RothCUK – a dynamic modelling system for estimating changes in soil C from mineral soils at 1-km resolution in the UK

Authors


P. Falloon. E-mail: pete.falloon@metoffice.gov.uk

Abstract

We describe the development and application of an integrated data and modelling system for estimating soil carbon (C) fluxes from mineral soils caused by changes in climate, land use and land management at 1-km resolution in the UK (RothCUK). The system was developed with the aim of improving methods for United Nations Framework Convention on Climate Change (UNFCCC) and Kyoto Protocol accounting and integrates national scale data sets of soil properties, land use and climate with the Rothamsted carbon model (RothC). A preliminary estimate of soil C fluxes because of land use change (LUC) over the period 1990–2000 is presented as an example application of the system. RothCUK shows LUC to be a net source of CO2 from 1990 to 2000 although the RothC estimate was smaller (6488 kt C) than the estimate from the single exponential model (SEM) method currently used to calculate C fluxes due to LUC for the UK National Greenhouse Gas Inventory (mean: 9412 kt C). Based on previous studies, an uncertainty range in our estimates of ±50–100% seems plausible. In agreement with the SEM, RothCUK suggests that the largest single contributor to soil C fluxes from LUC was conversion of grassland to arable land. Differences between the results may be attributed to differences in the two models and the assumptions and underlying data used in making the calculations. The RothCUK system provides a powerful method for estimating changes in soil C stocks, enabling areas and management systems with particularly large changes in soil C stocks to be located at fine resolution.

Introduction

Maintaining and regularly updating the UK Greenhouse Gas Emissions Inventory (GHGEI) category on land use change (LUC) and forestry is a commitment under the United Nations Framework Convention on Climate Change (UNFCCC) and the EU Monitoring Mechanism for greenhouse gas emissions. In addition, the UK Department for the Environment, Food and Rural Affairs (Defra) is required to report emissions to the EU Monitoring Mechanism, whilst Article 3 of the Kyoto Protocol lays the foundations for inclusion of human-induced changes in fluxes of C from agricultural soils and LUC and forestry in Annex I Parties’ commitments. The current method for estimating changes in soil organic carbon (SOC) stocks because of LUC for the UK National GHGEI uses a modification of the Intergovernmental Panel on Climate Change (IPCC) method (IPCC, 1996; Paustian et al., 1997) which is based on a first-order dynamic model for determining changes in SOC. SOC stocks (when decreasing) are assumed to change from one equilibrium level to another after a period of between 50 and 150 years (Cannell et al., 1999; Baggott et al., 2004). This follows the pattern of a single exponential and this current method is hereafter referred to as the single exponential model (SEM). Increases in SOC similarly follow a single exponential pattern but take longer (100–750 years) to reach equilibrium. However, there is considerable uncertainty associated with C flux estimates, particularly on a regional basis, and changes may occur over larger or shorter periods. A spatially explicit dynamic soil carbon simulation model based on RothC (Coleman & Jenkinson, 1996) has been previously developed (Falloon et al., 2002; Falloon & Smith, 2002). This paper describes the further extension of this system with soil, land use and climate information at 1-km resolution in the UK (RothCUK). This will provide more process-based and spatially explicit estimates of changes in soil C stocks which explicitly account for difference in climate, land-use and soil type. This approach allows climate change and LUC to be examined either separately or together and for interactions to be predicted. The present work describes the first national scale comparison of RothC and the SEM, the first national scale application of RothC in the UK and at 1-km resolution, and the first modelling application of newly developed soil and LUC databases (Bradley et al., 2003) in the UK. As an example application, the modelled impacts of LUC on C stocks of mineral soils in Great Britain (GB) between 1990 and 2000 are presented.

Methods

The RothC model

The RothC model (Coleman & Jenkinson, 1996) is a monthly time-step dynamic SOC decomposition model that splits incoming plant residues into decomposable plant material (DPM) and resistant plant material (RPM); both of these decompose to form microbial biomass (BIO), humified organic matter (HUM) and evolved CO2. The model also includes a pool of inert organic matter (IOM). With the exception of IOM, each compartment decomposes by first-order kinetics and each has an intrinsic decomposition rate. The actual rate of decomposition is determined using modifiers for soil moisture, temperature and plant cover, operating on the intrinsic rate. The clay content of the soil affects the apportioning of SOC between the evolved CO2, the BIO and the HUM pools. The main input data required to run the model are soil clay content, initial SOC, soil bulk density, total monthly precipitation, mean monthly temperature, total monthly evaporation, residue quality, soil cover, residue C inputs and manure C inputs. Plant inputs of C to soil for RothC can be set in one of two ways. First, RothC may be run using ‘inverse modelling’ of SOC (Coleman et al., 1997): by adjusting the annual plant input of C to soil to obtain the best fit between modelled and measured SOC values (‘fitted’). Model best fit can then be assessed using the root mean square error (RMSE), a measure of the mean error between simulated and measured data points irrespective of the direction of the error (Smith et al., 1996). Secondly, RothC can be run using plant inputs set independently of SOC data (‘defaults’). Initial pool sizes for RothC may be estimated as described below. If both total SOC measurements and soil 14C dates are available, an inverse modelling procedure can be used to derive the IOM content of the soil for the RothC model. In the absence of soil 14C dates, the equation derived by Falloon et al. (1998) may be used to estimate IOM from total SOC content. RothC has been extensively evaluated using SOC data from long-term experiments covering a wide range of land uses, soil types and climates (see, for example, Coleman et al., 1997; Falloon & Smith, 2002), generally with the RMSE between modelled and measured values below 6.8–8.5% although the lack of estimates of error about the mean of SOC measurements makes it difficult to separate model error from measurement error (Falloon & Smith, 2003).

Input data sets and model linkage

RothCUK is an interface built around the RothC compartmental model of soil carbon turnover in soils. It uses 1-km resolution soils, land use and LUC data, along with relevant current and future climatic data sets, and can be used to investigate the effects of changes in land use, land management and climate change on national C stocks. The model may be run for GB or Northern Ireland (NI), and runs separately for different land uses [arable, grass, (semi)-natural and woodland] and two soil depths (0–30 and 30–100 cm) based on the available data sets, as well as for LUC matrix data for GB. Output data is provided for soil total C and CO2 fluxes in formats suitable for import into geographic information systems (GIS) packages. The main sources of input data to the model are described in Tables 1 and 2. Although the 1-km soil databases detail soil parameters for each of the component soil series in each 1-km square, currently the RothCUK system is only set up to run with the dominant soil series of each 1-km square. Uncertainty in using dominant soil series data only has been assessed by Falloon et al. (2004) and is dealt with in the discussion of this paper.

Table 1.   Model input data sources – Great Britain
Data setSource
1-km soil series extentsNational soil map of England & Wales
National soil map of Scotland
Soil parameter tables for soil seriesLandIS & NSI (England & Wales)
Scottish soil databases
1-km land use classes1990 Corine land cover map
1-km land use change matrices1990 & 1998 land use maps
Climate dataBaseline (1961–1990 average) climatology and future climate scenarios:
Baseline + SRES scenarios (CRU)
Baseline + UKCIP scenarios (UKCIP)
Table 2.   Model input data sources – Northern Ireland
Data setSource/supplier
1-km soil series extentsSoil maps of Northern Ireland (1:50 000)
Soil parameter tables for soil seriesDARD Soil Survey Database
1-km land use classes1990 Corine land cover map
Climate dataBaseline (1961–1990 average) climatology and future climate scenarios:
Baseline + SRES scenarios (CRU)
Baseline + UKCIP scenarios (UKCIP)

Land use.  The 1990 CORINE land cover database for UK (Brown et al., 2002) was derived from high-resolution satellite images at a scale of 1:100 000. The legend identifies 44 classes grouped in a hierarchical nomenclature and is landscape and ecology oriented. Most CORINE classes are by nature heterogeneous and/or are determined by their functional land use and thus consist of various land cover types with minimum mappable units of 25 ha. For the purposes of the current study, classes grouped into six broad land use types within each 1-km square were used as follows: (1) Arable – agriculture including managed short-term grass; (2) permanent grassland (Grass) – improved managed grass; (3) semi-natural (natural) vegetation – all vegetative cover outside woodland and agriculture; (4) woodland – including coniferous, deciduous and planted forest and other trees; (5) urban – no vegetation; all built up land including in rural locations; quarries; (6) Other – no vegetation; quarries, etc.

Land use change.  In the UK GHGEI, a LUC matrix has been developed for the whole of each of the four countries. These have been estimated from national data from the Monitoring Landscape Change project (1947, 1980) (MLC, 1986) and Countryside Surveys (1984, 1990, 1998) (Barr et al., 1993; Haines-Young et al., 2000; Cooper & McCann, 2002). RothCUK has been applied to the 1990–2000 period by developing LUC matrices for each 1 × 1 km grid cell in GB from 25-m resolution land cover maps (1990 and 1998) based on Landsat data (Barr et al., 1993; Haines-Young et al., 2000). There are difficulties in estimating LUC from these two land cover maps because of differences in the methodologies used for each. For the 1990 map, each 25-m pixel of spectral data from the satellite sensor was classified into one of 25 ‘Target Classes’. For the 1998 map the pixels were classified in groups or ‘parcels’ with similar spectral characteristics and ‘Broad Habitats’ were used for the land cover classes (Figure 1). Developing land use or cover change matrices cannot be successfully completed by comparing the same pixels in the two years because the per pixel classification in 1990 tends to cause noise on data and hence spurious land change and there are additional difficulties because of the different basis for the land cover types in the two years. To minimize these problems the land classifications were: (a) broadened to the six broad types used for soil carbon classification, i.e. arable, grassland, semi-natural, urban, woodland, other – land with zero carbon, and (b) the pixels were grouped together into blocks of 2 × 2, 4 × 4, 5 × 5, 8 × 8, 10 × 10, 20 × 20. A LUC matrix was then estimated for each 1 × 1 km grid cell from these ‘degraded’ pixels. A ‘best’ block size for each of the 32 landscape classes (ITE Land Classes) was decided by comparison of the results from this approach with field data from 501 Countryside Survey sample squares. The ‘best’ block size thus represented the blocking, or aggregation of 25-m pixels from the land cover maps which best represented the coverage of each land use by comparison with Countryside Survey data (based on ground surveys). This, therefore, allows assessment of the accuracy of using the two digital land use maps to estimate LUC. ‘Best’ block size varied with ITE Land Class and is shown in Table 3. The differences in the transitions in the land use matrix from remotely sensed data for different block sizes and the field data for two example ITE Land Classes are shown in Figure 2. The matrices developed for the period 1990–1998 were assumed to apply also to the period 1990–2000. Table 4 shows a typical matrix, for an example square. Currently no comparable information is available for NI.

Figure 1.

 Illustration of the effect of different classification methods for data for 1990 and 1998 from a single 1 × 1 km grid cell. In 1990 classification was per individual pixel but in 1998 pixels were classified in groups or ‘parcels’.

Table 3.   ‘Best’ size for each ITE Land Class for block degradation of 25-m pixels from land cover maps
Pixel block size (25-m pixels)ITE land classes
1 × 14678101721252730 
2 × 21529         
4 × 420          
5 × 512359121314182632
8 × 81931         
10 × 10111622232428     
Figure 2.

 Total mean square difference between transition in a land use change matrix developed using different block sizes for pixel groups and data from field survey data for sample locations of the same ITE Land Class. Different colours in a column relate to different entries in land use change matrix, e.g. woodland to urban, natural to woodland (36 in all). (a) LC 1 England Central South Downs – gently rolling farmland, both grassland and arable, (b) LC 22 Southeast uplands of Scotland – open moorland on rounded hills.

Table 4.   Land use change matrix 1990–2000 for an example square – values are % of the 1-km square under each land use change
From 1990To 2000
GrasslandUrbanWoodlandZeroArableNatural
Grassland0.000.370.000.000.08
Urban0.410.100.150.000.00
Woodland0.000.000.000.000.00
Zero0.000.000.000.000.00
Arable0.170.000.020.000.34
Natural0.620.000.720.000.00

Soil spatial data.  The soil series spatial data for England and Wales were based on the revised national soil map (2001) derived from the 1:250 000 scale map published in 1983 (Soil Survey Staff, 1983). This shows the distribution of 296 soil associations, each comprising between one and eight soil series which are found together within a particular landscape type. Soil series are characterized by a precisely defined set of soil and substrate properties. They form the basic unit of soil classification and mapping at field scale (Clayden & Hollis, 1984). An estimate of the average proportion of each main soil series within each soil association was made nationally, based on field data collected and recorded during the making of the 1:250 000 scale soil maps. A total of 410 soil series are recognized on the map. Using these data, a soil series data set at 1-km resolution was created by integrating the proportion of each soil association within each 1-km square with the proportion of soil series within each soil association. The spatial soil series data for Scotland were based on the 1:250 000 scale national soil map published in 1981 (Soil Survey of Scotland Staff, 1981). This map shows the distribution of 580 soil map units, plus miscellaneous units comprising rock, scree and urban. Each map unit comprises a number of soil series that are found associated within a particular landscape. These are described in the Soil Memoirs of Scotland. The component soil series were assigned to each of the 1:250 000 scale map units and the proportion within each map unit estimated. Using these data a soil series data set at 1-km resolution was created by multiplying the fraction of each map unit within each 1-km square with the fraction of each soil series within each map unit. A total of 540 soil series are recognized on the map. The spatial soil data for NI were based on the 1:50 000 scale NI soil map series (Cruickshank, 1997). A total of 476 mapped series are recognized, and using these, a 1-km grid data set of soil series in each square was created.

Soil parameter data – general principles.  The following principles were applied when producing soil parameter data.

  • 1All soil series were parameterized for organic carbon, clay, silt, sand and bulk density by standard layers (0–30 and 30–100 cm). The driving variables required by RothC were soil clay%, organic carbon%, reference depth (cm) and bulk density (t m−3);
  • 2H horizons were included but C and R horizons excluded;
  • 3the stoniness of each horizon was identified so that the mass of organic carbon could be derived on a <2-mm fraction basis;
  • 4the depth of the profile to rock, extremely stony or C horizons was identified so that the depth over which the mass of carbon was calculated could be known;
  • 5the various parameters were identified according to four land use types: (a) arable – cultivated land (mainly arable and ley (short-term) grassland); (b) grass – managed permanent grassland; (c) natural – semi-natural vegetation and grassland that receives no management; (d) woodland (deciduous and coniferous trees);
  • 6where no ‘actual’ soil property data were available for a particular soil/land use combination, surrogate data were used as a first option, based on the data from a similar soil series, using expert judgement. If no such series with appropriate data were found then ratios of the parameters derived from those series with data were used (Cruickshank et al., 1998). Further details of the derivations of the data are in Bradley et al. (2003).

For modelling purposes all soils where any of the parameter values were missing were excluded. This mostly affected organic soils, which had no corresponding texture values although soils with texture values and high SOC were included. This meant that a considerable portion of the GB and NI SOC stock was not included in our simulations. However, as the RothC model is not currently applicable to organic soils, our estimates of C fluxes for such soils would have been subject to large errors.

Climate.  Two sources of climatic data were selected for use with RothCUK, and the climate scenarios were chosen to cover the range of projected climate change in the UK:

  • 1UKCIP02 (Hulme et al., 2002) 50-km resolution, 1961–1990 average baseline climate plus two scenarios, low and medium high. These scenarios were based on regional model (HadRM3) ensemble runs with emissions scenarios from the IPCC Special Report on Emissions Scenarios (SRES: IPCC, 2000) – B1 and A2, respectively. The data were available for three time slices – 2020s (2011–2040), 2050s (2041–2070) and 2080s (2071–2100). Predicted mean temperature increases by the 2080s across the UK ranged from 1.5 to 2.5 °C under the low scenario, and 2–4 °C under the medium-high scenario; the corresponding changes in annual mean total precipitation were 0 to −20% and 0 to −10%, respectively. Variables used were mean monthly temperature and rainfall. Potential evapotranspiration (PET) was not available with the raw data set and hence was calculated from the available variables following the simple equation of Franko et al. (1995), based on grassland land use.
  • 2CRU HadCM3 output data plus 1961–1990 average baseline climate, using two SRES climate change scenarios – A1F1 (high) and B1A (low) scenarios. The variables used were mean monthly temperature and rainfall. A wider range of variables were available which allowed PET to be derived from the equation of Penman (1948) based on grassland land use.

The baseline climatology for GB was based on the GB data set of Barrow et al. (1993) at 10-km resolution whilst that for NI was based on the global data set of New et al. (2002) at 0.5° resolution. These baseline climatologies were used in the present study and the climate scenario data for investigating climate impacts on soil C stocks (Falloon et al., 2004).

An example application – modelling carbon fluxes because of land use change in GB

To model the effect of LUC on soil carbon stocks in GB, the RothCUK system uses LUC matrix data at 1-km resolution as described above. The RothCUK model was applied to these data, along with the 1-km soil database and average 1961–1990 climatology. As discussed before, the RothCUK system currently only uses soil data from the dominant soil series of each 1-km square.

The 1-km LUC matrix database provided no information on precisely when between 1990 and 2000 a particular LUC occurred, or whether it occurred as a step change over the whole 25-m sub-square, or in increments through time over the sub-square. For example, a particular grid square could be arable land in 1990, and grassland in 2000, but the change could equally have occurred in 1990 following the mapping exercise, shortly before mapping in 2000, or any date in between. It is important to determine when LUC occurred between these time points, as this has considerable influence on the predicted C flux following LUC but this is made difficult where no information is available between the two sample points. Previous work investigating LUC in hypothetical 1-ha squares showed that for all LUC scenarios, a gradual change in increments of 0.1 ha year−1 resulted in a SOC concentration at the end of 10 years that were close to values assuming a step change in the whole 1 ha at year 6 (Ashman et al., 2000). For this study, LUC was, therefore, implemented at year 6 (1995), assuming the whole fractional area of the 1-km square changed to the new land use instantaneously, although different results would be expected if different assumptions were used.

Assumptions made about the nature (or intensity) of LUC could also have significant effects on predicted C fluxes. For certain LUCs, this is not an important issue as LUC can be assumed to occur instantaneously – for example, conversion of arable land to newly seeded grassland, or ploughing up of old grassland and conversion to arable land. Lack of further information from available data makes it difficult to specify how LUCs that could occur more gradually might have occurred and how to model them. Examples of gradual LUCs could be afforestation of arable land – if this was natural regeneration, it would involve a period with grassy species, followed by a period dominated by scrub, and finally establishment of trees. By contrast, if the change involved plantation of fast-growing conifer species, it might be considered almost instantaneous. A set of preliminary LUC transitions were developed for use with the RothCUK system based on this assumption and on pre-existing knowledge of modelling soil carbon fluxes due to LUC (Table 5).

Table 5.   Assumed land use change transitions for initial RothCUK runs 1990–2000
 Equilibrium land useLand use years 1–5Land use years 6–10Rationale
Arable to arableArableArableArableNo change
Arable to naturalArableArableGrasslandTransition to natural preceded by grassy species
Arable to grasslandArableArableGrasslandInstallation of grassland instantaneous
Arable to urbanArableArableUrbanInstallation of urban instantaneous
Arable to woodlandArableArableGrasslandTransition to wood preceded by grassy species then scrub
Arable to zeroArableArableZeroTransition to zero instantaneous
Natural to arableNaturalNaturalArableInstallation of arable land instantaneous
Natural to naturalNaturalNaturalNaturalNo change
Natural to grasslandNaturalNaturalGrasslandInstallation of grassland instantaneous
Natural to urbanNaturalNaturalUrbanInstallation of urban instantaneous
Natural to woodlandNaturalNaturalNaturalTransition to natural preceded by scrub
Natural to zeroNaturalNaturalZeroTransition to zero instantaneous
Grassland to arableGrasslandGrasslandArableInstallation of arable land instantaneous
Grassland to naturalGrasslandGrasslandGrasslandTransition to natural preceded by grassy species
Grassland to grasslandGrasslandGrasslandGrasslandNo change
Grassland to urbanGrasslandGrasslandUrbanInstallation of urban instantaneous
Grassland to woodlandGrasslandGrasslandGrasslandTransition to wood preceded by grassy species
Grassland to zeroGrasslandGrasslandZeroTransition to zero instantaneous
Urban to arableUrbanUrbanArableInstallation of arable land instantaneous
Urban to naturalUrbanUrbanGrasslandTransition to natural preceded by grassy species
Urban to grasslandUrbanUrbanGrasslandInstallation of grassland instantaneous
Urban to urbanUrbanUrbanUrbanNo change
Urban to woodlandUrbanUrbanGrasslandTransition to wood preceded by grassy species
Urban to zeroUrbanUrbanZeroTransition to zero instantaneous
Wood to arableWoodWoodArableInstallation of arable land instantaneous
Wood to naturalWoodWoodNaturalInstantaneous change/management
Wood to grasslandWoodWoodGrasslandInstallation of grassland instantaneous
Wood to urbanWoodWoodUrbanInstallation of urban instantaneous
Wood to woodlandWoodWoodWoodNo change
Wood to zeroWoodWoodZeroTransition to zero instantaneous
Zero to arableZeroZeroArableInstallation of arable land instantaneous
Zero to naturalZeroZeroGrasslandTransition to natural preceded by grassy species
Zero to grasslandZeroZeroGrasslandInstallation of grassland instantaneous
Zero to urbanZeroZeroUrbanInstallation of urban instantaneous
Zero to woodlandZeroZeroGrasslandTransition to wood preceded by grassy species
Zero to zeroZeroZeroZeroNo change

The RothCUK system was set up to allow modelling of LUC using C inputs data in one of two ways. First, the user could supply information for C inputs to soil under each component land use (default C inputs), or secondly, the system could be used to estimate C inputs at equilibrium for each component land use based on the soil parameter data supplied at 1-km level (fitted C inputs). The results presented here used the latter approach. As soil parameter data were unavailable for urban and zero land uses, a fixed value for SOC under urban land use was assumed (5.58 and 13.02 t C ha−1 for the 0–30 and 30–100 cm layers, respectively –Bradley et al., 2003), because urban areas do contain gardens, parks, verges and other carbon-containing soils. As land use conversion to and from ‘zero’ land uses (i.e. lake, concrete, etc.) would require initial or final carbon values, 0 t C ha−1 was assumed. For both urban and zero land uses, the soil texture under arable land use was used. Based on these assumptions, RothCUK was run to simulate soil C fluxes because of LUC from 1990 to 2000 as follows. For each 1-km square, each LUC transition was run separately for each soil depth (0–30 and 30–100 cm). The model was first run to equilibrium under the ‘to’ land use (e.g. grassland in an arable–grassland transition), so as to estimate the C input required. The model was then run to equilibrium under the ‘from’ land use (e.g. arable in an arable–grassland transition), and then in dynamic mode for 10 years following the transitions in Table 5. The DPM/RPM ratios used with each land use were as follows: arable 1.44, grassland 1.44, natural 0.67, wood 0.25, urban 1.44, zero 1.44 (irrelevant because of zero C content).

Results

The total SOC stocks to 1-m depth modelled at equilibrium (1990) by fitting C inputs were 2825 Tg C for GB and 177 Tg C for NI; the component depth data for 0–30 cm were 1816 Tg C for GB and 125 Tg C for NI, for 30–100 cm were 1010 Tg C for GB and 52 Tg C for NI. SOC stock estimates to 1 m depth of 9838 Tg C for GB and 386 Tg C for NI are given by Milne & Brown (1997) and Cruickshank et al. (1998), respectively – also based upon dominant soil series (and dominant land use in the case of NI) and not considering peat over 1 m depth in Scotland, or revised bulk density estimates for peats. The total SOC stocks from the databases (which did include peat over 1 m depth in Scotland, and revised bulk density estimates for peats) were 4267 Tg C for GB and 296 Tg C for NI (to 1 m depth); corresponding figures for 0–30 cm were 2370 Tg C for GB and 124 Tg C for NI – for 30–100 cm the values were 1895 Tg C for GB and 172 Tg C for NI (Bradley et al., 2003). However, the RothC modelling exercise did not include soils where any of the parameter values were missing (largely texture values for organic soils), so that the present approach would provide significantly lower estimates. As approximately 43 and 56% of the GB and NI SOC stocks are attributed to peat soils, respectively, the corresponding SOC stocks in mineral soils to 1 m depth are 1828 and 165 Tg C for GB and NI, respectively (Bradley et al., 2003) – much closer to our modelled values.

Table 6 shows modelled average C inputs at equilibrium for each soil depth and major land use category in GB and NI. A considerable range of C inputs for all land uses was observed, and C inputs differed considerably between land uses, generally increasing from arable > grass > natural > woodland in agreement with previous work (Falloon et al., 2002; Falloon & Smith, 2002). The great range of C input values obtained reflects the variability between sites, land management and climate. Greater C input values were predicted for NI than for GB, and generally slightly greater C input values were required than would be expected from former RothC modelling experience modelling long-term experimental data (e.g. Coleman et al., 1997). C inputs to subsoils were generally less than those for topsoil layers, and some very large and unreasonable C input values were obtained for soils with high C contents, indicating that the current model has limited applicability under these conditions (Falloon et al., 1998).

Table 6.   Mean C inputs (t C ha−1) fitted to measured soil organic carbon (SOC) values for all 1 km squares of Great Britain (GB) and Northern Ireland (NI)
Land useSoil depth (cm)GBNI
  1. Figures in parentheses are minimum, maximum and standard deviation.

Arable0–303.67 (1.41, 22.31, 1.71)5.86 (2.54, 19.92, 2.02)
30–1001.89 (0.05, 29.05, 1.33)2.31 (0.17, 20.23, 1.63)
Grass0–304.18 (1.22, 20.54, 2.16)5.61 (2.15, 17.52, 1.87)
30–1002.01 (0.04, 21.75, 1.31)2.40 (0.14, 19.34, 1.33)
Natural0–306.72 (0.66, 28.24, 3.65)8.09 (2.56, 17.01, 3.05)
30–1002.93 (0.03, 17.02, 2.00)3.53 (0.18, 17.83, 2.57)
Woodland0–304.86 (0.76, 26.35, 3.19)9.25 (5.27, 12.68, 2.63)
30–1002.36 (0.05, 15.76, 1.90)3.98 (1.87, 6.69, 1.48)
Urban0–300.19 (0.14, 0.42, 0.04)
30–1000.46 (0.31, 1.02, 0.10)

The national scale results from the initial RothCUK model runs for LUC from 1990 to 2000 are shown in Figures 3–6; Figures 3 and 4 additionally show comparisons between RothCUK and the SEM (data from Ashman et al., 2000; Milne et al., 2003, respectively). Figure 3 shows typical time series of SOC change for a typical 1-km square as predicted by RothC and the SEM method for LUCs showing extreme SOC changes, over a 300-year period, to demonstrate the SOC dynamics, and the timescale required to reach a new equilibrium SOC content (the ‘kink’ in Figure 3a at around 25 years represents a change in land cover from scrub to woodland as part of the LUC succession). This indicates that whilst both methods reach the same final equilibrium soil C value, the SEM method significantly overestimates C fluxes because of LUC when C stocks are both increasing and decreasing, as shown by Ashman et al. (2000). The national scale results from the initial RothCUK model runs for LUC from 1990 to 2000 are shown in Figures 5 and 6; the RothCUK model approach considered up to 36 LUCs for each of the 234 951 one-km squares in the soil and land use database. Comparing the total predicted C fluxes using the high-resolution spatial RothCUK approach with results from the SEM step-change method using highly aggregated data (Figure 4) showed a relatively small difference in predicted C fluxes because of LUC between the approaches, given their very different nature. The total C fluxes over the period 1990–2000 from RothC (1-km data, with LUC implemented in 1995) and SEM (highly aggregated data, LUC after 1990 – mean of Monte Carlo outputs from Milne et al. (2003) were 6488 and 9412 kt C, respectively. The SEM method and its application have been described briefly in the introduction of this paper, and fully by Milne et al. (2003).

Figure 3.

 SOC changes (0–30 cm) predicted by inline image RothC model and inline image SEM method over 300 years for a typical 1-km square for land use changes (a) of arable to woodland and (b) woodland to arable.

Figure 4.

 Comparison of 0 to 100 cm SOC fluxes because of land use change for GB 1990–2000 predicted by inline image RothC (using 1-km data, land use change implemented in 1995), and SEM using highly aggregated data (land use change after 1990), inline image mean of Monte Carlo results, inline image minimum of Monte Carlo results, inline image maximum of Monte Carlo results.

Figure 5.

 Soil C fluxes caused by land use change predicted by RothC model for bsl00000 0–30 and bsl00001 30–100 cm layers. A–G represents land use change of arable (A) to grass (G) – other land use codes as follows – (semi)natural (N), woodland (W), urban (U), zero (Z).

Figure 6.

 One-km land use change model outputs (a) % area of 1-km squares converted from grassland to arable land use 1990–2000, (b) 0–30 cm C fluxes from the conversion of grassland to arable land 1990–2000 (t C ha−1), (c) 0–30 cm C fluxes from the conversion of grassland to arable land 1990–2000 (tonnes per 1 km square), (d) Overall 0–100 cm soil C fluxes from LUC in 1-km squares 1990–2000 (tonnes per 1 km square).

Land use change both increased and decreased SOC stocks over the period investigated depending on local conditions (Figure 5), although overall GB was found to be a source of CO2 because of LUC, in agreement with the SEM method results. The largest contributor to C fluxes because of LUC was found to be conversion of grassland to arable land (a source of CO2) – the areas affected by this LUC, and resulting SOC changes are shown in Figure 6. This shows that although differences in the SOC change per unit area are important in determining the total SOC flux from a grid square (Figure 6b) the dominant factor is the amount of land changing to a different land use in each grid square (Figure 6a,c). Figure 6d shows the total C flux because of LUC from each 1-km square over the 1990–2000 period, which also indicates that the areas of greatest SOC loss are correlated with areas where considerable conversion of arable land to grassland was observed. The largest sink of C because of LUC from 1990 to 2000 was shown to be conversion of urban land to arable land (Figure 5), which is likely to be an artefact of land use map processing. In general, the 0–30 cm layer contributed the majority of soil C fluxes because of LUC, with the exception of conversion of (semi)natural land to woodland. This was generally because SOC stocks in the 0–30 cm layer were greater than those in the 30–100 cm layer, and as this study fitted C inputs to measured SOC data for each soil depth and land use in each 1-km square, C fluxes were determined by these differences in the amount of C inputs (as well as the quality of C inputs, defined by the DPM/RPM ratio).

Discussion

Differences between models for estimating SOC changes

Significant differences were observed between soil C fluxes due to LUC predicted by RothC and SEM. Possible reasons for this difference could include (a) differences in input data and (b) differences between the models. Differences in input data could include: resolution of LUC data; assumptions about timing of LUC; some 1-km squares were not modelled with RothC because of missing data as discussed above; the RothC model only used data for the dominant soils to represent 1-km squares; and as the RothC approach assumed gradual LUC transitions following LUCs implemented in 1995, smaller predicted C fluxes would be expected. The differences between the SEM and RothC models including SOC turnover rates, and the influence of soil type and climate on SOC turnover have been discussed in greater detail by Ashman et al. (2000). This comparison of methods provides some measure of uncertainty between SOC models. Previous intercomparisons between RothC and other SOC models at plot scale showed differences of 33–146% in the fit to measured long-term changes in SOC (Falloon & Smith, 2002); at the regional scale differences in C sequestration potential over 100 years of 2–250% were found, and global scale estimates of changes in soil C due to climate change differed between models by 48% over 100 years (Jones et al., 2004).

Limitations in data sets and their application

A number of limitations related to the underlying data sets and their application need also to be considered. The current RothCUK approach uses soil parameter data for the dominant soil of each 1-km square, which could lead to significant errors particularly where component soils are very heterogenous and have a disproportionate effect on C fluxes from the square, such as organic soils. Falloon et al. (2004) investigated the impact of using component or dominant soil data only for thirty 1-km squares in North Derbyshire following conversion of grassland to either arable land or woodland. Estimated differences in mean soil C changes over 20 years were: grassland to arable land 10% (range −30 to +64%, SD 24%), and grassland to woodland 360% (range −78 to +7430%, SD 1366%). It should be noted, however, that the squares investigated included organic soils which were modelled with RothC unmodified – which is likely to exaggerate these estimates of uncertainty. Aggregation of the soil horizon level data into the arbitrary 0–30 and 30–100 cm reference layers could also cause errors in predicted C fluxes particularly where large differences in C concentration between neighbouring horizons occur (e.g. shallow organic layer overlying mineral soil). The derivation of the 1990–2000 LUC matrices at 1-km level is also subject to error, as (a) the source data were derived using different methods and (b) different classification legends were applied. We attempted to minimize these differences through the approaches described earlier, but uncertainty of the matrices may still be significant and difficult to quantify. A considerable remapping exercise would be required to make the 1990 and 1998 land use maps truly comparable.

Soil series were excluded from the study where any parameter values were missing, and as discussed previously, this (a) excluded a large portion of national SOC stocks and (b) mostly affected organic soils. This is likely to have mitigated the uncertainty introduced by using dominant soil data only discussed above. Without a suitable model for SOC dynamics in organic soils (see later discussion), it is difficult to postulate the impact that LUC or climate change might have on SOC stocks in these soils. However, given the overwhelming importance of organic soils (accounting for 43 and 56% of the total SOC stocks of GB and NI, respectively), this clearly deserves further investigation. Obviously, LUC has the potential to alter C stocks in organic soils as in mineral soils – by affecting the balance between C inputs (amount and quality) and decomposition (which LUC could influence by altering soil moisture, temperature, disturbance and aeration), albeit by different mechanisms. Available information on the impact of cultivation on British fenland soils suggests that considerable losses could occur (Bradley, 1997), whilst the impact of afforestation, drainage and other management practices on SOC stocks in organic soils are unclear (Chapman et al., 2001). Very little information was available regarding uncertainty in the national scale input data sets used here. New et al. (2002) used cross-validation to estimate interpolation errors in the climate data used here of 13–20% in precipitation and 0.8–1.1 °C in temperature in Europe. Ground-based surveys have been used to assess the accuracy of the 1990 and 2000 land cover maps at around 80–85% (i.e. 15–20% error: Fuller et al., 1994, 2000). The only published estimate relating to the soil databases used here is an uncertainty in measured SOC pools of around 25% (Milne & Brown, 1997) although this related to an older data set than that used here. For comparison, Jenkinson et al. (1999) reported coefficients of variation (CV) in the UK soil C measurements at the plot scale ranging from 9 to 16%, in rough agreement with the findings of Milne & Brown (1997) and Janik et al. (2002) found three Australian soil types to have CV values ranging from 37 to 40%.

Limitations of the RothUK system

The RothCUK system is based upon the RothC-26.3 model and as such is subject to the same limitations as the RothC model. As discussed above, the model is not generally applicable to permanently waterlogged or organic soils and has not been parameterized for use with subsoils. Thus considerable errors in predicted C flux could occur when applying the model to organic soils, wet soils or subsoils in GB and NI. The only available information on uncertainty relevant to organic soils relates to soil acidity –Jenkinson et al. (1999) fitted the DPM rate constant of RothC to labelled decomposition experiments in strongly acid soils which required the value to be reduced by a factor of four, only resulting in a difference in estimated C input of 0.5% compared with the normal rate constant. The RothC runs in this exercise simply applied the standard RothC model to the 0–30 cm and 30–100 cm layers, although the model was developed for topsoils and has not been specifically calibrated for use with depths much greater than 30 cm. There is evidence that C turnover rates at depth are different from those in the topsoil (e.g. Falloon & Smith, 2000). Sensitivity tests showed that reducing the standard RothC rate constants to 0.25 or 0.5 of their usual value would result in increases in subsoil radiocarbon dates between 8 and 35% (Smith et al., 1999) or errors in estimated C inputs of 7–12% (Jenkinson et al., 1999). A number of processes in topsoil models may, therefore, need altering to allow accurate prediction of C fluxes when subsoils are included. Factors which may need further development for subsoils could include (a) turnover rate of SOC, (b) the proportion of IOM at depth, (c) water balance (the current model uses the same data as the topsoil), (d) temperature balance (the current model uses air temperature), and (e) C input quality, which may be different to that for the topsoil.

Sensitivity analyses of model predictions in relation to input data can provide insights into uncertainty in model predictions. A comprehensive sensitivity analysis of RothC to input data was performed by Janik et al. (2002) in preparation for using the model in the Australian National Carbon Accounting System. Changes in SOC stocks at the end of 40-year simulations because of 10% changes in initial inputs showed that the potentially most sensitive model variables were the initial size (0.72 t C ha−1) and decomposition rate (0.64–0.72 t C ha−1) of the RPM pool and variables associated with carbon input (0.19–1.10 t C ha−1), the size of the original carbon pools (0.24–0.72 t C ha−1), highest for RPM, lowest for HUM and IOM). Relatively little sensitivity to clay content (0.03–0.04 t C ha−1) was shown; moderate sensitivity to monthly rainfall (0.09–1.89 t C ha−1) and temperature (1.28–1.57 t C ha−1), and very small sensitivity to evaporation (<0.1 t C ha−1) was shown. Janik et al. (2002) also considered likely uncertainty in these variables, assessing decomposition rates, initial pool sizes and clay content as moderately uncertain, C input parameters with varying uncertainty (low to high), and climate data as having low-moderate uncertainty, especially evaporation. This could affect the contribution to overall uncertainty – for example, model response to the rate constant for the BIO pool was very low, although uncertainty in the parameter considered moderate – although these would produce only very small uncertainty overall in combination.

Ideally, IOM values for RothC would be set from radiocarbon dates although these data are expensive to obtain and rarely available, particularly at regional or national scales. However, relatively small errors are introduced by the use of default IOM values in RothC simulations such as those presented here. Falloon et al. (2000) showed that errors in SOC changes introduced by estimating IOM from the 95% confidence limits of the equation used in this paper (Falloon et al., 1998) only resulted in errors in a SOC change of 3–19% at the site scale, or 2–36% at the European scale depending on the area affected by LUC. Similarly Jenkinson et al. (1999) found that an 89% increase in the size of the IOM pool only resulted in an 8% difference in estimated C inputs at the plot scale. There is little information on model pool sizes at the regional scale, although CV for the IOM and RPM pools of three Australian soil types were shown to vary from 41 to 80% and from 66 to 114%, respectively (Janik et al., 2002).

We implemented a step change in land use over the whole grid square in the work presented here, although the timing and spatial allocation of LUC could affect results. Falloon et al. (2004) showed that over a 10-year period, the maximum soil C change that could be expected from conversion of grassland to arable or arable to woodland was −14 or +30%, respectively, assuming the change in land use to occur over the whole plot in year 1. The minimum soil C change because of the same land use conversions would result from changing land use in year 10, resulting in soil C changes of −2 and +5%.

Uncertainty in C inputs

Fitted C inputs across GB and NI showed considerable variation, and in some cases large and unreasonable values. Possible reasons for errors in predicted C inputs could include (a) measured SOC not being at equilibrium (as assumed in the model runs), (b) management practices not considered in modelling runs, or (c) soil/climatic conditions outside the range of model applicability. Whilst it is clear from experimental data that soil C stocks do reach equilibrium in the long term (e.g. Jenkinson, 1990; Poulton et al., 2003), and that the RothC model is sensitive to previous land use (Romanyáet al., 2000), it is possible that a number of the soils from the 1-km database were not at equilibrium. Indeed without time-series data on SOC, it would be difficult to conclude that SOC was at equilibrium at the time of sampling as land use history could have a significant impact on SOC changes. Jenkinson et al. (1999) showed that a 50% reduction in C inputs previous to the present day could result in errors in estimated present day C inputs of 90% if the reduction occurred 1 year before, or almost zero if the reduction occurred over 200 years ago. If SOC was actually increasing at sampling time (e.g. recently converted grassland from arable land), this would result in an over prediction of C input; conversely, if SOC was decreasing (e.g. conversion of grassland to arable land), this would result in an under prediction of C input values. General land management practices were assumed for the model runs presented here – for example, additions of organic amendments and cereal straw to arable land were not specified explicitly. If the model was being used to estimate C inputs for a soil receiving additional organic inputs which were not modelled, this would result in an underestimate of C input. Finally, the RothC model was not developed for permanently waterlogged or organic soils or subsoils, and the application of the model to these conditions would be expected to lead to erroneous values of C input – for example, very large C inputs for organic soils as a consequence of over-estimated decomposition rates. Considerable research would be required to develop a dynamic model suitable for organic soils.

Errors in predicted SOC values because of uncertainties in the size and quality of C inputs have been assessed by Falloon (2001). Halving or doubling the quantity of C inputs when simulating a 120-year experiment resulted in differences in SOC of −41 and +80%, respectively, reducing the fit to measured data by 28–71%. Altering the quality of C inputs via the DPM/RPM ratio resulted in errors in modelled SOC of 0.2–11.6%, similar to the errors in C input of 5–8% calculated by Jenkinson et al. (1999). Long-term data sets of changes in SOC are often used to evaluate SOC models and estimate C inputs. However, most long-term data sets have only mean SOC values and no estimate of the error about the mean. Falloon & Smith (2003) showed that when using data sets that do not include estimates of the error about the mean, it is not possible to reduce the error between modelled and measured SOC below 6.5–8.5% even with site specific calibration. Equivalent errors for model runs using regional default C input values are 12–34%. RothC does provide reasonable estimates of initial SOC stocks globally (Jones et al., 2004) and regionally (P. Falloon, unpublished data) when driven with C inputs and land cover history from plant production models rather than using default or fitted C inputs.

Future work and uncertainty in perspective

The work described suggests that recent LUC in GB has the potential to decrease SOC stocks. Currently no comparable 1-km resolution LUC matrices are available for NI. We have attempted to collate information relating to the various components of uncertainty in our study which has indicated that only moderate uncertainty is likely to be introduced from climate and soil data compared with the size of estimated soil C changes (1–6% of our estimates on an equivalent time basis), although using dominant soil data only could introduce large errors (10–360%); the timing of LUC has only small impacts on results (−2 to +30%). Uncertainty in input soil, climate and land use databases is likely to be <25%. The largest source of error is likely to be estimates of C input, perhaps in the order of 12–34% (Falloon & Smith, 2003). It is not possible to postulate how these component sources of uncertainty could combine and propagate in a system such as RothCUK without a detailed uncertainty analysis. However, recent applications of modified IPCC methods for soil carbon accounting in the UK (Milne et al., 2003), Canada (VandenBygaart et al., 2004) and the US (Ogle et al., 2003) have used a Monte-Carlo approach for uncertainty analysis. This involved varying key input parameters between defined ranges and examining the effect on model outputs. In summary, these three studies found uncertainty around the mean estimates of 40–60% (range, Milne et al., 2003), 43–45% (95% confidence intervals, VandenBygaart et al., 2004), and 39–40% (95% confidence intervals, Ogle et al., 2003). Application of RothC to Belgian cropland soils combined with spatial soil C measurements in three time slices (1960, 1990 and 2000) used comparison between modelled and observed SOC values at the three time periods to give uncertainty ranges of 7.5–14% (van Wesemael et al., 2005), similar to errors in modelling long-term experimental sites (e.g. Falloon & Smith, 2002). In light of these studies, a tentative uncertainty range of 50–100% in our estimates seems reasonable.

The 1-km LUC modelling approach has been applied in a simple manner and there is a considerable scope for further studies investigating the effect of LUC and climate change in combination, and investigating different assumptions on the timing, nature and spatial allocation of LUC. In particular, only the effects of one LUC event were examined (1990–2000) without taking into account the continuing effects that earlier LUCs might have had on soil C fluxes. Finally, considerable improvements in the prediction of dynamic interactions between vegetation, climate and soil under LUC and climate change would result from the application of a linked soil-vegetation model. There is also scope for investigating how scenarios of LUC for C mitigation might be implemented and how they might interact with LUC and climate change.

Conclusions

We have developed an integrated data and modelling system for estimating changes in C stocks from mineral soils caused by changes in climate, land use and land management at the 1-km scale in the UK. This system integrates national scale data sets of soil properties, land use and climate with the RothC model. A simple example application of the system to estimate changes in soil C stocks due to LUC over the period 1990–2000 predicted LUC to be a net source of CO2 from 1990 to 2000. This agrees with the SEM method currently used to calculate C fluxes because of LUC for the UK National Greenhouse Gas Inventory, although the RothC estimate was smaller than the SEM estimate. Depending on the particular LUC in question, LUC could either be a source or a sink of CO2. The largest single contributor to soil C fluxes from LUC was conversion of grassland to arable land. Differences between the results may be attributed to differences in the two models, and assumptions and underlying data used in making the calculations. Based on previous studies, an uncertainty range in our estimates of ±50–100% seems plausible. Errors in predicted C fluxes from RothCUK could arise from: (a) using only dominant soil data to represent soil properties of 1-km squares; (b) the inability of the RothC model to adequately predict soil C dynamics in organic and waterlogged soils and subsoils; (c) assumptions regarding the nature, spatial allocation and rate of LUC; (d) bulking data from component soil profiles into the 0–30 cm and 30–100 cm reference layers currently used; (e) excluding soils with missing parameter data from our studies; and (f) uncertainties in input data. Nevertheless, the system provides a powerful method for estimating C fluxes, enabling areas and management systems with particularly large changes in soil C stocks to be located at fine resolution, and allowing for interactions between LUC and climate change to be predicted. A preliminary assessment of the impact of climate change on UK soil C stocks using RothCUK is presented by Falloon et al. (2004).

Acknowledgements

This work contributes to the UK Defra projects CCO2421 ‘Modelling soil carbon fluxes and LUC for the National Carbon Dioxide Inventory’ and GA01054 ‘UK Emissions by sources and removals by sinks due to land use, LUC and forestry (LULUCF)’. Rothamsted Research receives grant-aided support from the Biotechnology and Biological Sciences Research Council. Part of the contribution of PF was supported by the Defra contract PECD 7/12/37.

Ancillary