et al., 2017) can be estimated through streamflow recession analysis (Kirchner, 2009).Yet, estimates of S dyn have hardly been used to derive E C (Aulenbach & Peters, 2018) and assess uncertainties associated with the neglection of ΔS in E C trends.
Here we characterize variations of water-balance derived E C for Germany and the last six decades, which we further compare to variations from plot-scale E data.To increase the robustness of E C estimates, we use a homogeneous P data set and estimates of changes in S dyn (ΔS dyn ).We further assess uncertainties in the estimates of E C variations, by using different P products and alternative S data for the more recent decades.Finally, we evaluate variations in main climatic drivers of E C, that is, P, as a proxy for available moisture, and radiation (R), as main contributor of past changes in atmospheric evaporative demand (Duethmann & Blöschl, 2018).

Study Area and Streamflow Data
We selected Germany as study area and focused on catchments completely within it to exploit national-scale climatic data which do not cover transboundary areas (Section 2.1.2).We collected daily Q data and catchment boundaries from the environment agencies of the German Federal States, the Global Runoff Data Center, and the Global Streamflow Indices and Metadata Archive (Do et al., 2018a(Do et al., , 2018b)).We selected catchments with an area of 50-1000 km 2 to focus on near-natural conditions (Stahl et al., 2010), excluding catchments with known direct human impacts (e.g., water withdrawals and transfers) from information by data providers.We only included catchments with less than 5% of missing data per year and at least 28 years of data over the study period , with hydrological years starting in November throughout the manuscript).To ensure high data quality we retained only catchments with annual Q less than annual P, and without change points in annual Q (Pettitt test, Pettitt, 1979, p ≤ 0.05) which frequently indicate problems in Q data (Slater et al., 2021).This led to 461 partly nested study catchments with median area (interquartile range) of 173 (105/320) km 2 and elevation of 348 (145/ 502) m a.s.l.(Figure S1 and Table S1 in Supporting Information S1).We grouped these catchments into three

Geophysical Research Letters
10.1029/2023GL107753 regions (Figure S1 in Supporting Information S1), following the German river classification system (Pottgiesser & Sommerhäuser, 2004), and long-term variations in P and S dyn (Figure S2 in Supporting Information S1).

Climatic Data
To focus on long-term consistency in P data, we used a German-wide gridded data set which relies on a constant station network over time (P1).It applies the SPHEREMAP method (Shepard, 1968;Willmott et al., 1985) to interpolate daily data from 1300 rain gauges of the German Weather Service (DWD) to a resolution of 0.11°, as used by Hoffmann et al. (2018).To quantify the uncertainties in E C variations from uncertainties in P, we used additional P products.The HYRAS data set, provided by DWD at 1 km and daily resolution (DWD, 2023b) for Germany, is based on all available stations (up to ≥6,000, Rauthe et al., 2013).This ensures high resolution and station density, which however varies over time.E-OBS is a European-wide data set at daily and 0.1°resolutions from station data of the European Climate Assessment and Data set (v26.0e, Cornes et al., 2018), comprising an ensemble of 20 interpolations to reflect uncertainties from spatial upscaling.Finally, ERA5-Land provides global P fields at resolutions of 9 km and 1 hr (Muñoz-Sabater et al., 2021).As a reanalysis product, it overcomes potential data inhomogeneities over time.
To avoid systematic P underestimations for gauge undercatch, we corrected the observational products following Richter (1995), who proposed corrective coefficients for the study area depending on precipitation type and gauge exposure.To discriminate between rain and snow, we used air temperature (T ) data from the E-OBS data set (v26.0e, Cornes et al., 2018).We assumed corrective coefficients for moderately sheltered locations for all cells, given low impact of using alternative assumptions on E C trends (Duethmann & Blöschl, 2018).Finally, for each catchment we computed catchment-average P time series to estimate E C (Section 2.2.2) and analyze variations in P as driver of variations in E C (Section 2.2.4).
To investigate variations in R as additional climatic driver, we used data from 32 stations provided by DWD (DWD, 2023a) with <25% missing daily data per year and ≥20 years of data over the study period.We computed annual anomalies for each station and averaged them among the study region to derive regional average annual R anomalies.

Benchmark Storage Data
We obtained regional anomalies in S from GRACE (S GRACE henceforth) over 2003-2019.To minimize uncertainties from specific GRACE products, we retrieved the CSR mascon product (RL06v02, Save et al., 2016) and the GFZ Level-3 product (RL06v05, Boergens et al., 2020), and we averaged them over the study regions.Furthermore, we used GWL data from the environment agencies of the German Federal States compiled by CORRECTIV.Lokal (2022).We selected 1052 wells located within our study catchments and with no more than 1 month of missing data in each year over 2003-2019.

Additional Evapotranspiration Data for Comparison
We used E data from grass-covered and forested monitoring sites (flux towers and lysimeters, Figure S1 and Table S2 in Supporting Information S1).Since long-term data are sparse, we included sites in neighboring countries.Specifically, we used lysimeter data from Britz (Müller, 2009), St Arnold (Harsch et al., 2009), Rheindhalen, and Rietholzbach (Hirschi et al., 2017).We further selected 11 flux towers with ≥10 years of quality-checked and continuous data (<25% of missing daily data) over the study period from the Fluxnet2015 data set (Pastorello et al., 2020), the Warm Winter 2020 data set (Warm Winter, 2020 Team & ICOS Ecosystem Thematic Centre, 2022), the European Fluxes Database Cluster, Hörtnagl, Buchmann, et al. (2023), Hörtnagl, Shekhar, et al. (2023), and Pluntke et al. (2023).We pre-processed the data to avoid suspicious data and changes in land-cover or site management, basing on information from data providers and we associated each monitoring site with a study region (Table S2 in Supporting Information S1).

Water Balance-Derived Catchment Evapotranspiration
We estimated annual E C from the observed water balance (Equation 1).As a first-order approximation of ΔS, we derived (ΔS dyn ) from the analysis of streamflow recession data as proposed by Kirchner (2009) for catchments

10.1029/2023GL107753
where Q generation is mainly controlled by S dyn (details in Text S1 in Supporting Information S1).Briefly, we followed Brutsaert (2008) and Stoelzle et al. (2013) for the selection of recession periods and the derivation of catchment-specific recession parameters, which we then used to estimate S dyn from Q data (Kirchner, 2009).We calculated annual series of ΔS dyn from S dyn at the beginning and end of each hydrological year.We removed unrealistic values which can occur in individual catchments and years, due to uncertainties in streamflow recession analyses (Stoelzle et al., 2013) and violation of the assumption of Q mainly controlled by S dyn (Text S1 in Supporting Information S1).
For estimating inter-decadal E C variations, we derived E C from the precipitation data set P1 and ΔS dyn , as a "best estimate".We used additional P products (Section 2.1.2) and different assumptions on ΔS (negligible, approximated by ΔS dyn or by changes in S GRACE , ΔS GRACE ) for alternative E C estimates to evaluate uncertainties in inter-decadal E C variations.

Trend Analysis
For trend detection we adopted the Mann-Kendall test (significance level of 0.05, Mann, 1945;Kendall, 1975) with trend-free prewhitening to remove lag-one autocorrelation (Yue et al., 2002) and the Sen's slope estimator (Sen, 1968) to quantify trend magnitudes (in the following trends).We performed a multi-temporal trend analysis to explore E C variations over multiple subperiods within 1964-2019.We considered all subperiods ≥20 years to ensure sufficient length for trend detection.We estimated trends for E C averaged across the study catchments of each region and for individual catchments.For each subperiod, we included all catchments with a maximum of two years missing (removed before trend detection).This resulted in a variable number of catchments over the subperiods, but the number of catchments within each subperiod was kept constant to avoid artifacts in trend detection.
As a benchmark for trends in E C , we quantified trends in plot-scale E data over the whole record period of each site.Given the low number of catchments with long-term monitoring of E and Q, we did not perform a comparison at the scale of specific catchments, but we visually compared E and E C temporal dynamics by region, and we verified the coherence of their trends.

Uncertainties in Variations of Catchment Evapotranspiration
We quantified uncertainties in E C trends from uncertainties in P as the standard deviation in E C trends from different P products (weighted by 0.05 for trends from the 20 E-OBS members, and 1 for others).We assessed potential uncertainties in E C trends neglecting ΔS as the range of E C trends when ΔS is neglected or approximated by ΔS dyn (using the precipitation data set P1).We presented these uncertainties as percentage of the absolute E C trend from the "best estimate".
For evaluating our S dyn estimates, we compared S dyn , S GRACE , and GWL over 2003-2019, in terms of regional average monthly deseasonalized anomalies (baseline period 2004-2009) and Pearson's correlation coefficient (r).
To allow comparison in case of different storage capacity/water yields, we first deseasonalized the time series of S dyn and GWL, following Güntner et al. (2023), and then aggregated catchments/wells within each region.To evaluate the influence of uncertainties from ΔS dyn on E C variations, we further derived E C estimates using regional ΔS GRACE and the P1 data set (E C,GRACE ) over 2003-2019.To this end, we used 271 catchments with realistic E C,GRACE data (Equation S8, Text S1 in Supporting Information S1).

Contribution of Main Climatic Drivers
We calculated partial correlations between E C and two main climatic drivers (P and R).To focus on long-term variability, we smoothed the time series through a Gaussian kernel with a 2-year standard deviation, which we also used for visualization purposes throughout the manuscript.We calculated regional averages over the study catchments for E C and P, and over the stations for R. Furthermore, we fitted a multi-linear regression (MLR) model with E C as dependent, and P and R as independent variables, over the whole study period  and a period with strong E C increases (1970-2000, Section 3.1).

Inter-Decadal Variations in Catchment Evapotranspiration
Increases in regional averages of E C dominated across Germany in the first part of the study period, while decreases occurred over recent decades (Figure 1).Significant strong increases (>2 mm year 2 ) were observed between 1970s and 2000s, with an increase of 3 mm year 2 or 5.1% decade 1 on average over the three regions for 1970-2000 for instance.The timing of the transition to negative trends differed across the regions.Significant strong negative trends were identified for subperiods already starting in mid-1980s in the Pre-Alpine region (Figure 1a) and in early 1990s in the Western one (Figure 1b).Similar results were obtained when excluding nested catchments to avoid potential redundant information (Figure S3 in Supporting Information S1).
For individual catchments, trends in E C were unavoidably heterogeneous in terms of magnitude and significance, but the sign of significant trends was broadly consistent among catchments, especially for subperiods with strong significant trends at the regional scale (Figure S4 in Supporting Information S1).As an example, for 1970-2000 significant positive (negative) trends were observed for ≥33% (≤3%) of catchments within each region and the standard deviation of trends among catchments was lower than the average trend for each region (Table S3 in Supporting Information S1).
Only 4 (out of 15) monitoring sites showed significant trends in E over their record period.We detected increases at the lysimeter St Arnold over 1967-2015 and decreases at the flux towers CH-Lae, FR-Hes, and DE-Obe over the last two decades (Figure S5 and Table S2 in Supporting Information S1).
Figure 1.Inter-decadal variations in catchment evapotranspiration (E C ) at the regional-scale.Multiple trend analysis for regional E C for the Pre-Alpine (a), Western (b), and Eastern (c) regions, and (d) regional average annual anomalies in E C .In (d) we smoothed the time series for visualization purposes, according to the details provided in Section 2.2.4 (for unfiltered data refer to Figure 3a).Hatched cells in (a-c) indicate significant trends at the 5% level.

Uncertainties in Variations of Catchment Evapotranspiration
For subperiods with significant strong trends (absolute magnitude ≥2 mm year 2 ), uncertainties in regional E C trends from P were lower than estimated trends, with median (maximum) uncertainties of 26% (63%) of trend magnitude (Figures 2a-2d).Relative uncertainties slightly increased to a median of 36% when focusing on all subperiods with moderate trends (absolute magnitude ≥1 mm year 2 ).
Potential uncertainties by assuming negligible ΔS were lower than uncertainties from P, with a median of 13% for all subperiods with moderate trends, and relevant for the sign of estimated trends (i.e., ≥100%) only for subperiods with specific start/end years or short duration (i.e., along the diagonals of the matrices, Figures 2e-2h).
Anomalies in S dyn generally agreed with those from both GWL (r > 0.6 for all regions) and S GRACE (r > 0.43, Figure S6 in Supporting Information S1). S GRACE showed strong negative anomalies over 2018-2019 and replacing ΔS dyn with ΔS GRACE led to higher E C values, but still plateaued or decreasing (Figure S7 in Supporting Information S1).

Discussion
Multi-temporal trend analysis showed widespread increases of regionally averaged E C across Germany during 1970s-2000s, while mild changes and decreases over the last two decades (Figure 1).The increases between the 1970s and 2000s are consistent with Teuling et al. (2009), Duethmann and Blöschl (2018), and Pluntke et al. (2023).The time-varying approach complements previous works by providing a broader picture of E C variations over the last six decades and facilitating comparison among studies.At the regional scale and over the whole period, we found lower increases than those observed in Austria over 1977-2014 (Duethmann & Blöschl, 2018) and the Wernersbach catchment in Eastern Germany over 1969-2019 (Pluntke et al., 2023).However, we found strong increases for individual catchments and over specific subperiods (e.g., 1970(e.g., -2000, Figure S4 , Figure S4 in Supporting Information S1), pointing to regional differences in the timing and magnitude of E C increases in Central Europe.E C trends for individual catchments were generally coherent in sign (Table S3 in Supporting Information S1), suggesting that regional-scale trends are representative.However, the Eastern region comprises a comparatively lower number of catchments than others, due to widespread human impacts on Q in its central part (Figure S1 in Supporting Information S1).Trends in E from plot-scale data were broadly consistent with those in E C, with past increases and a turnaround around 2000s, despite a limited number of sites with significant changes.Differences between E and E C can be related to scale-differences and the fact that most E data did not coincide with the study catchments.Methodological challenges can further hamper the comparison, such as the difficulty in accounting for evaporation from interception in E data from flux towers (Pluntke et al., 2023).
For "best estimates" of E C trends, we used a homogeneous observational P data set (Section 2.1.2) and we estimated ΔS dyn from streamflow recession analysis (Section 2.2.1).We showed that uncertainties from different P products did not affect the sign of significant strong trends (median relative uncertainty of 26%, Figures 2a-2d).Yet, uncertainties from P were potentially relevant in some regions for subperiods with milder trends, which means that alternative P products may result in E C trends with even different sign.Uncertainties from P were generally higher than potential uncertainties that may have arisen by assuming negligible ΔS, though the uncertainty estimates were based on a different number of members (Section 2.2.3).Potential uncertainties by assuming negligible ΔS were relevant for short and specific subperiods, such as those starting or ending in the wet years 1981 and 1998 associated with large-scale floods in the study area (Uhlemann et al., 2010).We derived firstorder estimates of ΔS through streamflow recession analysis (Kirchner, 2009;Stoelzle et al., 2013).This approach relies on the assumption of Q mainly controlled by S (Kirchner, 2009), as done previously for many catchments in the study area (e.g., Berghuijs et al., 2016;Stoelzle et al., 2013), and it quantifies the portions of S connected to Q, neglecting those only connected to E and intercatchment groundwater flows (IGF, Dralle et al., 2018).A disagreement between S dyn and S GRACE over the more recent years (Figure S6 in Supporting Information S1) may be due to variations in S not connected to Q and to uncertainties in S GRACE .Thomas et al. (2016) quantified trends in groundwater storage for mesoscale catchments in the USA from S dyn estimates, GWL, and GRACE data.They found stronger agreement of trends from S dyn to those from GWL than to those derived from GRACE, which may further indicate uncertainties in GRACE data at small spatial scales.Uncertainties in estimates of S dyn are expected to be particularly relevant for individual catchments (e.g., where IGFs are significant) and specific subperiods (e.g., during heavy rainfall events associated with surface processes or in case of shifts in the recession parameters over time, Trotter et al., 2024).We aimed at reducing these uncertainties and we further checked that possible uncertainties in S dyn , as compared to S GRACE , did not affect the sign of the detected E C variations over the last two decades (Figure S7 in Supporting Information S1).Alternative methodologies for the derivation of S dyn from Q data during recessions (Stoelzle et al., 2013) or statistical approaches could be used in future work to explicitly quantify the uncertainty in E C trends due to uncertainties in ΔS dyn.
Correlations of P and R to E C (Figure 3) suggest that both drivers contributed to past variations in E C over the study area, similarly to previous findings for Central Europe (Duethmann & Blöschl, 2018;Teuling et al., 2009Teuling et al., , 2019)).The decreases in R until around 1980 and the increases afterward that we detected are in line with previous studies (Sanchez-Lorenzo et al., 2015), and they reflect climatic variations and changes in air pollution ("global dimming/brightening," Wild, 2012).Regional R variations may be affected by uncertainties related to the relatively low density of stations.P variations reflect a drying tendency over the study area during the last two decades, with summer droughts in 2003 (Pluntke et al., 2023;Teuling et al., 2013), 2015 (Ionita et al., 2017), and 2018-2019(Boergens et al., 2020).MLR underestimates high E C values around 2000 when fitted to the entire study period or overestimates E C during recent years when fitted over 1970-2000 (Figure S8 in Supporting Information S1), which suggests an increasing influence of P over the recent decades.The increasing importance of P over R on E C variations is also intuitively supported by decreasing E C with still high values in R over the last two decades and is in line with findings from global climate modeling showing widespread transitions from energy-to water-limited ecosystems (Denissen et al., 2022).We focused on climatic drivers of E C variations, building on previous studies which showed changes in climate, and in R in particular, as the main contributors for variations in E in large parts of Europe (see e.g., Duethmann & Blöschl, 2018;Teuling et al., 2019).While we focused on main climatic factors, future research should rigorously attribute the identified E C trends to their drivers, considering variations in P seasonality, in additional climatic variables, including relative humidity, wind speed, and T, and in land-use and -cover.Global changes in climate, atmospheric CO 2 concentration, and land-use and -cover recently promoted widespread vegetation greening, which was a major driver of E increases in many regions over 2001-2020 according to diagnostic products (Yang et al., 2023).
Understanding long-term E variations is essential to properly support forest and water management for society and ecosystems.Decreasing E under drying conditions points to increasing stress on vegetation during recent droughts over the study area (Pluntke et al., 2023;Senf et al., 2020).Long-term E variations may help contextualizing the role of E in surface water availability during droughts (Pluntke et al., 2023;Teuling et al., 2013) and hydrological non-stationarities triggered by them (Gardiya Weligamage et al., 2023;Massari et al., 2022).

Conclusions
We investigated (a) inter-decadal variations in data-based E C from a homogeneous observational precipitation P product and accounting for ΔS dyn of the catchments, (b) the robustness of these variations to the main sources of uncertainties, and (c) variations in the main climatic drivers of E C .E C largely increased across Germany between 1970s and 2000s, while it showed mild changes and tendencies to decreases over the last two decades (Figure 1).These variations were broadly coherent with sparse plot-scale data and robust to uncertainties.Uncertainties from P were in the order of 26% of trend magnitude and larger than those from neglecting ΔS dyn (Figure 2).To further reduce uncertainties from P, it is recommended to use homogeneous, observational P products for future E C trend analyses.If ΔS dyn are not accounted for, short study periods or those with strong storage anomalies at the start/end should be avoided.Increases in E C over 1970s-2000s reflected variations in P and R during the global brightening phase, while recent decreases in E C over the last drying decades with still high R values suggested an increasing influence of moisture variability on variations in E C (Figure 3).Our findings provide a framework to synthesize studies on variations in E C in Central Europe over recent decades, including their uncertainties and potential drivers, which is relevant for freshwater and forest management in a transient climate.
Drivers E C variations mirrored variations in P and R, despite a varying importance (ρ P,E|R = 0.84 and ρ R,E|P = 0.6 over 1964-2019, Figures 3a-3c).Periods with high P were the mid-1960s, the early 1980s, and years around 2000.R generally decreased until around 1980 and increased afterward.While high P values in the mid-1960s and around 2000 were also reflected by high E C values, high P in the early 1980s did not correspond to high E C , likely due to low

Figure 2 .
Figure2.Uncertainties in regional trends in catchment evapotranspiration (E C ). Uncertainty from different P products (a-d) and potential uncertainty by assuming ΔS negligible (e-h), as percentage of absolute trends from the "best estimate" of E C (Section 2.2.3).Uncertainties are considered only for subperiods with moderate and strong trends.

Figure 3 .
Figure 3. Variations in catchment evapotranspiration (E C ) and their main climatic drivers (P and R).Regional average anomalies in annual E C (a), P (b), and R (c).Shaded colors refer to unfiltered time series, whereas full colors to smoothed time series (Section 2.2.4).Note the different y-scales for visualization purposes.

R.
MLR fitted to the entire study period generally reflected the variability in E C (FigureS8in Supporting Information S1), despite underestimating high E C values around 2000.MLR fitted to the subperiod with increasing E C overestimated E C before 1970 and after 2000 (FigureS8in Supporting Information S1).