Modelling Mediterranean agro-ecosystems by including agricultural trees in the LPJmL model

In the Mediterranean region, climate and land use change are expected to impact on natural and agricultural ecosystems by warming, reduced rainfall, direct degradation of ecosystems and biodiversity loss. Human population growth and socioeconomic changes, notably on the eastern and southern shores, will require increases in food production and put additional pressure on agro-ecosystems and water resources. Coping with these challenges requires informed decisions that, in turn, require assessments by means of a comprehensive agro-ecosystem and hydrological model. This study presents the inclusion of 10 Mediterranean agricultural plants, mainly perennial crops, in an agro-ecosystem model (Lund-Potsdam-Jena managed Land – LPJmL): nut trees, date palms, citrus trees, orchards, olive trees, grapes, cotton, potatoes, vegetables and fodder grasses. The model was successfully tested in three model outputs: agricultural yields, irrigation requirements and soil carbon density. With the development presented in this study, LPJmL is now able to simulate in good detail and mechanistically the functioning of Mediterranean agriculture with a comprehensive representation of ecophysiological processes for all vegetation types (natural and agricultural) and in a consistent framework that produces estimates of carbon, agricultural and hydrological variables for the entire Mediterranean basin. This development paves the way for further model extensions aiming at the representation of alternative agroecosystems (e.g. agroforestry), and opens the door for a large number of applications in the Mediterranean region, for example assessments of the consequences of land use transitions, the influence of management practices and climate change impacts.


Introduction
The Mediterranean region is a transitional zone between the subtropical and temperate zones, with high intra-and inter-annual variability (Lionello et al., 2006).This region has been identified as one of the regional climate change hotspots, with a high likelihood of experiencing more frequent and more intensive heat waves, often combined with and strengthened by more intensive and longer droughts (IPCC, 2012;Kovats et al., 2014;Diffenbaugh and Giorgi, 2012).This will likely have adverse implications for the food and energy producing sectors as well as for human health, tourism, labour productivity and ecosystem services (Kovats et al., 2014;Skuras and Psaltopoulos, 2012).However, climate change is only a part of the challenges that the Mediterranean region will face in the near future.Environmental degradation involving soil erosion, biodiversity loss and pollution is negatively affecting natural and societal systems and is expected to intensify even more in future due to urbanisation, industrialisation and population growth (Doblas-Miranda et al., 2015;Lavorel et al., 1998;Scarascia-Mugnozza et al., 2000;Schröter et al., 2005;Zdruli, 2014).
Most aspects of climate change and environmental degradation will affect the Mediterranean agricultural sector di-Published by Copernicus Publications on behalf of the European Geosciences Union.
rectly.Agriculture plays a very important role, not only for food security in the region itself, but also through its economic integration in other regions, such as through the significant export of products to the rest of Europe (Hervieu, 2006).Agriculture therefore plays a key role for the national economies, making a part of the rural population rely on it for their livelihood and creating linkages with other issues and sectors, such as culture and tourism (Verner, 2012;Hervieu, 2006).Human population growth and socioeconomic developments on the eastern and southern shores as well as the already high dependence of the region on the international food markets will increase the need for local food production.Additionally, potential resource allocation trade-offs, especially for water and land, will put Mediterranean agriculture under increased pressure, calling for more and more efficient production practices (Verner, 2012;World Bank, 2013).
To support adaptation and mitigation efforts for climate change and environmental degradation, Mediterranean-wide assessments of the state of agriculture and the likely consequences of global change are required.These would have to be complemented by analyses of the potential developments and future difficulties of the agricultural sector and its interactions with the environment.The large-scale character of such assessments and the necessity of looking into possible future scenarios require the utilisation of modelling tools that cover the essential characteristics of the dominant agroecosystems in the region.At present, no suitable modelling framework for this task exists.Given the range of conditions in the region, such a tool should be process-based and integrate the major crop types, grasslands and natural vegetation, taking into account the carbon cycle and hydrology of them.Notably, the presence of perennial, woody species is a characteristic of Mediterranean agro-ecosystems, and they deliver 45 % of agricultural outputs (Lobianco and Esposti, 2006).Existing crop models have implemented some tree crops, and in some cases applications in Mediterranean environments (mostly small scale) were published.For example, the STICS crop model has been used to simulate the growth of vineyards and apple trees (García de Cortázar-Atauri, 2006;Nesme et al., 2006;Valdés-Gómez et al., 2009); and in the CropSyst model, pears, apples, vineyards and peaches are included (Marsal et al., 2013(Marsal et al., , 2014;;Marsal and Stöckle, 2012).Other modelling frameworks offer general and specific formulations for horticultural systems that have been applied in other regions, mainly in Anglo-Saxon countries.This is the case for the EPIC/SWAT/SWIM families (Neitsch et al., 2004;Gerik et al., 2014) for cotton and apple, and for the APSIM model for cotton and vineyards (Holzworth et al., 2014).In California, another region with a Mediterranean climate, there is a dynamic modelling community assessing climate change impacts on horticulture by process-based (Gutiérrez et al., 2006) and empirical models (Lobell et al., 2007).At the global scale, the GAEZ approach offers potential growing areas for citrus, olives and cotton (IIASA/ FAO, 2012).The GCWM model is probably the most complete model in terms of perennial crops, comprising citrus, cotton, date palm and grapes (Siebert andDöll, 2008, 2010).Other authors have developed models for fruit trees, such as the inclusion of kiwi, vineyards and apples in the SPASMO model (Clothier et al., 2012), walnuts in CAN-WALNUT (Baldocchi and Wong, 2006) and date palm by Sperling (2013).A model by Villalobos et al. (2013) focuses on transpiration of apricot, apple, citrus, olive, peach, pistachio and walnut trees.
The goals of these applications are diverse, from simple reproduction of experiments, via epidemiological analysis (causes and patterns of diseases), to the simulation of phenological change and the influence of management practices on agricultural production.Concerning the impacts of climate change, Moriondo et al. (2015) presented a detailed review of empirical and process-based models for olives and vineyards.They concluded that process-based models are better suited for climate impact studies, but they have to be completed and improved to account for the perennial nature of these crops, the effect of higher CO 2 on the atmosphere, dynamic growing periods and the effect of management practices.
Reviewing the existing studies reveals two important points.First, there is no single model or model family comprising all major agricultural plants of the Mediterranean region.Second, there is no model combining dynamic simulation of natural vegetation and agro-ecosystems for the Mediterranean region.Some models are very advanced in some processes, like the case of STICS for biochemical cycles.Other models have unique features, such as the detailed consideration of hydrology in the unsaturated zone, salt and leaching transport in the WOFOST model coupled with the SWAP and PEARL models (Kroes and Van Dam, 2003).The CENTURY model, which focuses on soil organic carbon computation, offers a forest module and general formulations that can be adapted to horticulture, but only one smallscale application was presented for the Mediterranean region (Álvaro-Fuentes et al., 2011).Without the integrated modelling of natural vegetation and agro-ecosystems in a comprehensive framework, there are many questions that cannot be answered.Notably, some of them are of extreme relevance for the Mediterranean region and concern water requirements and availability for the agricultural sector, sustainable food production potentials under climate change, environmental consequences of land use (including biodiversity loss), and soil carbon sequestration patterns, including responses to land use change.
To better assess the potential responses of Mediterranean agro-ecosystems to these forcings, we have extended the representation of Mediterranean agriculture in the Lund-Potsdam-Jena managed Land (LPJmL) model.LPJmL computes the dynamics of natural vegetation, annual crops and natural grasslands by considering carbon pools and fluxes, hydrological variables, and coupled photosynthesis and transpiration (Sitch et al., 2003;Gerten et al., 2004;Bondeau et al., 2007).The model has undergone major developments with respect to the hydrology, including a river routing and irrigation scheme (Rost et al., 2008), the management of dams and reservoirs (Biemans et al., 2011), and a five-soillayer hydrology (Schaphoff et al., 2013).The representation of agricultural systems has also been improved by including bioenergy systems (tree and grass bioenergy plantations, jatropha, sugar cane; Beringer and Lucht, 2008;Lapola et al., 2009).An agricultural management module has been added, representing the combined influence of management practices on plant and stand development (e.g.fertiliser inputs, mechanisation, use of high-yielding varieties, weed and pest control, etc.; Fader et al., 2010), and better representation of sowing dates and multiple cropping systems (Waha et al., 2012(Waha et al., , 2013)).LPJmL is widely recognised as a state-of-theart agro-ecosystem and hydrology model and has undergone the broadest possible range of validation efforts against experimental and observational data.In a recent intercomparison of global hydrological models and global gridded crop models for the assessment of future irrigation water availability, Elliot et al. ( 2014) indicated that LPJmL is unique in that it performs well in both categories.LPJmL is intensively used at the global and macro-regional scales in various research fields, particularly for questions related to future food security, land use change, and adaptation to climate change (Gerten et al., 2008;Rost et al., 2009;Lapola et al., 2010;Fader et al., 2010Fader et al., , 2013;;Waha et al., 2013;Müller et al., 2014).
Since Mediterranean-specific crops have been lacking in LPJmL, we present here an extension that includes 10 new crop functional types that are especially important in the region.Most of these may be called "agricultural trees": nut trees, date palms, citrus trees, orchards, olive trees, grapes, cotton, potatoes, vegetables and fodder grasses.Their inclusion made it possible to account for ∼ 88 % of irrigated areas of the Mediterranean instead of ∼ 50 %.
The following section outlines the methodology applied, including the compilation of a new input data set with land use patterns which was needed for the validation and application of the model, the description of the modelling approach and parametrisation of the new crops, as well as the computation of irrigation requirements and soil carbon densities.The results section details exemplarily the performance of the model in simulating yields, soil carbon and irrigation water requirements.Finally, the paper is closed by a discussion on perspectives for future developments, potential applications and further refinements.

Methods
As a function of climatic conditions and agricultural management, LPJmL simulates, spatially explicitly and at a daily to yearly temporal resolution, growing periods (sowing and harvest dates), net and gross primary productivity, carbon sequestration in plants' compartments and soil, heterotrophic and autotrophic respiration, agricultural production, as well as a number of hydrological variables, such as runoff, soil evaporation, plant transpiration, plants' interception, percolation, infiltration, river discharge, irrigation water requirements, water stress and soil water content.Required inputs are (a) gridded, monthly climate variables (temperature, cloudiness interpolated to daily values and precipitation and rainy days converted through a weather generator in daily values); (b) atmospheric CO 2 concentrations; (c) gridded soil texture as described in Schaphoff et al. (2013); and (d) a gridded data set of land use patterns prescribing which crops are grown where and whether they are irrigated or rain-fed.
For the present study, we used climate inputs at 30 arcmin spatial resolution and global CO 2 concentrations derived from the CRU 3.10 data sets (Harris et al., 2014).The land use patterns for the crops in LPJmL had to be compiled from different sources, as explained in the following section.The model was spun up for 5000 years with dynamic natural vegetation in order to bring the carbon pools into equilibrium and additionally for 390 years with natural and agricultural vegetation.The spun-up simulations were followed by a transient run from 1901 to 2010 using the land use patterns described in the next section.

Land use patterns for use in LPJmL
LPJmL needs irrigated and rain-fed physical (as opposed to harvested) areas for each simulated crop.Physical and harvested areas differ through multiple cropping practices; that is, when one area is used twice in a year, the harvested area is twice as large as the physical area.For the present model development a new land use data set had to be compiled from different sources: Portmann et al. (2011), hereafter MIRCA, Monfreda et al. (2008), hereafter MON, Klein Goldewijk et al. (2011), hereafter HYDE, andRamankutty et al. (2008), hereafter RAM.
The first step for the compilation of the land use data set was determining the harvested areas of all LPJmL classes, including the new Mediterranean crops, for the present time.For crops present in MIRCA, which differentiate between irrigated and rain-fed areas, all values in that study were used directly.For the missing crops, MON corresponding classes (see Table S1 in the Supplement) were compared at grid-cell level with MIRCA classes "other perennial" and "other annual", thereby splitting the harvested areas into the rain-fed and irrigated parts.This procedure was done for olives, noncitrus orchards, nut trees and vegetables.For the first three groups, MIRCA class "other perennial" was used for the grid-cell-specific splitting between rain-fed and irrigated areas.For vegetables, rain-fed and irrigated areas were derived through comparison with the "other annual" class.
Large inconsistences were found at the grid-cell level between MIRCA and MON, for example, cases where a single crop area in MON was larger than the sum of the rain-fed and irrigated corresponding "other" classes in MIRCA, and the absence or small extent of the class irrigated "other peren- nial" in areas that intuitively may be assumed to be irrigated, like in the case of orchards and olives in Egypt.The author of MIRCA (F.Portmann, personal communication, 2013) assumes that most of these inconsistences are due to scale, the grid-cell differences being potentially large but presenting a good agreement at the administrative level.Grasslands, representing meadows, were taken directly from RAM and assumed to always be rain-fed.
After deriving harvested areas for around the year 2000, the calculation followed the flowchart shown in Fig. S1 in the Supplement.Harvested areas were compared with cropland and grassland areas from RAM in order to exclude multiple cropping and derive physical cultivation areas.Decadal cropland data from HYDE were interpolated to derive annual values and then used for extrapolating the land use patterns of ∼ 2000 to the past, until 1700 (see below, Eqs. 1 to 5).Historical irrigation fractions were determined as explained in Fader et al. (2010).
Due to inconsistences between HYDE and the data set combining MIRCA, MON and RAM (hereafter MMR), proportionally changing cell fractions using the HYDE historical trend would have given an unrealistic overall trend of cropland areas.For this reason, crop-specific bias corrections had to be performed in between, as follows.
First, the global (g) area difference (D) was calculated: This procedure yields a gridded global data set at 30 arcmin spatial resolution of the cultivation areas of 24 crops from 1700 to 2010.Table 1 shows the resulting areas for each crop class of LPJmL.Expanding LPJmL for modelling Mediterranean crops made it possible to account for ∼ 88 % of irrigated areas of the Mediterranean instead of ∼ 50 %.For rain-fed areas, the improvement was from ∼ 21 to 73 %.The remaining areas are mostly fallow land and were included in the "others" model class which is parametrised as grasslands.
There is a general lack of information about the planting areas for agricultural trees at national and subnational level.Nevertheless, it was possible to make two comparisons.First, EUROSTAT offers information about olive tree areas at country level for eight countries of the northern Mediterranean.Our results are in good agreement with their numbers, with a mean absolute percent error of 26 % (MAPE, calculated as the sum of percentage differences between their values and our values, divided by their values, and finally multiplied by 1 over the sample size).Second, national harvested areas as reported by FAOSTAT for dates, olives, cotton seed, grapes and potatoes as an average of 2000 to 2009 could be compared with our data set.The agreement is high (MAPE < 30 %) for all classes except for olives and dates (MAPE 47 and 46 %,respectively) where our data set has mainly smaller areas.This is due to the fact that MIRCA, RAM and MON have been compiled for the year 2000 and the FAO shows strongly an accelerated expansion of areas from 2000 to 2010, for example 54 % for olives in Morocco and 45 % for dates in Egypt and Turkey.Overall, the input data set compiled here has no better alternatives and appears to be broadly suitable for applications until newer versions of the land use data used as sources are released.

Implementation, calibration and parametrisation of Mediterranean agricultural trees and crops
Twelve crops were already present in LPJmL (temperate cereals, rice, tropical cereals, maize, temperate roots, tropical roots, pulses, rapeseed, soybeans, sunflower, sugar cane, others).In this study we included nut trees, date palms, citrus trees, orchards, olive trees, grapes, potatoes, cotton, vegetables and fodder grasses.
For each new crop, a representative species was selected for which the parameterisation was performed (see Table 2 for details on the parameters described in the following sentences).Potatoes were introduced as an annual crop following the approaches as described in Bondeau et al. (2007) for other annual crops.Potatoes are planted in early spring in cooler climates and late winter in warmer regions (FAO, 2008).In LPJmL they are sown each year in the areas indicated by the land use input, taking into account the seasonality of rainfall and temperature and the experience of farmers (see Waha et al., 2012).In the case of no water stress, leaf area index (LAI) development follows a prescribed curve (as in SWAT) with inflexion points according to the parameters shown in Table 2 (Phu_1/2, Lmax_1/2, Phusen), but LAI is reduced in the case of water stress by scaling it to the difference between atmospheric demand and water supply.Phenology and maturity are modelled after the heat unit theory: when the accumulated difference between daily temperatures and base temperature reaches a prescribed total growing degree amount (called hereafter PHU for potential heat units), then the potatoes are ripe and are harvested.Absorbed photosynthetically active radiation drives assimilation.Carbon allocation to different parts of the plant is a function of PHU development.The PHU parameter used depends on the mean temperature for spring varieties and on the sowing date for the winter/autumn varieties, and ranges from 1500 to 2400 • Cd (with lower PHU in cooler climates).Agricultural trees -including grapes and cotton which are modelled in the present study as small trees -are planted as samplings with 2.3 g carbon in sapwood and a LAI of 1.6 in the growing areas indicated by the land use input.Each agricultural tree has a country-and tree-specific planting density, and a tree-specific parameter determines the number of years that are needed for trees to grow before the first harvest.The latter parameter depends not only on the varieties used and on the biophysical situation, but also on management, especially on the usage of fertilisers and irrigation.There are insufficient quantitative data on this issue, and we therefore assumed this parameter to be 4 years for all agricultural trees.After these years, a plant-specific portion (HI, the "harvest ratio" or "harvest index") of the net primary productivity (NPP) of the tree is harvested every year.Thus, fruit growth is represented by a carbon accumulation that equals the multiplication of HI and NPP.An additional tree-specific parameter determines the replanting cycles of trees.Since there are no data available on this, we assumed that plantations are renewed (replanted) after 40 years.Most agricultural trees have chilling requirements; that is, they need a period of low temperatures before flowering.This is modelled using the parameter T lim shown in Table 2.A 20-year running average of the coldest month maximum and minimum temperatures is compared with these values and defines the bioclimatic limits of each species.Hence, temperature warming above these limits would inhibit the establishment and survival of the perennial crops.
For deciduous trees, the active phase starts when the daily temperature is higher than the base temperature, and it is assumed that fruit growth occurs in the second half of the ac-Table 2. Key parameters of agricultural trees and potatoes.R: representative tree/plant for parametrisation; K_est: tree density range; HI: harvest ratio/index; T b : base temperature; GDD: growing degree day requirements to grow full leaf coverage (in deciduous trees) or to reach ripeness of fruit (in evergreen trees); T lim : lower and upper coldest monthly mean temperature; Ph opt : lower and upper temperature optimum for photosynthesis; HE: maximal height of tree; Phu_1/2: fraction of potential heat units accumulated in the first (1) or second (2) inflexion point of the optimal leaf area curve; Lmax_1/2: fraction of maximal leaf area index accumulated in the first (1) or second (2) inflexion point of the optimal leaf area curve; Phusen: fraction of growing season at which senescence becomes the dominant process; Lai_ha: fraction of leaf area index still present at harvest; PHU: potential heat units for ripeness; WCF: water content factor for conversion from dry to fresh matter.2008).e Ku et al. (1977).f Morales Sierra (2012).g FAO (2013a).For T lim of citrus, it was assumed that the rest can be induced by water deficit, not only by low temperatures.h Cannell (1985).i Based on FAO (2013a) information that indicates that citrus need 7 to 14 months from flowering to maturity.j Orwa et al. (2009).k Bastin and Henken (1997).l Aguilera et al. (2014).m California Rare Fruit Growers (1997).n FAO (2013b).o Roussos (2007).p Orlandi et al. (2014).q Kailis and Harris (2007).r Al-Khayri and Niblet (2012).s FAO (2002).t Elshibli (2009).u Large variations and lack of data lead to estimation of these parameters assuming relatively small, high-yielding varieties of palms, flowering around 1 month, developing fruits in around 4 months, with 1000 kg weight and yields of 500 kg per palm.v Zanotelli et al. (2013).w Toky et al. (1989).x Wünsche and Lakso (2000).y Parametrised as a standard apple tree and not as a dwarf.z Perry (2011).Apple, pear and cherry trees can survive lower winter temperatures, but other fruit trees (i.e.fig, peach and apricot trees) cannot.aa FAO (2011).ab Pontificia Universidad Católica de Chile (2008).ac Netafim (2013).ad Janick and Paull (2008).ae Duke (1983).af Alasalvar and Shahidi (2008).ag FAO (2013c).ah Meier (2001).ai Strik (2011).aj Ministry of agriculture, food and rural affairs (2013).ak Slovak Wine Academy Pezinok (2009).al Zamski and Schaffer (1996).am Sakin (2012).an Yan et al. (2007).ao Tsiros et al. (2009).ap Kiranga (2013).In wild conditions, cotton can be up to 5 m high, in crops, up to 1.5 m. aq Wright et al. (2004), Paytas and Tarrago (2011).Planting density varies largely.
tive phase of the year, i.e. when the phenological scalar (fraction of the maximum leaf coverage) is > 0.5 and before leaf senescence starts.Leaf senescence occurs when daily temperatures fall below the base temperature.
Following Sitch et al. (2003), evergreen trees are assumed to have constant leaf coverage and leaf longevities of more than 1 year.In this case, the accumulation of carbohydrates in the fruits occurs on days where temperature is above a tree-specific base temperature until a tree-specific threshold is reached (GDD).
Grass grows in the same areas of agricultural trees, except for cotton and grape plantations and orchards.For these three classes we assumed that grasses and weeds do not grow, thereby avoiding competition with the crops, implying that any ground cover is eradicated through some sort of weed control.This is the dominant practice in reality, although exceptions to this rule are gaining in importance.
Categories "vegetables" and "fodder grass" are modelled following the modelling approach of C3 grass described in Sitch et al. (2003).This is very appropriate for fodder grasses in the Mediterranean region since these are mainly alfalfa and clover.For vegetables, this parametrisation accounts for the very large physiological and allometric heterogeneity of vegetables, and also for multiple harvests per year, a fact that is well represented by a constant cover of the areas.Following the implementation of temperate herbaceous plant functional types (PFTs) in Sitch et al. (2003), the photosynthesis in vegetables and fodder grasses is assumed to be optimal between 10 and 30 • C. Vegetables and fodder grasses are harvested once their phenology is complete (i.e. the growing degree day accumulation determined by a parameter was reached) and the biomass increment is equal to or greater than 200 g C m −2 since the last harvest event.At that time, 50 % of the aboveground biomass is transferred to the harvest compartment.This assumption may be rather low for some vegetables such as lettuce, and rather high for others, such as beans.For the conversion from dry to fresh matter it was assumed that vegetables have an average water content of 40 %, which, again, is rather low for some cases, e.g.cucumbers, but rather high for e.g.garlic.The moisture content of fodder grass varies approximately between 10 and 75 %, depending on whether it is reported for hay, silage or fresh fodder.Here we represent fodder grass for hay production and assume thus a moisture content of 10 %.
The standard calibration process for agricultural management in LPJmL crops was extended in order to include agricultural trees.For annual crops this procedure consists in performing a set of runs with systematically modified management parameters representing the heterogeneity of fields, high-yielding varieties and the maximal achievable LAI (see more details in Fader et al., 2010).Similarly, 10 runs systematically modifying the tree-and country-specific plantation density parameter were performed to calibrate the management of agricultural tree plantations.Planting densities range from 25 to 230 % of the standard values, which were derived from literature research (see Table 2).For grapes the range was prescribed between 2000 and 15 000 vines per hectare.The tree density for each country was then chosen based on the best matching with reported FAO yields.

Irrigation water requirements and soil carbon
The computation of net and gross irrigation requirements (NIR and GIR, respectively; see below) in LPJmL is explained in detail in Rost et al. (2008) and Rohwer et al. (2006).The functioning of soil decomposition, soil biochemistry and soil hydrology, including soil organic carbon (SOC), is explained in Schaphoff et al. (2013) and Sitch et al. (2003).In the following paragraphs a short and simplified summary of these procedures will be given.
Irrigation is triggered in irrigated areas when soil water content is lower than 90 % of the field capacity in the upper 50 cm of the soil (here "irrigated layer").The plants' NIR is modelled in LPJmL as the amount of water that plants need, taking into account the water holding capacity of the irrigated layer and the relative soil moisture (Rost et al., 2008): where D (mm d −1 ) is the atmospheric demand, which depends on potential evapotranspiration and canopy conductance.Sy (mm d −1 ) is the soil water supply, which equals a crop's specific maximum transpirational rate if the soil is saturated or declines linearly with soil moisture.f Ril (fraction) is the proportion of roots in the irrigated layer.w il (fraction) is the water content in the irrigated layer.w r (fraction) is the water content weighted with the root density for the soil column.WHC (mm) is the field capacity of the irrigated layer (water holding capacity).GIR, also called water withdrawal or extraction, is obtained by dividing NIR by the project efficiencies (EP): EP is a country-specific parameter calculated for LPJmL by Rohwer et al. (2006) after the approach described in the FAO irrigation manual (Savva and Frenken, 2002).It takes into account reported data on conveyance efficiency (EC), field application efficiency (EA) and a management factor (MF): EA represents the water use efficiency in the fields and its increase from surface irrigation systems, via sprinkler systems, to drip irrigation systems.EC represents the water use efficiency in the distribution systems and is assumed to be linked to irrigation systems (lower for open channels than for pressurised pipelines).MF varies between 0.9 and 1 and is higher in pressurised and small-scale systems under the assumption that large-scale systems are more difficult to manage (see more details in Rohwer et al., 2006).
The soil column in LPJmL has five hydrologically and thermally active layers (20, 30, 50 cm, 1 and 1 m thickness) where roots have access to water.Infiltration depends on the soil water content of the first layer (water that does not infiltrate runs off), and percolation between the layers was simulated following the storage routine technique (see Schaphoff et al., 2013, for more details).Excess water over the saturation level is assumed to feed subsurface runoff.LPJmL has two soil carbon pools, with intermediate and fast turnover (0.001 and 0.03 rate of turnover per year at 10 • C).The maximum decomposition rate is reached around field capacity and decreases afterwards due to decreased soil oxygen content.A simple energy balance model is used for the thermal soil module.It includes a one-dimensional heat conduction equation, convection of latent heat, thawing and sensible heat (see Schaphoff et al., 2013, for more details).

Results
The performance of the improved LPJmL version was tested by simulating agricultural yields, irrigation water requirements and soil carbon density, and comparing the results to published observations.

Agricultural yields
Figure 1 shows LPJmL-simulated yields in metric tonnes fresh matter per hectare for the calibrated run for all new crops where FAOSTAT had data in the Mediterranean region, averaged for the period 2000-2009.LPJmL simulates all nuts, olives, fruits and potatoes in very good agreement with reported values, showing Willmott coefficients 1 of ≥ 0.6 in all cases.Only two cases with large planting areas and significant differences are visible, both in Turkey, for grapes and nut trees.The latter is due to the chosen representative tree for the parametrisation of this group (almonds), which does not represent the majority of nut plantations in Turkey.In 2010 almost 70 % of nut trees in Turkey were hazelnuts and only 3 % almonds (FAO, 2015a).The underestimation of grape yield in Turkey might be related to more than one factor, including the fact that the wine sector has been very dynamic in recent years there, with increases in production but decreases in area harvested (FAO, 2015a).FAO calculates yields by dividing national production by harvested area and calculates, thus, an increase in yields and a higher average over the years analysed.Our input data set shows a slight 1 The Willmott coefficient was developed by Willmott (1982) as a tool for testing model performance against independent data and is calculated by 1 where o is the independent data, p the LPJmL simulated data, and ō denotes the mean of independent data.increase in grape areas with relative constant production; hence, we calculated a lower yield average over the years.Also, the general parametrisation for European grapes probably cannot represent the special character of local Turkish varieties that are well adapted to sandy soils and high altitudes.
Validating subnational patterns of yields is very difficult due to a general lack of data on this and important differences with other estimations in terms of scale, methods and time frames.However, we included in Fig. S2 a comparison with the yields from Monfreda et al. (2008) for the new crops where this study offers subnational data (note that their estimates are for the period of time around the year 2000 and at the administration level).LPJmL reproduces correctly a number of spatial patterns, such as some high-yielding regions: olives in Greece, vineyards in Israel, Lebanon, southern Spain, the Po Valley and the Italian regions of Emilia-Romagna and Lazio, potatoes in Turkey, Greece, Egypt, Morocco, Israel, Lebanon and Algeria, as well as cotton yields in southern Spain, Greece, Turkey, Egypt, Israel and Lebanon.Also, some low-yielding zones are in good agreement, as is the case for potatoes in the Balkans, Portugal and Tunisia, olives in Morocco, Algeria and Tunisia, and cotton in Tunisia.However, a few patterns shown by Monfreda et al. (2008) are not shown in LPJmL simulations, including the north-south pattern of olives in Spain, the high-yielding zone of olives in southern Italy, and the grape and olive yields in Egypt.The first case is due to the extremely high management intensity in southern Spain which is not captured by the national calibration of planting densities (Scheidel and Krausmann, 2011).The latter case is originated by differences between the MON and MIRCA land use data sets that produce a lack of irrigated areas of grapes and olives in Egypt in our data set (see Sect. 2.2 for more details).The same is the case for the gaps in potatoes, e.g. in France.Overall there is a large agreement with no systematic differences between the spatial patterns shown by Monfreda et al. (2008) and the ones computed in the present study.

Irrigation water requirements
Figure 2 shows LPJmL-simulated NIR per hectare, which presents a clear north-south pattern that follows the climatedriven patterns of potential evapotranspiration.Konzmann et al. (2013) presented simulated irrigation requirements globally for around 10 crop functional types with a former version of the LPJmL model where tree plantations were represented as mowed grasslands.Their Fig. 1 shows a grid-cell pattern broadly similar to ours, but our more detailed representation of Mediterranean crops leads to higher values in various regions, including Algeria, Tunisia, Israel, Lebanon, Greece and the Iberian Peninsula.This is in good agreement with a general tendency of trees to absorb and transpire more water than grasslands (Belluscio, 2009).Siebert et al. (2010) computed present irrigation consumptive water use for subnational administrative units by means of the GCWM model that compares well with our NIR estimates (Fig. 3).However, they estimated higher values in Egypt, Libya, Greece and Portugal, and lower values in the Po Valley and southern France.These differences are mainly linked to disparities in the land use data set used as inputs: Siebert et al. ( 2010) areas equipped for irrigation are larger than the ones used in the present study in the first group of countries, and smaller in the second group.The comparison in absolute terms (Fig. S3) shows a similar pattern but with additional high differences in the Spanish province of Andalucía.These differences may be linked to the model approach, for example, the difference in the crops considered (olives, orchards, and nuts are only considered in the present study by LPJmL), different methods to model evapotranspiration (Penman-Monteith versus Priestley-Taylor) and differences in growing periods (e.g.dynamic versus static sowing dates for annual crops).Since Andalucía is a region strongly characterised by horticulture, and taking into account that we parametrised vegetables as C3 grasses, it is worthwhile looking in more detail into this class, also because neither Siebert et al. (2010) nor the present study account for cultivation of vegetables in greenhouses.We computed independent irrigation water requirements of 2357 m 3 ha −1 based on the values presented in Table 2 from Gallardo and Thompson (2013) that concerns various vegetables and water melon grown in greenhouses.Based on the crop cycles described in their publication, we assumed the possibility of planting three types of vegetables per year using the same area (multiple cropping).Vegetables are planted on around 1.6 Mha in the Mediterranean region.This yields a total water consumption of 11.3 km 3 .The present study computed 9.7 km 3 , that is, a very similar value.
In total, the agricultural sector in the Mediterranean was simulated to withdraw approx.223 km 3 of water per year for irrigation (average 2000-2009), with GIR being especially high in the Nile Delta, the eastern Mediterranean and in some Spanish regions (not shown).Our national GIR values are in good agreement with AQUASTAT data (FAO, 2015b) (Fig. 4, squares), with some differences.It is difficult to evaluate the quality of the AQUASTAT data.For example, the values of three countries with large differences from our estimates (Al- geria, Lebanon and Jordan) are in fact not reported data but modelled data.Assuming that the modelling was performed by the FAO's CROPWAT model, our estimates might be more accurate, since we perform a validated process-based calculation of transpiration instead of prescribing crop water coefficients.Another example of uncertainty is shown in the case of Egypt.In Fig. 4 (red symbols) it is evident that the estimates for Egypt vary largely; for example, Rayan and Djebedjian (2004) presented much lower estimates than AQUASTAT.Döll and Siebert (2002) were probably the first authors who quantified irrigation water requirements at the global level while distinguishing two crop classes (rice and nonrice).Their Table 5 shows GIR for Egypt, Israel and Spain according to independent data and their calculations (using irrigation areas from 1995 and climate from 1961 to 1990).Despite the difference in the period of time analysed and the methodology, Fig. 4 shows that our results agree well for Israel and Spain for the independent data (from the water commissioner of Israel and the executive secretary of the International Commission of Irrigation and Drainage, respectively), while they found their values to be overestimated for both countries.For Egypt, independent data deliver 30 % higher GIR than our calculations and 27 % lower than the values of Döll and Siebert (2002).However, the reliability of the reported water use data cannot be established, especially in the case of Egypt, where these numbers are relevant for negotia-tions on water allocation treaties with upstream countries of the Nile River.
A report by Cánovas Cuenca (2013), quoting an unpublished paper by Cánovas and del Campo (2006), shows in its Table 17 irrigation water requirements for Mediterranean countries assuming that every hectare agriculture needs 6176 m 3 of water per year.Our analysis shows that while this number delivers a fair estimate of irrigation requirements for some northern Mediterranean countries, it strongly underestimates irrigation requirements of dry Mediterranean countries (Fig. 4,dots).This confirms that the environmental and climate diversity of the Mediterranean region requires spatially explicit modelling approaches.
To summarise, LPJmL computes Mediterranean irrigation water requirements in the range of former studies, even if comparisons are a challenge due to inconsistent model inputs, differences in modelling approaches, and due to the fact that to our best knowledge, this is the first study with a complete representation of Mediterranean crops.

Soil carbon density
As mentioned in the introduction, some assessments can only be performed with a model that includes natural and agricultural ecosystems with a fair detail in the hydrological cycle.This is the case for carbon sequestration by soils in a unit of area that has both natural and agricultural vegetation -a very important variable for the climate debate and the ecosystem service research domain.
Soils under forest, grasslands, and cropland show different carbon densities depending on climate, vegetation, soil structure, and management.Generally speaking, forests have higher proportions of soil organic carbon (SOC) compared to mowed grasslands, and they, in turn, have higher values compared to cropland planted with annual crops (Jobbágy and Jackson, 2000;Eclesia et al., 2012;Werth et al., 2005).Also, evergreen broadleaved forests and plantations have usually higher SOC than deciduous forests and plantations in semiarid climates (Doblas-Miranda et al., 2013).Agricultural tree plantations have a lower tree density than forest, generally a regular distribution of trees that increase soil evaporation, and they are subject to removal of biomass (harvest).These factors lead to lower SOC values in tree plantations compared to forest.However, management of tree plantations, including irrigation input, planting density, presence or eradication of grass strips and mulching can strongly increase or decrease SOC.Putting all this information together and assuming that management and environmental factors are comparable, the SOC of agricultural tree plantations is expected to be generally higher than the SOC of mowed grasslands and generally lower than the SOC of natural forests and native grasslands.This is especially true for evergreen tree plantations and managed grasslands with high-frequency mowing.Hence, the implementation of agricultural trees in the LPJmL should produce higher SOC over the entire soil profile in many Mediterranean areas.This is because before the implementation of agricultural trees, the areas corresponding to these agro-ecosystems were simulated as mowed grasslands.
As expected, Fig. 5 shows that implementing agricultural trees in LPJmL increased the carbon stock in soils in the whole Mediterranean except France.This exception is due to the fact that there are only two new implemented crops in France with significant areas: non-citrus orchards and grapes.Both are deciduous trees, with high soil evaporation at the beginning and end of the active period affecting carbon decomposition.Also, orchards have a relatively low planting density in France (1300 trees per hectare), which reduces shadow effects and litter input that, in turn, results in lower soil carbon compared to high-density mowed grass.
Validation of the new SOC patterns is very challenging since SOC measurements are spatially discontinuous as well as dependent on local conditions, sampling method and small-scale drainage conditions.Comparison with empirically based or process-based modelled SOC is also difficult due to differences in approaches, parameters, processes considered and issues of scale.Nevertheless, we compared our SOC estimates before (LPJmL Old ) and after (LPJmL New ) the implementation of agricultural trees with the organic carbon density from the HWSD database (Hiederer and Köchy, 2012).These data are produced by establishing functions between SOC and soil type, topography, climate variables and land use situation.For this comparison we calculated the difference of absolute differences (/LPJmL Old − HWSD/ − /LPJmL New − HWSD/, Fig. 6).Considering significant differences (results <> 1 t ha −1 ), the number of grid cells with decreased differences to the HWSD estimates almost doubles the number of grid cells with increased differences (767 versus 460 grid cells).This means that the development presented in this study moved LPJmL's results for SOC closer to HWSD values.
As mentioned before, options for comparison with other SOC estimates are limited.The documentation of HWSD estimates (Hiederer and Köchy, 2012) offers an impressive effort in comparing their estimates with other assessments.They found large, spatially diverse differences from other es- timates; some of them could be associated with differences in approaches.When comparing LPJmL results with HWSD estimates, it is necessary to bear in mind that HWSD offers a more detailed spatial scale and representation of processes linked to soil types, while LPJmL has a more detailed influence of land use history, seasonality of temperature and types of crops in SOC formation.

Advances through consistent carbon-water-agriculture modelling
Environmental degradation, climate change, and population growth will put Mediterranean agriculture and natural ecosystems under pressure (IPCC, 2012;Kovats et al., 2014;Diffenbaugh and Giorgi, 2012;Skuras and Psaltopoulos, 2012;Doblas-Miranda et al., 2015).sible, but it also improved the potential for computation of irrigation requirements and soil carbon.The outcome is a model with a comprehensive representation of ecophysiological processes for all vegetation types (natural and agricultural) in a consistent and validated framework that produces estimates of carbon, agricultural and hydrological variables for the entire Mediterranean basin.As such, LPJmL is especially suitable for analyses of water issues.Taking into account the projected water scarcity due to climate change in the Mediterranean area, the continuously dropping groundwater tables due to overexploitation, and the projected increases in irrigation water demand (Fischer et al., 2007;Konzmann et al., 2013;Wada et al., 2010;Arnell, 2004), this constitutes a promising area for future model applications and further development.A first application of this model development was presented by Fader et al. (2015), pointing out that irrigation water needs of perennial crops in the Mediterranean region might increase significantly under climate change, and some countries may face constraints to meet the higher water demands.

Potential applications and perspective for further research
The inclusion of perennial crops in LPJmL presented in the current study opens up the possibility for a number of large-scale applications and research studies that cannot be performed with empirical and/or input-intensive agronomic models.These include assessments of climate change impacts on hydrological variables, agricultural production and carbon sequestration, as well as evaluation of consequences of land use change (including expansion of irrigated areas).Some of these applications could also be performed by land surface models, but LPJmL now has the advantage of considering perennial crops in detail, which allows more precise quantifications not only in the Mediterranean region, but also in other macro-regions where agriculture is partially dominated by tree crops, such as Australia, South Africa, Chile, western Argentina and California.Moreover, having a more accurate representation of perennial crops also allows studies on shifts in suitable growing areas for agricultural trees, diversity of diets, resilience of agricultural systems, needs for climate change adaptation and implications for food security, as well as assessments of ecosystem services provided by perennial cultures, for example habitat provision for avifauna.
Further improvements and refinements of LPJmL can be envisaged for applications in the Mediterranean area.We can divide these potential improvements into three groups, enumerated from least to most complex and work-intensive: (a) input-related, (b) parameter-related, and (c) inclusion of new processes.
The most important input-related improvement concerns national and subnational studies and is related to the need to increase the spatial resolution in all inputs used by LPJmL.
The limited availability of climate data and scenarios in a higher spatial resolution for the whole basin as well as missing detailed flow direction maps, especially for North Africa, have constrained this refinement until now.Nevertheless, work on data interpolation and downscaling is ongoing to bring the model to the 15 arcmin resolution.Another inputrelated issue is the need for scenarios of future crop patterns as a consequence of climatic and socio-economic change.With climate change very likely affecting the potential growing areas of agricultural trees and their profitability, studies aiming at a future quantification of agricultural, biochemical and hydrological variables would profit from coupling between models like LPJmL and land use models.
Small-scale application aiming at comparing and analysing single crops may require parameter-related changes such as re-parametrisation allowing differentiation of harvesting times after uses and varieties (e.g.varieties of grapes, difference between table olives and olives for oil), grid-cell-specific planting densities and its differentiation between irrigated and rain-fed conditions, as well as cropspecific setting up of fruits, which at present depends on the phenological development but is not differentiated for different crops.For national-scale studies, re-parametrisation with different representative plants for the groups (e.g. using hazelnuts instead of almonds for nut trees in Turkey) is possible without any difficulty.For this group of improvements, data on management are essential, including harvesting times, post-harvest uses, planting densities, planted varieties, etc.Another benefit of more precise management information would be the possibility of differentiating parameters that are assumed to be static and equal for all plants in the present study, such as the number of years that a perennial plantation stays in production before being renewed and the period of time from the planting until the first harvest (see Sect. 2.2).
Some refinements in modelled processes should be undertaken for studies aiming at detecting year-to-year phenological changes or sub-yearly patterns of carbon allocation in agricultural trees.These may include improved representation of chilling requirements, for example, implementing the chilling units approach (Byrne and Bacon, 2015), the variable harvest index depending on the special conditions of the year, implementation of dwarf trees, and daily update of carbon partitioning.Also, including a more differentiated approach to agricultural management, such as discretising practices, typology and processes affected, may be necessary for assessments of climate change adaptation and soil carbon.Also connected to soil carbon, the inclusion of erosion and salinisation would be essential since this process plays an important role in the semi-arid, hilly and terraced landscapes of the Mediterranean area (García-Orenes et al., 2012;Poessen and Hooke, 1997).
Finally, the inclusion of horticulture in LPJmL opens the door to further large developments aiming at the assessment of alternative agro-ecosystem managements and their envi-ronmental performance.One clear example of this would be the link between agroforestry systems and biodiversity conservation.

Code availability and technicalities
LPJmL is written in the C programming language and is run mainly under UNIX-like systems.Inputs and outputs are in binary format.Depending on the size of the region analysed and on the desired spatial resolution, it may require a highperforming computational structure.The version used as the basis for the present development was 3.5.003and the revision number 2133 from 17 May 2014.
The main site for downloads of different model versions can be found under https://www.pik-potsdam.de/research/projects/activities/biosphere-water-modelling/ lpjml/versions.
There, downloads are free of charge and possible after registration.However, the latest model version that includes the agricultural module and the Mediterranean development is not available yet since the different working groups are still compiling a complete technical documentation and merging the last developments into one unique model version.Please contact the first author of this publication if you plan an application of the model and envisage longer-term scientific collaboration.
The Supplement related to this article is available online at doi:10.5194/gmd-8-3545-2015-supplement.

×
Bias correction of HYDE global values was performed by HYDE g,y,bias_corrected = HYDE g,y − D 2000 , (2) where y represents years from 1700 to 2000.Cell (c) bias correction was done by HYDEcorrected, c, y = HYDEc, y HYDEg, y × HYDE g, y, bias_corrected.(3) Proportional temporal change of MMR cell values was done from 2000 backwards by MMRy, c, proportional = HYDEcorrected, c, y HYDEcorrected, c, y + 1 fractions from 2001 to 2010 follow the trend between 1950 and 2000.

Figure 1 .
Figure 1.Scatter plot of LPJmL-simulated yields (averages over 2000-2009) versus reported yields from FAOSTAT for Mediterranean countries.The line represents the 1 : 1 line.FM stands for fresh matter.The radius of the bubbles indicates the relative size of the harvested area in the respective country.

Figure 3 .
Figure 3.Comparison of irrigation consumptive water use from Siebert et al. (2010) and net irrigation water requirements computed in this study as a percentage of the Siebert et al. (2010) values.Negative (positive) values indicate higher (lower) values in their study.

Figure 4 .
Figure 4. Scatter plot of LPJmL-simulated gross irrigation water requirements (hm 3 , averages over 2000-2009) and the estimates presented by other studies.The line represents the 1 : 1 line.Egypt and Spain have been coloured in red and green, respectively, to visualise the disparities between different studies.(Note that Cánovas Cuenca, 2013, assumes a fixed requirement of 6176 m 3 ha −1 .)

Figure 5 .
Figure 5. Difference in LPJmL-simulated soil organic carbon (average 2000-2009) before and after the implementation of agricultural trees described in this study.Negative (positive) values indicate that the development decreased (increased) soil organic content.

Figure 6 .
Figure 6.Difference between the HWSD soil carbon density and LPJmL-simulated soil carbon density before and after the model development presented here (/LPJmL Old − HWSD/−/LPJmL New − HWSD/).Positive values indicate an improvement in simulation of soil carbon stock due to the implementation of Mediterranean crops described in the present study.

Table 1 .
Areas of LPJmL crops in the Mediterranean region.The stars indicate crops that are implemented in this study.
LPJmL is now suitable for supporting Mediterranean decision makers.The inclusion of Mediterranean crops in LPJmL not only increased substantially the proportion of agricultural areas for which quantitative assessments are pos- www.geosci-model-dev.net/8/3545/2015/Geosci.Model Dev., 8, 3545-3561, 2015