1. Introduction
Soil moisture state and dynamics are key for climate and water resources assessments, particularly in water-limited periods and regions [
1]. Soil moisture availability regulates numerous physical processes of the earth system; namely, the partitioning of energy fluxes at the land surface, agricultural growth and ecosystem dynamics, streamflow generation, subsurface drainage, and groundwater recharge [
2,
3,
4,
5,
6,
7]. The memory of soil moisture, understood as the persistence of soil moisture anomalies, can be of the order of weeks to months [
8] and, therefore, longer than the memory of atmospheric anomalies (hours to days). This difference in residence times gives soil moisture a buffering or intensifying impact on climate extremes at the surface such as droughts, floods, or heat waves [
9], as well as a role in the development of atmospheric processes at shorter timescales [
10]. Thus, the soil moisture response to meteorological droughts (periods of no precipitation), and how fast the lack of precipitation leads to hydrological/agricultural droughts (plant stress, groundwater reduction, low discharge), are crucial factors that we can use to assess model capabilities in a changing climate where droughts are likely to increase in severity and frequency [
11].
In this work, we set out to evaluate drydowns, understood as soil drying rates during periods of no precipitation, in global hydrological models (GHMs) and land surface models (LSMs) against observations of evapotranspiration (ET). GHMs and LSMs can work in stand-alone mode, driven by meteorological data, or coupled to the atmosphere in general circulation and earth system models. In these models, soil moisture controls and is affected by the land–atmosphere fluxes of water and, if resolved, energy and carbon. It is strongly model-specific [
12] because of the different formulations used to resolve surface processes dependent on soil moisture state, like evaporation or runoff, and to different model parameters. The JULES (Joint U.K. Land Environment Simulator) LSM, for example, does not explicitly consider the residual soil water content (amount of water that cannot be drained from the soil, because it is retained in disassociated pores and immobile films) as a component of soil moisture [
13], which may lead to differences in soil water content values when comparing with models of a different nature or field observations as a result of spatial variations of residual water content [
14].
Global products of observed surface soil moisture estimation from satellite missions are increasingly becoming available, for example, the works of [
15,
16,
17] and independent studies have shown good correlations with in situ soil moisture data [
18,
19]. Recent studies use satellite products to characterise observed surface soil moisture drydowns [
20], compare with in situ measurements [
21], and evaluate model performance [
22]. However, the model-specific nature of soil moisture and the depth limitation to these satellite estimates (typically accounting for the top 5–10 cm of soil) limit their usefulness in model evaluation, especially in the presence of vegetation that accesses deeper soil water.
Other studies use in situ observations, for example, the works of [
23,
24], to investigate soil moisture memory; however, the sparsity and the local nature of point-scale soil moisture observations call for caution on their potential use to evaluate global models’ soil moisture dynamics [
25].
After a given precipitation event over a vegetated area, a portion of the water input on the land surface is intercepted by vegetation, another portion runs off to surface waters, and the rest infiltrates the soil providing the source water for evapotranspiration (ET). The rate of drying of the soil during following hours and days until the soil moisture reaches the wilting point for the given vegetation in the area has three different regimes [
26,
27]. These regimes include the following: (1) the drainage stage dominated by gravity while the soil moisture remains above field capacity; (2) the ET first stage when water is held by gravity (no drainage below field capacity) and drying occurs at the rate demanded by the atmospheric potential ET (PET) while soil moisture remains above the critical point (threshold for limited soil water availability for roots); and (3) the ET second stage when the drying occurs at a rate limited by the soil moisture availability for vegetation in the root zone (between the critical point and the wilting point). ‘Drydown’ will refer hereafter to this latter stage of soil drying under water limitation.
These different timescales of drying are captured using new methods of model evaluation that overcome the limitations of observed soil moisture. For example, a study has focused on the energy response of the surface at different stages through the drying process, rather than the actual soil moisture evolution [
28]. These authors evaluate global climate models, analyzing the energy balance response to drought through warming rates of the land surface, using a methodology based on satellite land surface temperature products [
29,
30].
Another method that takes advantage of the timescales of the drying, but focusing on vegetated areas and water-limited periods, is described by Teuling, et al. [
7] and Blyth, et al. [
31]. These authors showed that the surface response to soil moisture depletion can be evaluated from time series of ET alone, allowing for a direct drydown assessment using flux tower data. However, only very few sites and single drying curves were used in these studies.
The following challenge remains: Can we use flux tower data to expand this latter idea in order to find a generic drydown metric that would capture the role of vegetation? And how well do models of surface flux exchange, like LSMs and GHMs, represent the drydown process when we apply this metric to evaluate them against flux tower data? Such a parameter will help understand and tackle the problem of model variability and shortcomings in drought propagation affecting surface exchange fluxes with the atmosphere.
In
Section 2, we first describe the flux data and models used in this study, then we define a metric of drydowns, and finally we apply this to time series from 35 flux tower observations during periods of no precipitation and to a suite of 10 global model assessments (LSMs and GHMs) from the eartH2Observe project (
http://www.earth2observe.eu/). In
Section 3, we analyse and compare the model and observations results, generalizing them through the responses under different vegetation types. We then discuss the relationship between the drydown rates and vegetation types and other outcomes of the experiment, in order to come to some conclusions about how to evaluate the models.
4. Discussion
Global model outputs like the WRR1 product analysed here tend to be evaluated using earth observation datasets for a characterization of the model skill to reproduce observed global or regional water budgets (for example, see the works of [
34,
36]), whereas flux tower data can be used to evaluate the model surface fluxes exchange (for example, see the work of [
60]). In this work, using flux tower data, we look for a process based evaluation that helps characterise the models in their capability to predict drought responses, and thus provide useful information to both model developers that are constantly working to improve the tools and water managers or other potential users of WRR1 that might be interested in drought response over particular regions. The reasons for the accuracy or shortcomings of the WRR1 models in their simulation of water-limited drydowns, however, will be dependent on particular model configurations and process representations (for instance, only two models incorporate groundwater interactions with the unsaturated zone in the soil).
4.1. Role of Vegetation in τ
The analysis in
Section 3 shows a strong link between drydown and vegetation cover. We can simplify our analysis to a first order vegetation cover differences using a three-type classification: trees, shrubs, and grasses. For the models, we use the dominant vegetation type map (
Figure S2) to identify any event over tree dominated regions (broadleaf trees, needleleaf trees), shrub dominated regions, and grass dominated regions. For the observation sites, we classify sites with vegetation cover reported of “Grasslands” or “Croplands” (
Table 1) as grasses, “Savannah” or “Woody savannah” as shrubs, and the rest of the sites as trees. Under this criterion, we have a simpler characterization of drydown rates by both models and observations (
Figure 7). As discussed in
Section 3.3, there is an issue with shrubs as the models dry too quickly (
Figure 7b), with SWBM being the only model slower than observations (higher τ). Focusing on trees and grasses (
Figure 7a,c), we can see that the models represent better the drydown rates over tree covered regions (six models within 10% of observed median τ: 27.8 days), whereas over grass covered regions, the model variability and differences with observations are higher (only two models, HTESSEL-CaMa and ORCHIDEE, within 10%).
The observations show a reduction in median τ from trees to grasses of 47% (from 27.8 to 14.8 days). However, as shown in
Table 6, the models show a smaller reduction depending on whether they are LSMs (between 21% and 38%) or GHMs (22% or below, with the exception of WaterGAP3, which presents very quick drydown rates for all vegetated areas). HTESSEL-CaMa shows this reduction with very good agreement with observations for both trees and grasses. Other two LSMs, ORCHIDEE and SURFEX-TRIP, which are significantly quicker than observations over trees, present a much better agreement over grasses. JULES shows a very good agreement with observations for trees (29.7 days versus 27.8 days in the observations), but the τ reduction for grasses is too small. SWBM also shows very good agreement with observations for trees (30.1 days), but then the representation of grasses is highly overestimated with a median τ of 26.5 days versus 14.8 days in the observations (too little reduction of τ from trees to grasses).
4.2. Other Factors
We acknowledge that the soil texture and not only the vegetation type should explain part of the observed and modelled variance of the drydown rates, as it determines the soil capacity to hold and conduct water at the surface and through the root zone (i.e., sandy soils with larger pores and lower water suction will evapotranspirate more easily and have quicker drainage down the soil column). This was shown by a clear systematic decrease of drydown rates with sand fraction by McColl, et al. [
20] when they analysed surface soil moisture drying. However, when we relate our global modeled data of drydown rates (using ET ratio) with soil water suction data from the Harmonized World Soil Database (HWSD [
61]), we do not find a relationship as clear among models as the results show when relating drydown rates with vegetation type (
Figure S3).
4.3. Global Maps of Median τ
Our methodology provides global maps of median τ from the WRR1 models, shown in
Figure 3. Some of the features are robust with conclusions already drawn from the analysis at the site level and by plant functional types: (1) LSMs (
Figure 3a–d) present very similar spatial patterns with SURFEX-TRIP drying quicker than the rest and JULES slower; (2) GHMs (
Figure 3e–j) present overall slower drying than LSMs and higher model variability; (3) the within model variability in spatial patterns is more apparent in LSMs pointing at different characterization depending on land cover; and (4) WaterGAP3 is a quick outlier.
These maps may be useful way to compare the features of the different models and to use satellite data for evaluation.
4.4. Identification of Drydowns and Characterization of Sites
The identification of water-limited drydowns detailed in
Section 2.3 is a departure from recent studies where the drydown evaluation had to be discretised in bins of antecedent rainfall [
28,
30]. In addition to the main purpose of this work of evaluating how quickly drydowns occur when the soil water limitation affects surface fluxes, we have used our drydown identification method to characterise sites/regions as dry or wet, considering the ratio of periods of no precipitation that fall on our ‘dry event’ definition.
A group of sites (Blodgett, El Saler, El Saler 2, Castelporziano, Rocca 2, Amplero, and Espirra) were found to be dry by most models, but considered wet by the observations (
Figure 4), even when they were located over semiarid regions, meaning that the models somehow stressed evaporative fluxes during periods of no precipitation, even though such stress was not observed. This points to other model processes being responsible for the observation/model mismatch in the identification of sites as dry, rather than the rate of drying, like irrigation (case of El Saler 2) or intensified soil moisture memory due to shallow groundwater connectivity [
62]. Three of these sites (Blodgett, Amplero, and Espirra) were used by Ukkola, et al. [
55] to analyse a range of LSMs (JULES and ORCHIDEE included) in their representation of droughts. These authors concluded that the models overestimate intensity and duration, falling in agreement with our findings here. We note that two particular GHMs (PCR-GLOBWB and SWBM) do not find these locations to be dry, however, these models tend to characterise all sites as wet.
Other sites were characterised as wet by both observations and models, as they are located in wet and cold northern latitudes (Hyytiala, Boreas, Degero, Quebecc) or under all year long humid climate conditions (Tumba, Willow).
Therefore, even though other processes beyond soil drying under water-limited conditions, like groundwater connectivity to the root zone, are not within the scope of this analysis as the theory does not hold on such conditions, the simple identification of dry/wet sites already gives some answers to the questions of what sites are actually water-limited during no precipitation periods and whether the models agree with observations in this identification.
4.5. On the Methodolody
We acknowledge that the assumption of constant PET during the dry event is a condition in order to obtain Equation (3). Possible oscillations of PET from day to day should be minimal during a dry period, and we have calculated ET/PET on a daily time step, so such oscillations are expected to be smoothed out on the exponential decay adjustment. Further, the restriction to the methodology of only considering events with 0 < τ < 50 days will neglect possible events where PET variations might result in the inaccuracy of the method.
Following this work, but slightly out of the scope of this publication, we consider separating events, for a particular site/region, into different PET categories. This would give us further insights into understanding under what atmospheric conditions the models represent more/less accurately the drydown process.
5. Summary and Conclusions
In this paper, we study the concept that we can characterise the rate of drydown after a rainfall event as a single value, and whether this is a useful metric to evaluate models. We set out to investigate if such a key quantity could be extracted from several years of direct measurements and whether existing large-scale models have a fixed value of this metric in time.
Ideally, we would be able to observe this quantity from satellite observations so that we could obtain global maps. However, there are problems with the use of satellite soil moisture observations to study this, as only the top surface soil moisture can be seen. In order to overcome the observed soil moisture limitations, other studies have focused on the surface responses rather than the actual soil moisture evolution, for instance, analyzing the energy balance response through warming rates of the land surface, using a methodology based on satellite land surface temperature products [
28,
30]. In this study, however, we use the most direct observation of evapotranspiration—the eddy-covariance method, commonly referred to as flux tower measurements.
The study reveals that it is possible to quantify the drydown process using the exponential curve of decrease of evapotranspiration (normalised by the evaporative demand or potential evaporation) when we have daily data of evapotranspiration, potential evapotranspiration, and precipitation. We define the lifetime parameter τ of this curve as our drydown metric, and characterise the land using the median τ obtained for all dry events identified within the time series of data available. A comparison of 35 sites showed a marked relation between the vegetation cover at the site and the drydown rates, with tree sites drying slower than grassland sites because of differences in the root system, and thus the soil water that they can reach.
We quantify the drydown metric for a suite of global land surface and hydrological models from the WRR1 assessment [
34] run globally at 0.5° resolution and compare it to the results from flux tower data. We find that a large part of the disagreement between the modeled and observed metric comes from discrepancies in the land cover type specified by the models compared with the actual vegetation at the flux tower. When we group the modeled metric into vegetation cover characterizations, we obtain a more robust metric for the models that correlates well with the findings at the site level. Land surface models show a strong difference between trees and grasses, in agreement with observations, while most large-scale hydrology models show a markedly lower difference.
We conclude that flux tower data can be used to evaluate drydown processes in global models using the median lifetime parameter τ, which is a property of the land and hence independent of the precipitation forcing, as long as the vegetation cover at the sites and in the models is taken into account.