Biogeosciences, 8, 1053���1065, 2011 www.biogeosciences.net/8/1053/2011/ doi:10.5194/bg-8-1053-2011 �� Author(s) 2011. CC Attribution 3.0 License. Biogeosciences Spatial distribution of soil organic carbon stocks in France M. P. Martin1, M. Wattenbach2, P. Smith3, J. Meersmans1, C. Jolivet1, L. Boulonne1, and D. Arrouays1 1INRA Orleans, �� InfoSol Unit, US 1106, CS 40001, Ardon, 45075, Orleans �� cedex 2, France 2Freie Universit�� at Berlin, Institute of Meteorology, Carl-Heinrich-Becker-Weg 6���10, 12165 Berlin, Germany 3Institute of Biological & Environmental Sciences, University of Aberdeen, Cruickshank Building, St. Machar Drive, Aberdeen, AB24 3UU Scotland, UK Received: 8 September 2010 ��� Published in Biogeosciences Discuss.: 18 November 2010 Revised: 10 February 2011 ��� Accepted: 1 April 2011 ��� Published: 4 May 2011 Abstract. Soil organic carbon plays a major role in the global carbon budget, and can act as a source or a sink of at- mospheric carbon, thereby possibly influencing the course of climate change. Changes in soil organic carbon (SOC) stocks are now taken into account in international negotiations re- garding climate change. Consequently, developing sampling schemes and models for estimating the spatial distribution of SOC stocks is a priority. The French soil monitoring network has been established on a 16 km �� 16 km grid and the first sampling campaign has recently been completed, providing around 2200 measurements of stocks of soil organic carbon, obtained through an in situ composite sampling, uniformly distributed over the French territory. We calibrated a boosted regression tree model on the ob- served stocks, modelling SOC stocks as a function of other variables such as climatic parameters, vegetation net primary productivity, soil properties and land use. The calibrated model was evaluated through cross-validation and eventually used for estimating SOC stocks for mainland France. Two other models were calibrated on forest and agricultural soils separately, in order to assess more precisely the influence of pedo-climatic variables on SOC for such soils. The boosted regression tree model showed good predictive ability, and enabled quantification of relationships between SOC stocks and pedo-climatic variables (plus their interac- tions) over the French territory. These relationships strongly depended on the land use, and more specifically, differed be- tween forest soils and cultivated soil. The total estimate of SOC stocks in France was 3.260 �� 0.872 PgC for the first 30 cm. It was compared to another estimate, based on the Correspondence to: M. P. Martin (manuel.martin@orleans.inra.fr) previously published European soil organic carbon and bulk density maps, of 5.303 PgC. We demonstrate that the present estimate might better represent the actual SOC stock distri- butions of France, and consequently that the previously pub- lished approach at the European level greatly overestimates SOC stocks. 1 Introduction The increasing concentration of greenhouse gases in the at- mosphere has led to the need for reliable estimates of the amounts of organic carbon that might be sequestered by soils (Batjes, 1996 Eswaran et al., 1993 Lal, 2004 Paustian et al., 1997 Post et al., 1982 Saby et al., 2008a Schlesinger, 1991). Indeed, the organic matter contained in the earth���s soils is a large reservoir of carbon (C) that can act as a sink or source of atmospheric CO2. The world���s soils represent a large reser- voir of C of about 1500 PgC (Batjes, 1996 Eswaran et al., 1993 Post et al., 1982). Accurate estimates of this pool are needed. However their reliability depends upon suitable data in terms of organic carbon content and soil bulk den- sity and on the methods used to upscale point data to com- prehensive spatial estimates. There are, therefore, few pre- cise assessments of soil organic carbon (SOC) stocks based on measurements over large areas since systematic sampling schemes including SOC, bulk density and rock fragment con- tent are quite rare (Morvan et al., 2008), and because high spatial variability of SOC requires a very high sampling den- sity to get accurate estimates (Bellamy et al., 2005 Saby et al., 2008b). Several approaches involving empirical mod- els to upscale SOC point measurements to the national level are found in the literature. These approaches range from Published by Copernicus Publications on behalf of the European Geosciences Union.

1054 M. P. Martin et al.: Spatial distribution of soil organic carbon stocks in France simple statistics or pedotransfer rules, relating SOC contents or stocks to soil type (Yu et al., 2007) or soil type and land use (Tomlinson and Milne, 2006 Arrouays et al., 2001), to multivariate statistical models (Meersmans et al., 2008, with multiple linear models and Yang et al., 2008, with gener- alized linear models). Recent studies have used techniques adapted from the data mining and machine learning litera- ture, with piecewise linear tree models (Bui et al., 2009) or multiple regression trees for regional studies (Grimm et al., 2008 Lo Seen et al., 2010). Despite the spatial dimension of such studies, few geostatistical approaches have been pro- posed for use at the national scale (but see Chaplot et al., 2009), mainly because of the difficulty of including the ef- fect of the different drivers of SOC dynamics in geostatistical models. Jones et al. (2005) developed a methodology for esti- mating organic carbon concentrations (%) in topsoils (oc- top) across Europe and recently published a map of SOC stocks by country. The information is available as a database which can be downloaded from the EU-soils web site (http: //eusoils.jrc.it). This methodology, based on pedotransfer functions, gave results which were validated using data from England and Wales and Italy (Jones et al., 2005). However, the match between country level estimates of SOC stocks us- ing this method and estimates based on national databases depends on the country and may sometimes be poor. For instance, SOC stocks for the first 1 m in Denmark was esti- mated to vary from 0.563 to 0.598 PgC, among which 60% is found in the 0���28 cm layer (Krogh et al., 2003). Thus, the amount can be rescaled to 0.338 to 0.359 PgC, for the first 28 cm layer, compared to the Joint Research Center (JRC)���s estimate of 0.6 PgC for only the first 30 cm (Hiederer, 2010). The issue of accurately assessing SOC stocks at the coun- try level is critical, because SOC stocks are used as input for studies on the impact of future land use change or climate change on SOC stocks dynamics, and on potential green- house gases (GHG) emissions (Chaplot et al., 2009). For instance, they may be used for defining the baseline state for SOC change simulations (van Wesemael et al., 2010), or setting some of the models��� parameters (Tornquist et al., 2009). In this paper, we apply a new methodology: boosted regression trees (BRT), already successfully applied in India (Lo Seen et al., 2010), to predict the geographical distribu- tion of SOC stocks in metropolitan France from a set of 1974 paired observations of SOC and bulk densities. We examine the effects of the main controlling factors of SOC stocks dis- tribution. We estimate the uncertainty of our national esti- mate and compare the results with those previously obtained by Arrouays et al. (2001) and Hiederer (2010) on the same territory. Fig. 1. Distribution of the 1974 sites within the French monitoring network which were used in the present study. 2 Materials and methods 2.1 Data 2.1.1 Site specific soil and agricultural data Soil Organic Carbon Stocks were computed for a subset of 1974 sites from the French soil survey network (RMQS), for which analytical data was available (Fig. 1). This dataset covered a broad spectrum of climatic, soil and agricultural conditions. In the near future, the RMQS will cover the entire metropolitan France. The network is based on a 16 km �� 16 km square grid and the sites are selected at the centre of each grid cell resulting in about 2200 soil sampling sites. In the case of soil being inaccessible at the centre of the cell (i.e. urban area, road, river, etc.), an alternative location with a natural (undisturbed or cultivated) soil is selected as close as possible, but within 1 km from the centre of the cell (for more information, see Arrouays et al., 2002). At each site, 25 individual core samples were taken from the topsoil (0���30 cm) using a hand auger according to a strat- ified random sampling design within a 20 m �� 20 m area. In- dividual samples were mixed to obtain a composite sample for each soil layer. Apart from composite sampling, at 5 m from the south border of the 20 m �� 20 m area, a soil pit was dug, from which main soil characteristics were described and 6 bulk density measurements were done, as described previ- ously (Martin et al., 2009). From these data, SOC stocks were computed for the 0���30 cm soil layer. Biogeosciences, 8, 1053���1065, 2011 www.biogeosciences.net/8/1053/2011/