Future increases in soil moisture drought frequency at UK monitoring sites: merging the JULES land model with observations and convection-permitting UK Climate Projections

13 Concerns exist about the viability of food security across Europe due to multiple, potentially adverse drivers. These include 14 economic, political and climate forcing factors, all of which require quantification. Here, we focus on the climate forcing, and 15 in particular, the soil moisture change component which crucially determines water availability for crop uptake. We estimate 16 future soil moisture levels at 34 sites of the UK COsmic-ray Soil Moisture Observing System (COSMOS-UK) network. We do 17 this by combining three platforms: the Joint UK Land Environment Simulator (JULES) land surface model, field-scale soil 18 moisture observations from the COSMOS-UK stations and 2.2 km convection-permitting UK Climate Projections (UKCP18). 19 We use COSMOS-UK data to optimise key soil moisture-related parameters in the JULES model, based on its performance in 20 the contemporary period. We then force the calibrated model with UKCP18 data to produce future soil moisture estimates. We 21 evaluate the modelled soil moisture for an average soil depth between 0 and 35 cm to match the depth of soil moisture 22 observations. Our main conclusions concern future soil moisture droughts which we compare with equivalent events in the 23 historical period, 1982-2000. We find that on average across all sites, there is an increase in the frequency of future extreme 24 soil moisture drought events of duration above 90 days. In 2062-2080, such frequency increase of between 0.1 and 0.6 events 25 per year (equivalent to at least 2 and up to 12 additional events in a 20-year period) is expected. We also show that, in 2062-26 2080, there is an increased risk of high or more intense soil moisture drought conditions in months between May and November, 27 with months between June and October being at especially high risk. The UKCP18 data corresponds to a high-emissions future 28 described by the RCP8.5 scenario. 29


Introduction
We calibrate the JULES model at 26 sites using soil 4 moisture observations from one selected full year at each 5 location (one site-year).The site-year selection (Suppl. 6 Section S1.1) is mostly based on strict completeness 7 criteria for precipitation data, recognising its importance as 8 a primary driver of soil moisture variations.We also avoid 9 peatland sites, as our modelling methodology is designed 10 for mineral soils, and woodland sites, as CRNS soil 11 moisture estimates are known to be less accurate there. 12 The future predictions of soil moisture are then    S4.This is consistent with an average decrease in precipitation 1 during this time of year and may also be partly due to an 2 increase in evapotranspiration due to higher temperatures.

3
In the winter, future precipitation is, on average, higher 4 than in the past period which reduces the negative soil  3).

33
Recent years have seen hotter and drier summer periods in 34 the UK (Met Office 2022; Turner et al 2021).Prolonged A c c e p t e d M a n u s c r i p t

SupplFigure 1 .
Figure 1.Map of COSMOS-UK sites used in this study.Each24

23 the data assimilation algorithm to produce a single 24 posteriorFigure 2
Figure 2 (blue) shows a schematic of the calibration 28

Figure 2 .
Figure 2. Schematic of the protocol for generating historical and future soil moisture estimates (for periods 1982-2000, 2022-2040 and 2062- modelled soil moisture at all 26 calibration 41 sites and the nine non-calibration sites.We use two 42 continuous years of COSMOS-UK observations, where 43 possible, to compare the measured and modelled daily soil 44 moisture.In the case of calibration sites, this includes the 45 calibration year.We use biases between the modelled 46 output and the observations, and the corresponding 47 unbiased root-mean-square errors as metrics to assess the 48 model (Suppl.Section S1.6).When comparing soil 49 moisture predictions using prior and posterior PTFs to 50 observations, there is an improvement in both metrics for 51 A c c e p t e d M a n u s c r i p t most of the sites following data assimilation (Suppl.
respect to the historical period, are 95 plotted in Figure 3. On average, across all 34 sites and 12 96 ensemble members, a decrease in soil moisture is expected, 97 especially in the summer, late spring and early autumn.

of soil moisture drought events 10 Figure 4
Figure 4 summarises the evolution of soil moisture 11

303.2 Probability of high stress months 52 Figure 6
Figure 6 shows probabilities of individual months being 53

90A c c e p t e d M a n u s c r i p t 1 Figure 3 .Figure 6 .
Figure 3. Projected changes in the modelled soil moisture and the UKCP18 precipitation for future time periods, 2022-2040 (labelled as

8
and years of a time period).The comparison metric is frequency difference (Eqn.4) in the former and probability difference (Eqn.6) in the 9 latter case.The minimum and maximum are calculated based on values obtained from 12 UKCP18 soil moisture simulations with and 12 10 without bias-correction (24 simulations in total) (Section 2.5.4).Probabilities used to produce probability differences have range between is the higher autumn stress which will 1 affect autumn sown cereals, for instance winter wheat, at 2 the beginning of their foundation phases, potentially 3 reducing yields.The autumn stress may also lead to the 4 prolongation of water-limited grazing productivity.The 5 drought conditions in the early spring and summer will 6 influence crops in their growing stages.Work in Slater et 7 al (2022) finds that on average, for broad UK regions, 8 climate change is likely to have beneficial impacts on 9 wheat yields.Nevertheless, the authors highlight that the 10 increased likelihood of prolonged, extreme weather will 11 generate conditions outside of the typical current climatic 12 envelope posing risks to future farming.13 Very dry soils will also have a negative impact on 14 grasslands which are important for biodiversity and as 15 grazing resources (Bengtsson et al 2019).The dry soils 16 may intensify heatwaves (Miralles et al 2019) and lead to 17 increasing wildfire risks, especially in the case of highly 18 organic soils and peatlands.19 Although our modelling strategy of first optimising the 20 PTF parameters provides an improvement in predictive 21 capability, some features of our reparameterization may 22 A c c e p t e d M a n u s c r i p t A c c e p t e d M a n u s c r i p t A c c e p t e d M a n u s c r i p t A c c e p t e d M a n u s c r i p t

Table 1 .
Meteorological variables and their units required for

Table S1 )
. We repeat the spin-up

Table 1 S5 ) .
For the posterior PTFs, most of the sites show negative 2 soil moisture biases, indicating an overall underestimation 3 of the modelled soil moisture.We note that one of the 4 calibration sites, Lizard (LIZRD), has a very significant 5 model bias and therefore we exclude it from the future soil 6 moisture analysis, leaving 34 sites in total for the future 7 runs.8 2.

4 Future soil moisture runs with the local 2.2
16 GC3.05).The same CPM structure and parameterisation 17 are used for all 12 simulations of the local UKCP18.18 However, parameterisations in regional and global models 19 differ between the ensemble members.The ensemble, 20 therefore, captures uncertainties due to alternative 21 parameter values describing the climate system and due to 22 interannual natural variability.The main advantage of the 23 CPM is that it allows the explicit representation of 24 convective storms, resulting in better estimates of the 25 statistical structure of localised, hourly rainfall.The CPM-26 generated data covers three time periods: 1 st December

Kendon et al 2023), but this
34was not the case during the time of producing the results.35Weuse the whole ensemble of the local UKCP18 data

Robinson et al 2020) (Suppl.
50and, therefore, is not included in the final analysis.We 51 perform the model spin-up three times and move to the 56 2, red).Given the dry model biases (Suppl.Table S5) 71 the fraction of available water (  ), accessible via roots, 72 falls below a certain threshold, commonly defined as 0.5 73 (Allen et al 1998; Grillakis 2019; Hunt et al 2009).It is 74 based on findings of (Baier 1969) which shows that 75 evapotranspiration is soil water-limited below this 76 threshold.We choose this generic threshold for our fixed 77 soil depth as a guide for the future PWS impact.With this, 78 we calculate a daily soil moisture index () (Hunt et al 83 where  is the soil water content,   is the field capacity 84 (FC) and   is the permanent wilting point (PWP).We 85 define PWP and FC as the soil water contents at a soil 86 matric potential of -1500 kPa and -33 kPa respectively 87 (Kirkham 2014).valuesdecreasing from zero 88 indicate increasing PWS up to PWP (when   = 0 and so 89  = −5).We apply three  bands to categorize the 90 intensity of different stress levels (Table2).

Table 2 .
Plant water stress categories.
92An alternative to the PWS index is a statistical index 93 (Samaniego

et al 2013; Sheffield et al 2004) which
16 duration and the total duration.Where the average  17 value of an event falls within the  range in Table 2, the 18 event is assigned the corresponding stress severity 19 category.For instance, if the average  value is -3, the 20 event is categorised as a high/ severe drought event.25For each site , ensemble member  and UKCP18 time 26 period  (1982-2000, 2022-2040 and 2062-2080), we 27 count the number of drought events,  _ , of a given 28 category.An average frequency of an event (per year, per 29 site) for each  and  can then be computed as 31 where   = 34 is the number of sites and   = 19 is the 32 number of years in a time period .We note that  _ can 33 be higher than one because more than one event of a given 34 category can occur within one year.When comparing 35 future  _ with the past period, we use an absolute 45 For each month  (January to December), site , 46 ensemble member  and time period , we count the 47 number of high stress months ( _ ).The probability 68 biases removed as a post-processing stage after running the 69 model ().They are defined as 70  _ = ′ _ −   , (7) 71 where  is the UKCP18 time period,  is site,  is the 72 ensemble member and   is the site-specific model bias 73 calculated in the contemporary period with respect to field-74 scale soil moisture observations (given in Suppl.Table S575 and defined in Eqn.S16 of Suppl.Section S1.6).We use 76 the bias-corrected soil moisture scenario alongside the non-77 bias-corrected version to address how the negative (dry)

Table 3 .
Temporal comparison of two types of events between time periods 2062-2080 and 1982-2000.The events are frequency of extreme 7 drought events above 90 days (per site, per year) and probability of a selected month being classified as a high stress month (across all sites