Soil moisture, starting from the terrestrial surface to the deepest soil layers, represents an invaluable parameter that has a fundamental role in the dynamics of the Earth system . Its variability results from the complex interaction between the atmosphere, vegetation, and soil processes. On the terrestrial surface, soil moisture is an essential component of the Earth surface energy budget, influencing the surface heat fluxes and evapotranspiration from land to atmosphere . From the climate point of view, showed that the number of hot days is largely determined by a precipitation deficit and, as a consequence, by small values of soil moisture. This coupling between atmospheric temperature and soil moisture is usually defined as soil moisture–temperature feedback, where drier soils determine a warmer atmosphere . Such feedback has the potential to exacerbate global warming by altering the surface heat balance . Other studies concentrated on the reciprocal influence between soil moisture and precipitation, which is referred to as the soil moisture–precipitation feedback. A number of processes may contribute to this feedback, acting both on a synoptic scale (by modifying synoptic settings and enhancing the large-scale transport of water vapor) and locally (by modifying boundary layer characteristics and influencing the organization of convection). Nevertheless, it remains challenging to ascertain an overall sign (positive or negative) for this feedback. The soil moisture available in the root zone is essential for vegetation and agriculture. Its values can be used as indexes for detecting hydrological drought . Through its impact on photosynthesis processes, found that the variability in soil moisture in climate model simulations drives 90 % of the inter-annual variability in the global land carbon uptake. The deep soil moisture is a fundamental feature with respect to the flux in the vadose zone towards shallow unconfined aquifers. For example, used the satellite-observed terrestrial water storage from the Gravity Recovery and Climate Experiment (GRACE) to determine the groundwater storage. Later, demonstrated that the initialization of seasonal forecast with such data improves groundwater forecasts in the USA. In addition, evaluated groundwater recharge from different land surface models and found that the seasonal cycle of simulated groundwater storage compared well with in situ groundwater observations.

Despite its fundamental role, in situ observations of soil moisture are scarce. Satellite and reanalysis products can provide a useful alternative to fill this gap. However, direct satellite observations are possible only for the first few centimeters below the surface . These surface observations can be propagated through the root zone by filtering operations, empirical models, or land surface models. Reanalyses offer a great alternative for studying soil moisture, and they are characterized by significant correlations with in situ observations. compared different reanalyses and found ERA5, the fifth-generation reanalysis of the European Centre for Medium-Range Weather Forecasts (ECMWF), to show the highest skill. Also, showed strong correlation between ERA5 flux and aquifer water table observations. The same was found by between the Global Land Data Assimilation System (GLDAS) and multi-satellite soil moisture anomalies. The utility of soil moisture data from land surface models employed within atmospheric general circulation models hinges not on the soil moisture value itself but on its temporal variations, which are particularly well represented when compared to observations . By analyzing different reanalysis and land surface models with respect to observational data in central Italy, found, on average, the best performances of the ERA5 reanalysis with respect to other well-established reanalysis. As a further feature suggesting the use of ERA5, its good performance in terms of water budget evaluation in closed lakes must be mentioned . For these reasons, in this paper ERA5 reanalysis, and its land component ERA5-Land , will be used as a reference soil moisture condition.

There is high confidence that the Mediterranean region will suffer from a larger aridity and an increase in hydrological droughts . Moreover, aridity can heavily impact the snowmelt recharge of the aquifers in the mountain ranges of the Mediterranean area, further affecting hydrological droughts as well as vegetation phenology . In this context of climate change, sub-seasonal to seasonal (S2S) forecasts are a fundamental tool for adaptation strategies, especially regarding water resources management. The accuracy of the S2S forecast system relies on the simulation of the response of the atmosphere to the slowly varying states of the ocean and land surface . demonstrate how SEAS5, the seasonal forecasting system of ECMWF, has a particular strength in the prediction of El Niño–Southern Oscillation (ENSO). found globally useful forecast skill when predicting the occurrence of marine heat waves (prolonged period of extremely warm sea surface temperature) for the two seasons after the forecast initialization date. analyze the forecast skill of SEAS5 for three key climate variables (temperature, precipitation, and wind speed) over Europe and found such forecasts useful for climate services after a proper bias-adjustment method was applied. found that seasonal forecasts from the SEAS5 system starting from the early May can provide useful information about the probability of occurrence of European summer heat waves. A recent study over the Mediterranean region by found that individual seasonal forecasting systems outperform elementary forecasts of precipitation anomalies based on persistence or climatology. However, the added value is not uniform over the Mediterranean area. The same inhomogeneity and potential usefulness of seasonal forecast in the Mediterranean area were found also by for agriculture and forestry. However, the same analysis could bring different results in regions with marked orographic impact and land–sea contrast such as the Mediterranean region. show that seasonal climate forecast by SEAS5 provides useful information for decision-making processes in the European winter-wheat-producing sector, by analyzing minimum and maximum daily temperature and daily total precipitation. In particular, drought events were better predicted than excessive wetness periods. On the scale of S2S forecasts, soil moisture is one of the most impactful land parameter and is crucial for the forecast skill . found that land initialization contributes to approximately a third of the total soil moisture predictability, while the remaining part is attributable to ocean conditions. Moreover, they found that the same initialization can provide limited skill in the precipitation forecast but enough skill in the soil moisture forecast. This result suggests that skillful seasonal prediction can be made on drought occurrence focusing on the soil state. This can be attributed to reduced variability in soil moisture which is an order of magnitude smaller than that of rainfall. The study by in North America suggested that this source of predictability is connected to the soil moisture reemergence process, in which moisture anomalies stored in the deep soil layer can “reemerge” to the surface, restoring the earlier root-zone anomaly and providing a year-to-year soil moisture memory. found that the seasonal forecast of standardized soil moisture anomalies (SSMAs) performs better than the forecast of precipitation using the CFSv2 (Climate Forecast System) in South America. Moreover, the performance was found to be higher for austral winter than summer and for dry events rather than wet episodes. This result shows the value of seasonal forecasts of SSMAs for their use for agricultural drought monitoring. A recent study by found that the Community Land Model (CLM5), forced by SEAS5 seasonal forecasts, satisfactorily reproduces the inter-annual variation of crop yield and also the high- and low-yield seasons in Germany and Australia. However, a systematic bias of soil moisture was found when comparing with satellite observations. Most of the above results apply to large continental regions in North and South America, while in Europe seasonal forecast performances are mostly evaluated for surface atmospheric variables. Accordingly, to fill this gap, this paper focuses on evaluating seasonal forecasts of soil moisture for water resources management, with particular attention to wet and dry events. The key questions addressed in this study are as follows: i

Can the seasonal forecast over the Central Mediterranean be used to predict the soil moisture behavior?

Monthly mean values of the soil moisture, θ, from seasonal forecasts are validated against monthly mean values of soil moisture from both ERA5 and ERA5-Land reanalysis. Both datasets are interpolated over a regular grid of 0.125° of horizontal resolution. The number and the depth of soil layers in each column are the same in SEAS5, ERA5, and ERA5-Land: four soil layers at a depth of 7 cm (soil layer 1), 28 cm (soil layer 2), 100 cm (soil layer 3), and 289 cm (soil layer 4), respectively. The soil type, when passing across different grids, is taken as the prevailing soil type in order to preserve soil hydraulic properties . The evaluation of seasonal forecasts and also the discrimination of dry and wet periods are performed over the standardized soil moisture anomaly (SSMA). Following the approach by , the SSMA is calculated at each grid point (i,j), month (m, from January to December), year (y), and soil layer (k) as 1SSMAk(i,j,m,y)=θk(i,j,m,y)-θk‾(i,j,m)σθk(i,j,m) where ⋅‾ is the time average operator over the whole reference year, and σ is the standard deviation operator. The time period considered for the forecast validation spans 20 years from 2001 to 2021, while the reference time period considered for evaluating the monthly climatology and standard deviation ranges from 2001 to 2016. The same reference period is also considered for the bias adjustment of seasonal forecast. The method used in this work is the simple mean and variance adjustment (MVA) method as described by . Each member mean and variance over each grid point is bias-adjusted with respect to the ERA5 observation mean and variance over the period 2001–2016, in the following form: 2θk′(l,m,y,n)=(θk(l,m,y,n,)-θk^(l,m))σobs(m)σf(l,m)+θkobs‾(m), where l is the forecast lead time, m is the forecast month, n is the index representing each ensemble member, θk^(l,m) is the ensemble and time average of forecasts for each lead time and month over the reference period, σf(l,m) is the standard deviation of the complete ensemble for each lead time and month over the reference period, θkobs‾ is the time average of all observation for the considered month over the reference period, and σobs(m) is the standard deviation of all observations for the considered month over the reference period. The bias adjustment is computed for each forecast lead time (l, from 1 to 6 months). In this way, the bias and variance adjustment take into account both the forecast month and the forecast lead time, which has been found to be beneficial in previous work by . Although the simplest among different methods, demonstrated that MVA methods represent a good compromise between computational cost and performance. This is particularly relevant, since the final aim of this study is to develop real-time applications for climate services. The final effect of the bias adjustment on the forecast ensemble mean is shown in Fig. , where the Umbria reference grid point (see Fig. ) is shown as an example.

In order to analyze the variability in soil moisture and to compare it across different soil layers, we compute the memory timescale of each layer as the e-folding time of the temporal autocorrelation function. The autocorrelation is evaluated by calculating the Spearman correlation coefficient, shifting the time series by a temporal lag that is between 0 and 365 d. The corresponding time when the correlation coefficient becomes lower than e-1 is taken as the memory timescale of that grid point and soil layer. This timescale is evaluated by considering ERA-Land daily mean soil moisture data over all the domain. An example of this procedure for the Umbria reference point is reported in Fig. . As expected, the deeper the soil layer, the longer the memory timescale. This behavior can be observed for all grid points of the study domain, as will be shown later in Sect. . A further interesting feature is pointed out by the autocorrelation structure. Precisely, after an initial decay (as expected), the autocorrelation shows a rebound with a secondary statistically significant maximum at a lag of approximately 300–350 d. Such a rebound could be indicative either of the seasonal cycle or of the reemergence of soil moisture anomalies as hypothesized by . However this behavior is not representative of all regions and therefore merits further explorations in future research.

The performance of SSMA forecasts is evaluated through three different metrics, two deterministic and one probabilistic. First, the average magnitude error of SSMA ensemble mean is evaluated trough the root-mean-squared error (RMSE). This metric, by definition, puts greater influence on large errors than smaller errors. The RMSE is commonly used in both weather forecast performance assessment and seasonal streamflow forecasting . Successively, the anomaly correlation coefficient (ACC) is used to measure the correspondence between forecasted and observed ensemble mean SSMA. The ACC is the most widely used skill metric for evaluating the skill of deterministic forecast , and it is not sensitive to forecast bias. Then, the ability of SEAS5 ensemble system to discriminate between different event type is measured by the area under the relative operating characteristic (ROC) curve . An example of the procedure for the evaluation of the ROC curve is reported in Fig.  for the Umbria reference point. In particular, dry and wet events have been defined as those with the SSMA being smaller or larger than 1, respectively (see Fig. a). An ensemble (probabilistic) forecast will have a certain probability of detecting that event. Using a set of increasing probability thresholds, we build a contingency table (true and false positive, true and false negative). Then we calculate the true positive rate (or probability of detection) and the false positive rate (or false alarm rate) for each probability bin. The ROC curve is obtained by plotting the true positive rate against the false positive rate as shown in Fig. b for the different probability bins. For each probability bin, the true positive rate should be larger than the false alarm rate, otherwise the forecast is not useful. Therefore, the area under the ROC curve can be used as a score to evaluate the usefulness of a forecast. The diagonal line in Fig. b indicates no skill (ROC area close to 0.5), while the perfect forecast would have a ROC area equal to 1. Each metric has been evaluated for the different forecast lead times. The first two metrics (RMSE and ACC) are evaluated by considering the ensemble mean SSMA values, while the latter (ROC) is evaluated by considering all the ensemble members. All metrics calculations rely on the xskillscore Python package (https://github.com/xarray-contrib/xskillscore, last access: 17 February 2025).

In this section, the obtained results are analyzed in terms of soil moisture variability and forecast performance metrics.

In this section, we show a possible application of seasonal forecasts of soil moisture for groundwater management. Figure a shows the water table level observations (expressed as standardized anomalies with respect to their mean and standard deviation) in two different locations of Italy, the Umbria and Veneto region, in the central and northern part of Italy, respectively (Fig. ). The monitored aquifers are selected to be shallow (depth smaller than 10 m) and unconfined in order to be directly influenced by atmospheric conditions rather than other groundwater processes . From such observations we only detect three dry periods (in 2007, 2012, and 2017) where the standardized anomalies of the water table level were less than -1 for both regions. On the other hand, wet periods in the water table observations, where values larger than 1 are observed, seem to happen more frequently.

The water shortage in 2007, 2012, and 2017 in different regions of Italy is an indication of the synoptic-scale character of such drought periods. The variability in water table level is well captured by the variability in deep soil moisture as extracted from ERA5 reanalysis, as shown in Fig. b. In 2007, 2012, and 2017, negative anomalies are also observed for SSMA4 in both the analyzed piezometers, with the 2017 anomaly being weaker with respect to the others. In the analysis below, we focus on the 2012 dry period and the following wet period in order to test the ability of seasonal forecasts to predict such events. Figure a, b, and c show the spatial distribution of SSMA4 over the Central Mediterranean in June 2012, December 2012, and June 2013, respectively. These periods are taken as a reference for the start of the dry period, the end of the dry period, and the start of the wet period, as observed in Fig. b for northern–central Italy.

June 2012 is characterized by a large negative anomaly over all the domain expect for the Alps, Sicily, Tunisia, and the Adriatic coast of the Balkan Peninsula. Seasonal forecasts for lead time 1 month predict smaller negative anomalies over central Italy and the Balkan Peninsula while largely underestimating the positive anomalies over the Alps (Fig. d). The forecast slightly improves in northern Italy and Balkan Peninsula going to lead times of 3 and 6 months (Fig. g and l), while it gets worse for Sicily and Tunisia. December 2012 shows a similar spatial distribution of SSMA4 except for larger positive anomalies on the Alps, the southwestern coast of Italy, and the southwestern coast of the Balkan Peninsula (Fig. b). Also, the amplitude of negative anomalies of SSMA4 decreases in central and northern Italy. The seasonal forecasts perform well in central and northern Italy, on the southeastern coast of Italy, in Sicily, and in Provence for all lead times (Fig. e–h–m). However, the large increase in positive anomalies over the Alps, the western coast of the Balkan Peninsula, and the Tyrrhenian coast of Italy was not detected. The larger negative SSMA4 in the internal regions of the Balkan Peninsula was also not detected. The wet period of June 2013 especially involved the northern part of the domain, with large positive anomalies of SSMA4 (Fig. c). Seasonal forecasts show a good performance in general, especially in central and northern Italy at a lead time of 1 month (Fig. f), while they tend to underestimate such a positive anomaly at larger lead times, especially over Hungary (Fig. i–n).

This study provides a first assessment of seasonal forecast of soil moisture for the Central Mediterranean region. The seasonal model analyzed in this study is SEAS5, whereas the reanalysis is ERA5-Land, both produced by the ECMWF. ERA5-Land is considered a reference dataset for soil moisture observations. A total of 25 member seasonal forecasts with lead times from 1 to 6 months have been analyzed from 2001 to 2021, by considering the hindcast period 2001–2016 for climatology. In this reference period, the standardized soil moisture anomaly (SSMA) was evaluated, and the forecasts were bias-adjusted through the mean-variance adjustment method. Then, the root-mean-squared error (RMSE) and the anomaly correlation coefficient (ACC) were evaluated for SSMA for all soil layers considered in ERA5-Land. To test the ability of the forecast to discriminate between dry and wet events, the relative operating characteristic (ROC) area has been calculated. Finally, a case study of the dry and wet periods during 2012–2013 was studied in detail, to show the potential usefulness of the seasonal model. The outcomes of the research can be summarized as follows.

As indicated by the RMSE, the average magnitude of the forecast errors decreases as we go deeper into the soil. Only in the deepest soil layer at 289 cm depth can the RMSE reach values below 0.5, even for a lead time of 6 months. However, this is only valid over certain regions like central and northern Italy, Hungary, some internal regions of the Balkan Peninsula, and the Provence region. The RMSE remains too high in other regions, even when only considering the deepest layer. Significant values of the ACC, with values larger than 0.8, can be found over the mentioned regions even at a lead time of 6 months. The analyzed performance depends on the memory timescale of the soil layer: the higher the memory timescale, the higher the forecast performance. The main physical factors affecting the spatial variability in memory timescale are various: soil depth, orographic complexity, local climatology (e.g., soil aridity, mean precipitation), and soil hydraulic properties. In general, we found better forecast performance in the deepest soil layer and in regions with low orographic complexity, corresponding to regions with larger memory timescale. This is in agreement with previous studies on the spatial variability in soil moisture memory. For example, found a large sensitivity of soil moisture memory to soil hydraulic parameters (e.g., ) and found a longer memory in the deepest soil layers. The dependence on soil depth was ascribed to the smallest influence of the throughfall precipitation, which is partly absorbed by evapotranspiration before penetrating into the deepest soil layers. Moreover, analyzed the influence of altitude, topography, and dryness index on soil moisture memory timescales, finding that memory timescales decrease with elevation and increase with topography (measured by a topographic index which is a function of the slope) and aridity. Our study identified comparable signals in the Central Mediterranean. However, further investigations are required to ascertain which factors (soil properties, altitude, orographic complexity, climate) are most influential in shaping the soil moisture memory in a given region. Such an investigation could inform a more robust modeling approach, incorporating additional parameter uncertainty into the forecasting system, which may ultimately enhance the skill of seasonal forecasts . The ability of seasonal forecasts to detect wet and dry events exhibits a large variability within the domain. However, a ROC larger than 0.8 can also be found in certain regions for the deepest soil layer for a lead time of 6 months. This indicates that in those regions, like Provence, central and northern Italy, the southeastern coast of Italy, and the internal regions of the Balkan Peninsula, seasonal forecasts can be used to detect such events in advance. The area under the ROC curve for dry and wet events in the two uppermost soil layers is about 0.5 when lead times exceeding 3 months are considered. This suggests that seasonal forecasting is not a reliable method for predicting the evolution of upper soil moisture beyond 3 months. A small ROC area for dry and wet events is found at a lead time of 6 months, especially in mountainous regions (Alps and Dinaric Alps), confirming the spatial variability already found for RMSE and ACC indicators. In general, for all soil layers, dry events are generally better captured than wet events. From a water management point of view, this indicates that information provided by seasonal forecasts on soil moisture should be trusted more for supporting drought-risk management rather than flood-risk management. The most useful forecasts are produced for events where the antecedent and present conditions are aligned (e.g., dry after dry, wet after wet). This further validates the significance of soil moisture memory and soil moisture pre-condition for the predictability of the system.

In the areas with large correlation coefficients, larger correlations are found between April and October, while a minimum correlation is found in December and January. Such a feature is of great relevance in terms of water resources management as the critical period is late spring and summer, when the water demand is the largest in the year for both agriculture and civil activities. As an example, the case study of 2012 drought period shows how the SEAS5 model is able to predict such an event for central and northern Italy 6 months in advance. Moreover, the strict connection between the deepest soil moisture and the water table of shallow unconfined aquifer in Italy highlights the large potential usefulness of seasonal forecasts of soil moisture for water management purposes. A local water management service, especially located in the most effective areas, could monitor the forecasted soil moisture anomalies across all soil layers, as publicly provided by the Copernicus Climate Data Store. The seasonal forecasting system can provide a probability of either a wet, dry, or normal month at different lead times, thus assisting in the decision-making process for the management of drought or flood risks. Moreover, groundwater models or simpler methods such as those in could be run, starting from forecasted soil moisture products for monitoring groundwater levels in unconfined aquifers. In this case, when using these data in very local application, an evaluation of the influence of irrigation input could be of great importance. However, the volume of water used for irrigation – a critical quantity towards water resources management – is a data that is very difficult to find for a number of reasons. One of the most important is the poor measuring instrumentation installed in irrigation systems. Consequently, it is difficult to estimate the contribution to soil moisture. The irrigation volume being equal, the dispersion towards the aquifer depends on the type of irrigation practiced. Maximum dispersion occurs in flowing systems, whereas in the case of drop irrigation in pressurized networks, dispersion can be considered negligible. For the Veneto irrigation systems, due to the extreme relevance of aquifers, reliable quantitative assessments are available . These irrigation systems, active for the entire year, are supplied by surface water and are of the flowing type with unlined channels in 70 % of the cases. Groundwater withdrawals are carried out by water utilities for drinking water use. On the basis of the data provided by the land reclamation consortia and water utilities, it is shown that the entity of dispersion of the irrigation systems, minus withdrawals for drinking water use, is comparable to that of effective rainfall . In the Umbria region, even if analogous documentation is not available, the same situation can be assumed. However, since in this study the analysis is focused on the soil moisture anomaly, the contribute of the irrigation volume is not effective, as it is almost constant over the year. Conversely, in cases where irrigation was only active in a few months, its effect should be taken into account provided that data availability allows. To refine the proposed approach, two possible paths can be followed in future research. The first is to use different seasonal forecasting models and different reanalysis and observation products (e.g., MERRA-2 and GLEAMv3 ). The second path is to analyze the behavior of the autocorrelation function of soil moisture anomalies across different soil layers in more detail. This would evaluate the role of the seasonal cycle and the reemergence of soil moisture as hypothesized by . Then, this would allow us to compare the seasonal forecast performance with that obtained by memory-prediction models, following the approach proposed by .