A global inventory of N2O emissions from tropical rainforest soils using a detailed biogeochemical model

Authors


Abstract

[1] Beside agricultural soils, tropical rainforest soils are the main source of atmospheric N2O. Current estimates of the global N2O source strength of tropical rainforest soils are still based on rather simplistic upscaling approaches and do have a large range of uncertainty. In this study, the biogeochemical ForestDNDC-tropica model was recalibrated and intensively tested on the site scale prior to inventory calculations. For this, the model was coupled to a newly developed global GIS database holding relevant information on model initialization and driving parameters in 0.25° × 0.25° resolution. On average, the mean annual N2O emission source strength of rainforests ecosystems worldwide for the 10-year-period 1991–2000 was calculated to be 1.2 kg N2O-N ha−1 yr−1. Using a total rainforest area of 10.9 × 106 km2, this amounts to a total source strength of 1.34 Tg N yr−1. The result of an initialization parameter uncertainty assessment using Latin Hypercube sampling revealed that the global source strength of N2O emissions from tropical rainforests may range from 0.88 to 2.37 Tg N yr−1. Our calculations also show that N2O emissions do vary substantially on spatial and temporal scales. Regional differences were mainly caused by differences in soil properties, whereas the pronounced seasonal and interannual variability was driven by climate variability. Our work shows that detailed biogeochemical models are a valuable tool for assessing biosphere-atmosphere exchange even on a global scale. However, further progress and a narrowing of the uncertainty range do crucially depend on the availability of more detailed field measurements for model testing and an improvement of the quality of spatial data sets on soil and vegetation properties.

1. Introduction

[2] Nitrous oxide (N2O) is one of the most important radiatively active trace gases in the atmosphere contributing approximately 6% to the observed additional global warming. Its atmospheric concentration is increasing at a rate of approximately 0.25% yr−1 [Khalil et al., 2002]. N2O is not only a potent greenhouse gas, but is also involved in the depletion of stratospheric ozone [Crutzen, 1970]. N2O is mainly of biogenic origin and soils are one of the main sources, contributing approximately 60% to the total atmospheric N2O budget. Besides agricultural soils, tropical forest soils have been identified as a major source of N2O. The mean estimate of 3.0 Tg N yr−1 accounts for about 18% of all atmospheric N2O sources [Intergovernmental on Panel Change, 2001]. However, this number is highly uncertain owing to the limited number of field measurements [Breuer et al., 2000].

[3] N2O is produced in soils through the microbial processes of nitrification and denitrification [Conrad, 1996], the key processes of N cycling in terrestrial ecosystems. Nitrification and denitrification can occur simultaneously in soils, although the rates of the two processes depend on soil aeration and micro-site availability of substrates. N2O production, consumption and emission is affected by various biotic processes (e.g., mineralization and plant N uptake), and a wide range of abiotic factors such as soil temperature and moisture, pH, and soil texture. The significant variability of these factors across space and time [Davidson et al., 1998; Papen and Butterbach-Bahl, 1999] has also been demonstrated to feedback on N2O emissions from tropical forest soils [Verchot et al., 1999; Kiese and Butterbach-Bahl, 2002] causing a substantial variation on seasonal and interannual as well as on spatial scales.

[4] Owing to the importance of tropical rainforest soils as a major source of atmospheric N2O a number of upscaling attempts have been conducted. This includes compilation, comprehensive statistical analysis and finally extrapolation of available measurements to global scales by considering the spatial extent of vegetation types and soil properties [Matson and Vitousek, 1990; Bouwman et al., 1993; Breuer et al., 2000; Stehfest and Bouwman, 2006]. However, such upscaling approaches are hampered by their dependence on a relatively limited number and temporal as well as spatial coverage of predominantly chamber based measurements (see compilation of flux measurements given by Breuer et al. [2000]). Mechanistic models coupled to Geographic Information Systems (GIS) have been increasingly used for calculating regional and global trace gas emission inventories [Parton et al., 1996; Potter et al., 1996; Butterbach-Bahl et al., 2001; Kesik et al., 2005; Kiese et al., 2005; Li et al., 2005, Del Grosso et al., 2006]. Since mechanistic models calculate the effect of changes in driving variables on processes, which are assumed to function the same way in all ecosystems (e.g., control of soil microbial C and N turnover by moisture and temperature), these models have the potential to account for a wide range of natural occurring parameter combinations. If one postulates that process-oriented models such as ForestDNDC-tropica are capable of capturing the major biogeochemical processes within ecosystems, the quality of their spatial extrapolation should be superior as compared to empirical methods. Furthermore, as already pointed out by Del Grosso et al. [2006], the use of biogeochemical models to simulate N trace gas emissions from soils on site and regional scales implies the use of daily simulation time steps to account for feedbacks of rapid changes in environmental conditions on the production, consumption and emission of trace gases.

[5] Therefore the main objective of this study was (1) to calculate a global N2O emission inventory for tropical rainforest ecosystems using a detailed mechanistic model, (2) to identify spatial and (3) temporal variability of emissions, and (4) to assess the uncertainty of estimates. This required the further development of the ForestDNDC-tropica model, involving Bayesian calibration techniques, intensive model testing and the development of a global GIS database for initializing and driving the model.

2. Materials and Methods

2.1. ForestDNDC-Tropica Model Description

[6] The ForestDNDC-tropica model used in this study is a follow-up version of the PnET-N-DNDC model, which was initially developed to predict soil carbon and nitrogen biogeochemistry, including N-trace gas emissions, in temperate forest ecosystems [Li et al., 2000; Stange et al., 2000; Butterbach-Bahl et al., 2001], and was recently adapted for tropical forest conditions [Kiese et al., 2005]. The model consists of submodels for the simulation of soil climate, decomposition and forest growth. Within the soil climate submodel, daily climate data is used to calculate soil temperature, moisture and oxygen profiles from one-dimensional thermal-hydraulic flow and gas diffusion equations [Li et al., 2000]. This is done by considering soil physical properties (texture) and plant and microbial turnover processes of C, N and water. Forest growth is calculated depending on solar radiation, temperature, water and nitrogen availability. Litter production, water and nitrogen demand of plants and root respiration is linked with the soil climate and the decomposition submodel. Decomposition of organic matter increases concentrations of dissolved organic carbon (DOC), ammonium (NH4+), and carbon dioxide (CO2). Decomposition is based on soil environmental conditions and specific decay rates for different organic matter pools. N-trace gas production, consumption and emission are calculated within the submodels nitrification and denitrification. The processes are based on simulated soil microbial activities, which depend on soil environmental conditions and a series of biochemical and geochemical reactions determining the transport and transformation of C and N components [Li et al., 2000]. Aerobic nitrification and anaerobic denitrification are simultaneously calculated using the concept of a dynamic “anaerobic balloon.” By using this approach, substrates (DOC, NH4+ and nitrate [NO3]) are allocated into aerobic and anaerobic soil compartments on the basis of the oxygen concentration in the respective soil layer [Li et al., 2000].

[7] Compared to the ForestDNDC-tropica version used by Kiese et al. [2005], we revised the distribution of soil organic carbon (SOC) in the soil profile. The previous distribution for Australian rainforest soils was based on an exponential decay function derived from SOC concentration measurements in 5 and 30 cm soil depth. Since the ISRIC-WISE profile data used in this study only provides information on a mean SOC content (%) for the top soil (0–30 cm; see section 2.2.1) a revised exponential decay function was introduced.

equation image

where SOCdepth [%] = SOC content in a given soil depth [cm].

[8] The dimensionless shape parameters y0, A and t were extracted from data provided by Jobbagy and Jackson [2000] and differed between the two forest types considered in this study (evergreen tropical rainforest; seasonally dry tropical rainforest),

[9] Evergreen tropical rainforest

equation image

[10] Seasonally dry tropical rainforest

equation image

[11] On the basis of the simulated SOC profile, we recalculated the bulk density provided for 0–30 cm (ISRIC-WISE, BD0–30) for individual soil depths (BDdepth). We first calculated the density of mineral soil (DBM [g cm−3]) at 0% SOC content,

equation image

where DBO = density bulk organic matter = 0.3 g cm−3 [Li et al., 1992, 2000].

[12] We then considered the recalculated SOC profile for the calculation of bulk density (BDdepth [g cm−3]) with depth,

equation image

[13] Finally, the corresponding soil porosity for each soil layer (POROdepth) was calculated using a value of 2.65 g cm−3 for the specific weight of soil minerals [Schachtschabel et al., 1992],

equation image

[14] In accordance with previous applications of PnET-N-DNDC and ForestDNDC [e.g., Kesik et al., 2005; Kiese et al., 2005], the simulation depth was limited to 30 cm and subsoil data (30–100 cm) from the ISRIC-WISE data set was not used.

[15] Since soil physical characteristics of tropical soils (e.g., Andosols and Ferralsols) substantially deviate from temperate soils, specific pedo-transfer functions (PTF) are essential to correctly derive soil hydraulic properties like field capacity, wilting point and saturated hydraulic conductivity. PTFs specifically designed for tropical soils were recently developed by Hodnett and Tomasella [2002] and were newly implemented in the ForestDNDC-tropica model. The Hodnett and Tomasella approach comprises a series of algorithms using information on texture (sand: SA [%], silt: SI [%], clay: CL [%]), soil organic carbon (SOC0–10 [%]), bulk density (BD0–10 [g cm−3]), cation exchange capacity (CEC [cmol kg−1]) and pH in order to derive parameters of the van Genuchten equation [van Genuchten, 1980, equation (5)], which was finally used to calculate soil specific field capacity and wilting point for soil climate submodel initialization [Li et al., 1992, 2000].

equation image

where

Θ

= volumetric water content [m3 m−3];

Θs

= saturated water content [m3 m−3];

Θr

= residual water content [m3 m−3];

ψ

= absolute value of the matric potential [kPa];

α, n, m

= van Genuchten shape parameters.

equation image
equation image
equation image
equation image
equation image

[16] The initialization of wood, leaf, and floor mass for ForestDNDC-tropica was extracted from the output of a vegetation model (see section 2.2.3), thus replacing the former model internal initialization, which was specifically designed for Australian rainforest ecosystems. The fine root mass was estimated as 0.7 × leaf mass [Kiese et al., 2005] (also data compilation of Vogt et al. [1996]). The ForestDNDC-tropica model was run for a two years simulation period, using the first year as spin-up period in order to allow model internal C and N pools to equilibrate.

2.2. GIS Database

[17] For the calculation of a global N2O emission inventory soil, vegetation and climate data was integrated into a raster GIS database covering land areas from 30°S to 30°N (grid resolution 0.25° × 0.25°) and stored in a plain geographic projection. An area correction based on latitude was applied in order to report the true spatial extent of each grid cell. The model was only applied for grid cells with forest cover >10%.

[18] To delineate actual tropical rainforest areas we combined a high-resolution, satellite-based land cover map (GLC2000) [Bartholomé and Belward, 2005] (also Joint Research Center, GLC2000: The Land Cover of the World in the Year 2000, available at http://www-gvm.jrc.it/glc2000/objectivesGLC2000.htm) with an eco-zonal classification [Blasco et al., 2000]. The eco-zonal classification allows to account for differing bioclimatic limits and conditions of tropical rainforests worldwide and to discriminate between humid (seasonally dry) and perhumid (evergreen) tropical rainforest (Figure 1). The location of the GLC2000 land cover classes “broadleaved-evergreen”, “broadleaved-deciduous”, “mixed leaf type” and “regularly flooded” were considered as potential tropical rainforest areas. By using the long-term climatic criteria [see Blasco et al., 2000], we derived the areas covered either by humid or perhumid tropical rainforests (Terrestrial Air Temperature and Precipitation: Monthly and Annual Climatologies, Version 3.02; 0.5° resolution [Willmott and Robeson, 1995] (also C. J. Willmott and K. Matsuura, Terrestrial air temperature and precipitation: Monthly and annual climatologies (version 3.02), 2001, http://climate.geog.udel.edu/∼climate/). Montane tropical rainforests, shown in Figure 1 for consistency with work of Blasco et al. [2000], were not simulated as an independent ForestDNDC-tropica forest class, as temperature is already explicitly affecting vegetation growth and microbial process activity in the ForestDNDC-tropica model.

Figure 1.

Bioclimatic forest classification.

[19] In our study, total rainforest area (humid and perhumid tropical rainforest) was 10.93 × 106 km2. This number is close to the reported 11.16 × 106 km2 of tropical rainforests for the year 1997 [Achard et al., 2002]. However, some previous global N2O emission inventories for tropical rainforest soils were based on a considerably larger area (e.g., 14.96 × 106 km2 [Matson and Vitousek, 1990; Breuer et al., 2000]), which is partly due to a different ecological classification scheme. To compare our estimates with previous estimates of N2O emissions from rainforests we scaled the estimates of other studies to the spatial extend of tropical rainforest ecosystems used in this study.

2.2.1. Soil Data

[20] We used the FAO Digital Soil Map of the World (version 3.6 [Food and Agriculture Organization, 2003], soil properties linked using the 1974 classification scheme [FAO74]) as the spatial base layer of our GIS database and converted the original vector file into a grid file with 0.25° × 0.25° resolution. Topsoil (0–30 cm) information of sand, silt and clay fraction, gravel, SOC, mineral soil pH, bulk density and CEC was taken from the ISRIC-WISE database (version 2.2 [Batjes, 2002a]). This database was constructed from homogenized profile information of 9607 profiles sampled worldwide with focus on tropical and subtropical regions [Batjes, 2002a]. The homogenized ISRIC-WISE soil database does not distinguish between agricultural and natural land cover. Since agricultural soils tend to have higher soil bulk densities as compared to soils under natural vegetation, a correction factor (0.907; r2 = 0.56) was applied on the basis of the correlation of the mean bulk densities of FAO74 soil subunits (predominantly agricultural soils) with a specific collection of homogenized primary soil profile data for tropical rainforests soils (ISRIC-WISE Global Soil Profile Data Set, Version 1.1 [Batjes, 2002b]). Even though we have to assume that SOC content may be also biased, we could not find a significant correlation to derive a SOC correction factor using the same approach as carried out for bulk density. Soil organic carbon content was thus used as given in the original ISRIC-WISE soil database. The soil properties were finally linked to the FAO soil map by classification codes (FAO74 soil subunits; including texture class, e.g., xanthic Ferralsol; FAO texture class b). Characteristics and spatial extent of the most dominant soil types of our simulation domain are presented in Table 1 and Figure 2.

Figure 2.

Spatial distribution of selected model input parameters: (a) aboveground biomass, (b) soil organic carbon, (c) soil clay content, and (d) soil pH.

Table 1. Soil Characteristics of the Most Prominent Soil Types of the Simulation Domaina
FAO ClassFAO SubclassFAO TextureArea, 106 km2Cover, %Sand, %Silt, %Clay, %SOC, %pHCEC, cmol kg−1BD, g cm−3Gravel, %
  • a

    Simulation domain is >1% of total simulated forest area. Abbreviations: BD, bulk density corrected by 0.904; CEC, cation exchange capacity; SOC, soil organic carbon. Soil percentages include ± standard deviation.

  • b

    No standard deviation is given in soil database.

FerralsolsxanthicB1.7415.7464.2 ± 12.211.9 ± 10.223.9 ± 4.30.94 ± 0.344.6 ± 0.63.8 ± 2.41.21 ± 0.108.0 ± 7.0
 C0.978.7831.8 ± 16.213.8 ± 7.254.4 ± 16.31.84 ± 0.554.3 ± 0.36.0 ± 3.01.02 ± 0.151.0 ± 0.0
 A0.403.6479.6 ± 5.66.9 ± 4.613.6 ± 3.70.78 ± 0.254.9 ± 0.52.8b1.25 ± 0.112.6 ± 1.4
orthic C1.5914.3527.5 ± 16.320.5 ± 13.152.1 ± 12.01.91 ± 0.764.9 ± 0.611.4 ± 8.11.13 ± 0.173.9 ± 2.9
 B0.575.1259.6 ± 13.114.0 ± 13.126.4 ± 5.31.36 ± 0.505.0 ± 0.65.1 ± 3.11.22 ± 0.132.8 ± 1.5
 Bc0.141.3143.6 ± 17.717.2 ± 13.639.3 ± 8.41.63 ± 0.634.9 ± 0.68.2 ± 5.51.18 ± 0.143.4 ± 2.2
AcrisolsorthicBc0.696.2539.1 ± 15.926.0 ± 14.334.9 ± 7.51.24 ± 0.494.7 ± 0.611.0 ± 3.91.19 ± 0.148.6 ± 5.8
 B0.363.2849.3 ± 16.827.1 ± 16.223.6 ± 5.71.21 ± 0.534.8 ± 0.79.8 ± 3.91.24 ± 0.1411.6 ± 8.8
 C0.201.8429.0 ± 13.624.9 ± 12.546.2 ± 8.81.28 ± 0.434.7 ± 0.512.2 ± 3.71.14 ± 0.135.5 ± 3.3
plinthicB0.171.5247.6 ± 19.129.6 ± 18.922.9 ± 6.61.40 ± 0.704.7 ± 0.75.3 ± 0.31.29 ± 0.1329.1 ± 19.2
 C0.121.0622.2 ± 14.631.8 ± 14.646.3 ± 13.01.88 ± 0.864.5 ± 0.621.1b1.17 ± 0.195.8 ± 5.2
Arenosols A0.423.8289.5 ± 5.45.3 ± 4.35.3 ± 2.90.48 ± 0.146.0 ± 0.94.1 ± 2.01.36 ± 0.134.0 ± 3.4
NitosolsdystricBc0.181.5933.2 ± 15.828.1 ± 11.338.8 ± 10.71.30 ± 0.385.3 ± 0.67.7 ± 3.91.12 ± 0.107.7 ± 4.0
 C0.181.5822.3 ± 13.623.1 ± 7.654.8 ± 12.11.52 ± 0.475.7 ± 0.59.5 ± 3.81.08 ± 0.061.0 ± 0.0
GleysolsdystricC0.332.9915.9 ± 13.228.9 ± 11.955.3 ± 14.92.07 ± 0.854.7 ± 0.420.4 ± 15.51.00 ± 0.2414.0b
CambisolsdystricB0.211.8941.7 ± 15.438.7 ± 13.119.9 ± 8.21.99 ± 1.175.1 ± 0.717.5 ± 11.71.14 ± 0.1810.3 ± 7.4

2.2.2. Weather Data

[21] Daily weather data covering the tropical belt (30°S–30°N) for the years 1991–2000 was provided by the European Centre for Medium-Range Weather Forecasts (ECMWF). Six-hourly postprocessed data of maximum and minimum temperature and photosynthetically active radiation (PAR) was aggregated to daily fields (ECMWF ERA40 atmospheric forecast [Uppala et al., 2005]). Daily total precipitation at surface level was also provided by ECMWF from the Tropical Ocean and Global Atmosphere (TOGA) forecast product (ECMWF/WCRP TOGA Level III/operational archive). All data sets were resampled to 0.25° × 0.25° grids and converted to ForestDNDC-tropica climate input file format.

2.2.3. Vegetation Data

[22] We used the Lund-Potsdam-Jena Dynamic Global Vegetation Model (LPJ-DGVM [Sitch et al., 2003]) with an updated hydrology module [Gerten et al., 2004] to derive a global distribution of aboveground biomass estimates for tropical rainforests ecosystems. The LPJ-DGVM model simulates vegetation dynamics and structure based on plant functional types (PFT) which represent individual physiological, morphological, phaenological, bioclimatic and fire-response attributes [Smith et al., 2001; Sitch et al., 2003].

[23] The landmasses stretching from 30°N to 30°S were initialized as bare ground with the FAO texture classes in 0.5° × 0.5° resolution. Using interannually varying climate conditions of annual average temperature, precipitation and cloudiness [see Sitch et al., 2003] the LPJ-DGVM model was run for 900 simulation years in order to reach equilibrium conditions for the soil and vegetation pools. For the following 100 simulation years, LPJ-DGVM was run with monthly climate data of the Climate Research Unit (CRU 2000, University of East Anglia [Mitchell et al., 2004]). The wood mass and leaf mass of the PFT composition (Figure 2a), as well as the floor mass, were extracted from the LPJ-DGVM output and used as input drivers for ForestDNDC-tropica, whereas the soil C concentration was taken from the ISRIC-WISE database. The leaf C:N ratios for the two rainforest classes (humid or perhumid tropical rainforest) were set to 45 and 30, respectively. These ratios were based on an updated literature review by Kiese et al. [2005]. An overview of all data sets and models used in this study, as well as their connection, is given in Figure 3.

Figure 3.

Study setup for calculation of the global N2O inventory for tropical rainforest soils.

2.3. Model Calibration and Uncertainty Analysis

2.3.1. Model Evaluation

[24] Changes and improvements of various sections of the model code and model structure affecting vegetation initialization, soil hydrology and vertical soil carbon distribution, required a profound recalibration of the model. This was done by the Bayesian Calibration (BC) technique using a range of detailed field data sets. The strength of this method in calibrating and evaluating process-based models was outlined recently by van Oijen et al. [2005]. We used the algorithm described by van Oijen et al. [2005] and based the model calibration on three data sets (Bellenden Ker/Australia 2002 [Kiese et al., 2003]; Kakamega/Kenya 2004 [Werner et al., 2006]; and Xishuangbanna/Southwest China 2005 [Werner et al., 2007]; see Table 2). These N2O emission data sets feature daily temporal resolution and, are therefore well suited for the calibration of the dynamic process interactions simulated in ForestDNDC-tropica also on daily time steps. Furthermore, the sites cover different geographic locations, environmental conditions and show different N2O emission levels. Data sets (n = 12) with low temporal resolution or shorter measurement duration were not used in the calibration but in an independent testing procedure.

Table 2. Characteristics of Sites Used for Model Calibration and Testinga
SiteNumberLocationSimulated YearsClimateSoil PropertiesSource
AT,°CAP, mmSand, %Silt, %Clay, %SOC, kg kg−1pHBD, g cm−3
  • a

    Annual average temperature (AT) and annual precipitation (AP) are given for each simulation year (separation by comma indicates independently treated data; data separated by slash was treated as one data set). Brackets around texture values: values partly missing and replaced with plausible data. Letters in parentheses in some climate and soil properties entries denote that data are from entries in source column followed by the same parenthetical letters.

  • b

    Climate data are from http://www.ots.ac.cr/rdmcnfs/datasets/meteoro/ls_met/.

  • c

    Climate data are from corresponding grid cell (ECMWF climate data set).

  • d

    Climate data are from corresponding grid cell (ECMWF climate data set); as the simulation year 2003 was not available we simulated the years 1994, 1997, 1998 which feature a total annual precipitation close to the long-term mean of 1590 mm reported by Purbopuspito et al. [2006].

  • e

    Entry is adapted to 0–10 cm.

Calibration sites
   Bellenden Ker, Australia1145°54′E 17°16′S200224.343955721220.0314.11.09Kiese et al. [2003]
   Kakamega, Kenya234°52′E 0°15′N200424.916604323340.0756.31.00Werner et al. [2006]
   Xishuangbanna, SW-China3101°12E 21°58′N200521.814935923180.0195.11.30Werner et al. [2007]
Validation sites
   Bellenden Ker, Australia4145°54′E 17°16′S200124.343955721220.0314.11.09Kiese et al. [2003]
   Kauri Creek, Australia5,6145°38′E 17°17′S1997, 200020.2, 20.81374, 2013689230.0325.21.05Breuer et al. [2000]; Kiese and Butterbach-Bahl [2002]
   Lake Echeam, Australia9145°36′E 17°08′S199721.015006811210.0234.80.99Breuer et al. [2000]
   Massey Creek, Australia7,8145°34′E 17°36′S1997, 199819.9/21.41384/18735214340.0525.50.69Breuer et al. [2000]
   La Selva, Costa Rica1084°0′W 10°26′N1990/199125.8/25.9b4463/4548b(12)(15)730.0303.60.65Keller and Reiners [1994]
   Guacimo, Costa Rica1184°0′W 10°26′N1992/199326.0/26.2b3757/3736b(12)(15)730.0303.60.65Liu et al. [2000]
   Central Rondônia, Brazil1262°30′W 10°30′S1992/199324.6/25.0c1950/1513c(60)(19)210.0134.91.28Neill et al. [1995]
   Jambi Province, Indonesia13102°6′E 1°5′S1997/199825.1/25.7c2988/2754c61 (a)17 (a)22 (a)0.0304.21.12Ishizuka et al. [2002]; Ishizuka et al. [2005](a)
   Wuasa, Indonesia14120°17′E 1°25′S1994/1997/1998d24.8/25.6/25.6d1416/1648/1934d6425110.034e6.50.99Purbopuspito et al. [2006]
   Pará, Brazil1547° 31′W 2° 59′S1995/199625.1/24.72595/2419c6 (b)18 (b)76 (b)0.0254.40.99Verchot et al. [1999]; Davidson et al. [2004](b)

[25] The BC program was initialized with the default parameter setup of ForestDNDC-tropica prior to our model changes [Kiese et al., 2005]. Without changing the original process descriptions [Li et al., 2000; Kiese et al., 2005] we defined upper and lower bounds of selected parameters affecting SOC partitioning. The bounds were based on a literature review and were set deliberately wide to enable the exploration of a parameter setup to avoid parameter “lock-in.” A list of the selected parameters, the initial and final values and the upper and lower parameter bounds are given in Table A1 (parameters are referenced by their index key in the manuscript). The target parameters can be grouped into carbon partitioning, nitrification, denitrification and decomposition related parameters. It is important to note that in ForestDNDC-tropica, as in any DNDC-based model, the amount and availability of N is derived from the carbon pools. Therefore carbon partitioning (P01–P05) has a direct impact on the N availability and quality for the microbial processes involved in the production and consumption of N2O. We also included scaling parameters (SP) in this calibration exercise as the lack of measured information for those process steps complicates the manual definition of these process parameters (P06–P08; P13–P15). The drought induced denitrification activity reduction measure introduced by Kiese et al. [2005] was included into the recalibration (P09–P012) as the authors only tested the approach for forest ecosystems of the Australian tropics. For the same reasons we included the factors defining the effect of temperature and moisture on denitrification (P16–P19).

[26] The optimal parameter vector with regard to N2O emissions (but also soil CO2 emissions and soil hydrology) after approximately 10000 chain iterations was finally used for site and regional applications. Model performance was documented for individual sites and the entire test data set using the following measures:

[27] Coefficient of variation

equation image

[28] Model efficiency

equation image

[29] Logarithmic model efficiency

equation image

[30] Normalized root mean square prediction error

equation image

where xmod is the simulated value, equation imagemod is the average of all simulated values, xmeas is the value obtained from field data, equation imagemeas is the average of measured field data, and SD is the standard deviation of measured field data.

[31] We varied soil, vegetation and climate parameters for 1000 randomly selected grid cells to assess model sensitivity (Figure 4). We used the climate data of the year 1995 and input data was changed in three levels (small: ±10%; medium: ±25%; large: ±50%). Model sensitivity was calculated as the variation of predicted annual N2O emission in response to changes in major input parameters (leaf mass, floor mass, wood mass, sand, silt, clay, gravel, pH, bulk density, SOC, CEC, temperature, precipitation, PAR; note that temperature and pH were altered by ±1.0°C/ ±2.5°C/ ±5.0°C and ±0.25/ ±0.5/ ±1.0, respectively). Each parameter was individually increased (P1) or decreased (P0) for the three uncertainty ranges and the sensitivity index β [Friend et al., 1993] was calculated from the resulting change in N2O emissions (P0 → N2O0 or P1 → N2O1),

equation image
Figure 4.

Rainforest simulation area, location of measurement sites (circles, sites used for model calibration; squares, sites used for model testing) and spatial distribution of randomly sampled grid cells used in Latin hypercube and model sensitivity analysis.

[32] Note that the distance of β from zero is proportional to the sensitivity of a given parameter whereas the sign of β indicates the direction of correlation. We reported the average β values for all tested grid cells.

2.3.2. Data Uncertainty

[33] Data sets used for bottom-up upscaling approaches are known to be a major source of prediction uncertainty for biogeochemical models [Li et al., 2004], since subgrid spatial variability in land use and topography and theirs effect on physical and chemical processes do result in large uncertainties for soil and vegetation parameters. Monte Carlo analysis has become a common tool for the assessment of model input parameter induced uncertainty [Li et al., 2004]. A probability-based sample procedure is generally used to derive a parameter vector (model input). In this study we applied the Latin Hypercube Sampling (LHS), which stratifies the parameter range of each input variable in N disjoint segments of equal probability [McKay et al., 1979]. A random number is drawn within each interval and, thus, one random number is guarantied to fall into the bounds of each of the N segments. The obtained N random realizations of each parameters combined to sequences by random permutation are forming the final model input parameter vectors. It has been shown that this method is capable of exploring the input parameter space more exhaustively [e.g., McKay et al., 1979]. When computation time of the model is significant or a large number of parameters are used the LHS permutation process enables the user to use significantly less model runs as compared to simple random sampling in order to asses model uncertainty.

[34] An uncertainty assessment for effects of soil properties on N2O emissions was carried out for the following parameters: sand, silt, clay, gravel, SOC, pH, BD and CEC. Parameter mean, standard deviation and lower and upper limit of the soil parameters were extracted from the ISRIC-WISE soil database (see Table 1 for overview of most prominent soil types). As wood mass, leaf mass and leaf litter mass for each grid cell were simulated using the LPJ-DGVM, lower and upper parameter bounds as well as the standard deviation were set manually. For these parameters we assumed normal distributions, i.e., the standard deviation (SD) was set to ±33% of the mean value (1 SD), whereas the lower/upper bounds were set to ±3 SD.

[35] On the basis of 106 model runs (1000 randomly sampled grid cells ×200 repetitions ×5 realizations), the first quartile (Q1) and third quartile (Q3) of all calculated N2O emissions are considered to be the lower and upper bounds of data induced model uncertainty and were thus applied to the final inventory calculations as lower and upper uncertainty bounds for data-induced uncertainty.

[36] ForestDNDC-tropica uses daily weather parameters (minimum temperature, maximum temperature, precipitation, PAR) as model drivers. These parameters are highly variable throughout the year and season and it needs to be assumed that also these meteorological parameters have significant subgrid variability. Climatic parameters also predominately modulate the effects of the defining soil and vegetation parameters on the production, consumption and release of N2O and a simple variation of daily temperature (e.g., ±3°C) or precipitation (e.g., ±25%) is not suitable to evaluate climate data induced uncertainty. Since we used 1000 randomly sampled grid cells from throughout the GIS database covering all climatic regions (same sample as used for the sensitivity analysis; see Figure 4 for spatial distribution of those grid cells), thus covering various meteorological regimes, we decided not to include meteorological parameters explicitly in this analysis. Please see section 3.2 for information about the general sensitivity of ForestDNDC-tropica on changes of climatic parameters instead.

3. Results and Discussion

3.1. Model Calibration and Testing

[37] By applying the Bayesian Calibration technique, we obtained optimized parameter settings for SOC partitioning and N2O production during nitrification and denitrification for the three calibration data sets Bellenden Ker 2002, Kakamega 2004 and Xishuangbanna 2005. Model performance given as coefficient of variation (r2) after optimization was 0.48, 0.78 and 0.70, respectively. As illustrated by Figure 5, ForestDNDC-tropica was capable of simulating temporal fluctuations of N2O emissions as induced by precipitation events, but also captured the significant differences in the magnitude of N2O emissions between the sites (mean measured N2O emission [g N ha−1 day−1]: Bellenden Ker 3.8 ± 0.2, Kakamega 9.9 ± 0.9, Xishuangbanna 1.4 ± 0.1; see also Table 3). Up-to-date, the most detailed N2O emission data set was obtained by Kiese et al. [2003] at the tropical lowland rainforest site Bellenden Ker, Queensland, Australia in 2002. This site showed a strong seasonality in N2O emissions as driven by a distinct wet and dry season [Kiese et al., 2003; Butterbach-Bahl et al., 2004]. As compared to the model version used by Kiese et al. [2005], the new parameterization performed better (this study: r2 eff = 0.45; RMSPEn, = 0.7; Kiese et al. [2005]: r2 eff < 0; RMSPEn, = 1.99) and the model did not tend to overestimate N2O emissions under wet season conditions as it was the case in the previous version. However, under prevailing dry season conditions, N2O emissions tended to be overestimated in the course of low-intensity rainfall periods (e.g., Bellenden Ker 08/02–10/02, Figure 4).

Figure 5.

Measured and simulated N2O emissions from the three calibration sites Bellenden Ker/Australia, Kakamega/Kenya, and Xishuangbanna/Southwest China (±standard deviation; daily mean N2O emission).

Table 3. Modeled and Measured N2O Emissions From the Calibration and Validation Field Sites and Model Performancea
Data SetNumberDuration of MeasurementbData ResolutionMeasured N2O emission [g N ha−1 day−1]Simulated N2O emission [g N ha−1 day−1]Model PerformanceReference
Mean ± SEMinimumMaximumMean ± SEMinimumMaximumRMSPEnr2r2 Efficiencylog r2 Efficiency
  • a

    Abbreviations are as follows: SE, standard error; d, daily N2O measurements; m, monthly N2O measurements; nm, no model performance given as climate data for 2003 were not available.

  • b

    Dates are given as mm/yyyy.

Calibration Data
Bellenden Ker, Australia101/2002–12/2002215/d3.8 ± 0.20.414.33.2 ± 0.10.912.90.700.480.450.46Kiese et al. [2003]
Kakamega, Kenya204/2004–06/200459/d9.9 ± 0.91.730.99.6 ± 0.83.623.70.470.780.770.82Werner et al. [2006]
Xishuangbanna, SW-China302/2005–05/200559/d1.4 ± 0.10.52.21.1 ± 0.10.52.00.760.700.410.53Werner et al. [2007]
 
Validation Data
Bellenden Ker, Australia411/2001–12/200159/d3.9 ± 0.21.79.33.7 ± 0.21.55.90.740.460.440.37Kiese et al. [2003]
Kauri Creek, Australia507/199712/d3.0 ± 0.71.28.72.4 ± 0.21.64.00.790.560.320.52Breuer et al. [2000]
Kauri Creek, Australia601/2000–02/200011/2000–12/200044/d8.6 ± 1.11.431.511.0 ± 0.65.423.00.840.390.280.24Kiese and Butterbach-Bahl [2002]
Massey Creek, Australia705/1997–06/199718/d4.1 ± 0.32.86.03.7 ± 0.13.04.50.960.200.020.03Breuer et al. [2000]
Massey Creek, Australia812/199814/d15.8 ± 2.84.231.516.2 ± 1.99.930.40.630.570.570.47Breuer et al. [2000]
Lake Eacham, Australia905/199716/d2.1 ± 0.30.84.93.0 ± 0.22.24.21.010.68<0<0Breuer et al. [2000]
 
Validation Data (Monthly Measurements/No Climate Station Data)
La Selva, Costa Rica1010/1990–09/199112/m16.0 ± 3.71.236.712.9 ± 2.52.925.50.700.530.460.78Keller and Reiners [1994]
Guacimo, Costa Rica1103/1992–11/19927/m12.4 ± 2.26.723.312.4 ± 2.63.421.20.820.460.220.10Liu et al. [2000]
Central Rondônia, Brazil1206/1992–12/199311/m4.9 ± 0.91.410.93.4 ± 0.51.76.70.780.620.330.26Melillo et al. [2001]
Jambi Province, Indonesia1309/1997–08/19989/m0.6 ± 0.20.22.01.1 ± 0.10.71.41.20.07<0<0Ishizuka et al. [2002]
Wuasa, Indonesia1411/2002–10/200312/m3.1 ± 0.60.88.65.9 ± 0.44.28.4nmnmnmnmPurbopuspito et al. [2006]
Pará, Brazil1502/1995–05/199615/m7.7 ± 1.71.519.97.1 ± 1.20.716.30.870.240.190.39Verchot et al. [1999]

[38] The test of model performance with other N2O emission data sets from tropical rainforest sites not included in the calibration procedure was partially hampered by low-frequency measurements (e.g., monthly measurement intervals), lack of information on sampling dates and/or meteorological conditions (Table 3; data sets 10–15). To overcome these uncertainties, we were forced to aggregate daily model outputs to monthly values (La Selva 1990/1991 [Keller and Reiners, 1994]) and to use climate data from the ECMWF climate data set for the respective region (grid cell) if no other data was available [Verchot et al., 1999; Melillo et al., 2001; Ishizuka et al., 2002; Purbopuspito et al., 2006]. However, it should be noted that the comparison of monthly N2O emission measurements with monthly aggregated model output is critical, as N2O emissions in tropical rainforest ecosystems are highly depending on rainfall events [e.g., Butterbach-Bahl et al., 2004]. Therefore a low model agreement with data sets obtained by weekly or monthly measurements is not necessarily a result of insufficient model performance. Despite the described difficulties measures for model performance for the 12 test data sets were in the same range than in comparison with the three calibration data sets with r2 up to 0.68, r2 eff up to 0.57 and RMSPEn of around 0.7 (Table 3).

[39] Whereas Table 3 gives site specific information of model performance, Figure 6 shows mean simulated versus mean measured N2O emissions for all simulated sites. It is apparent that simulated and measured emissions were close to the 1:1 line over a wide range of N2O emission levels (<1–16 g N ha−1 day−1). However, for two data sets larger deviation of measured and modeled N2O emissions are visible (data sets 10: La Selva 1990/1991; 14: Wuasa 2003; see Tables 2 and 3 and Figure 6). For the Wuasa site, Indonesia, we could not retrieve meteorological data for the year 2003. To overcome this problem we simulated N2O emissions for this site for the years 1994, 1997 and 1998, since the annual precipitation sum of those years was comparable with the long-term mean precipitation (Table 2). In view of the strong effect of precipitation on soil climate and thus on soil microbial processes, the 40% overestimation by the model is not a principle failure of the model to reproduce N2O emissions at this site but rather induced by precipitation differences. The underestimation (−20%) of N2O emissions for the La Selva site (data set 10 [Keller and Reiners, 1994]) could be due to the above-mentioned comparison of monthly measured fluxes (assuming that those values indeed reflect mean monthly fluxes) with daily simulated N2O emissions aggregated to monthly means (see also discussion by Kiese et al. [2005]).

Figure 6.

Comparison of measured and modeled N2O emission for the three calibration sites (circles) and the 12 testing sites (squares). Given values are mean daily N2O emissions (±standard error). Numbers correspond to the data set numbers of Tables 2 and 3.

[40] The overall r2 calculated for all 15 simulated sites was 0.94, thus providing a sound basis for the application of ForestDNDC-tropica within an upscaling approach in order to quantify N2O emissions from tropical rainforest on the global scale (section 3.3).

3.2. Model Sensitivity

[41] The sensitivity of simulated N2O emissions by the ForestDNDC model on variations in input parameters has been tested by Stange et al. [2000] on site scale and by Kiese et al. [2005] on site as well as on regional scales. Both studies indicated that SOC, texture and pH are the most sensitive factors finally affecting simulated N2O emissions. In this study the sensitivity of modeled N2O emissions on variations in input parameters was determined by randomly selecting 1000 grid cells out of the global GIS database and by varying the individual parameters by ±10, 25 and 50% (Figure 6). The most sensitive parameters were soil pH and soil bulk density (β > 1). The pronounced sensitivity of simulated N2O emissions to changes in pH is in agreement with earlier studies and is due to the explicit consideration of pH effects on N2O production by nitrification and denitrification [Li et al., 2000; Kesik et al., 2005]. This strong sensitivity of microbial N2O production even to small changes in pH has recently been demonstrated in studies involving pure cultures of heterotrophic nitrifiers [Kesik et al., 2006], but was also shown in earlier field and laboratory studies [Yamulki et al., 1997; Ellis et al., 1998; Stevens et al., 1998]. The sensitivity of simulated N2O emissions on changes in bulk density is a consequence of the implemented dependency of nitrification and denitrification on oxygen availability and, thus, on the extent of anaerobic zones in the soil profile [Li et al., 2000]. An increase of bulk density will decrease total pore volume and, thus, oxygen diffusion into the soil profile. This is notable following rainfall events, when in soils with high bulk density the anaerobic volume fraction of the soil increases more rapidly as compared to soils with lower bulk density [Stange et al., 2000; Kiese et al., 2005].

[42] Sensitivity of N2O emissions on leaf mass and floor mass changes was significantly lower (β = 0.39–0.6) as compared to changes in soil pH and bulk density. While the effect of floor mass is direct, by affecting the amount of substrates available for mineralization and consecutive nitrification and denitrification processes [Li et al., 2000], the effect of leaf mass is indirect, by its impact on the magnitude of litter fall and, thus, on forest floor mass. That field N2O emissions tend to increase with increasing availability of organic matter and soil organic carbon content has been shown in several field studies, as, for example, reviewed by Li et al. [2005].

[43] Most notable is the relatively low sensitivity of simulated N2O emissions with ForestDNDC-tropica to changes in SOC. In earlier studies with the ForestDNDC and DNDC models SOC was always one of the most sensitive parameters [Stange et al., 2000; Li et al., 2005]. However, this is an apparent insensitivity, which mainly depends on the rather low mineral soil SOC contents of the most prominent soil classes (area corrected average of 1.38%). Therefore even changes by ±50% (0.69–2.07%) have not necessarily significantly stimulated N2O emissions via increased microbial C and N turnover, since, in the frame of the sensitivity study, the absolute changes in total available labile C (N) fractions within the soil were rather low. Furthermore, as also found in field studies [e.g., Kiese and Butterbach-Bahl, 2002], the uppermost soil layers and the litter layer are the most important layers driving N2O emissions.

[44] The calculated sensitivity index for changes in precipitation was very low (see Figure 7). This is a consequence of the underlying controls on the microbial process involved with the production and consumption of N2O. As stated previously, precipitation is a factor modulating the N2O source strength defined by the soil and vegetation properties. Changes of precipitation can therefore have different effects on the level of emitted N2O at a given site within a single year, depending on the fact if nitrification or denitrification is dominating the N2O production or in case of denitrification even N2O consumption.

Figure 7.

Model sensitivity for changes of individual input parameters on N2O emissions. Details for the calculation of the sensitivity index (β) are given in the text. The index was calculated using three sets of parameter variations: strong (±50%), medium (±25%) and small (±10%) parameter variation. Temperature and pH were altered by ±5°C/±2.5°C/±1.0°C and ±1/±0.5/±0.25, respectively. The distance of β from zero is proportional to the sensitivity of a given parameter whereas the sign of β indicates the sign of correlation (asterisk: ±50% variation of bulk density results in unrealistic values, thus not calculated).

[45] In contrast to these findings, a positive β value was calculated for the effect of air temperature on N2O emissions (Figure 7). As temperature controls the process activity of decomposition, nitrification and denitrification in ForestDNDC-tropica, the sensitivity of the simulated N2O emissions on temperature is pronounced. However, an adverse effect can also be observed if elevated temperatures are leading to substantial reductions of soil moisture due to increased transpiration.

3.3. Global N2O Emissions From Tropical Rainforest Soils

[46] Following intensive model sensitivity tests on site scale, we finally applied ForestDNDC-tropica at a global scale to reassess N2O emissions from tropical rainforests worldwide. ForestDNDC-tropica was linked to the newly developed global GIS database holding the relevant information for initializing and driving the model (Figure 3). Figure 8 shows the spatial distribution of simulated mean annual N2O emissions from tropical rainforest soils for the years 1991–2000 (see also Figure 9 for the standard deviation of N2O emissions for a 10-year period). The map shows that N2O emissions ranged from <0.5 to 5.4 kg N ha−1 yr−1, which covers the reported range of annual N2O emissions from tropical rainforest ecosystems worldwide (see, e.g., overview given by Verchot et al. [1999], Breuer et al. [2000], and Werner et al. [2007]). In our combined model-GIS application highest N2O emissions of >3.5 kg N ha−1 yr−1 were simulated for the Southeast Asian tropics, central Africa, north and the southwest Amazon and Central America. Significantly lower emissions in a range of 0.5–1.0 kg N ha−1 yr−1 were simulated for large regions in South America, Africa and Oceania (Figure 8). Regional variations in N2O emissions were mainly resulting from differences in the spatial distribution of soil and climate characteristics. For instance, in Southeast Asia tropical forests are receiving highest amounts of annual rainfall and are growing on soils characterized by high clay and organic carbon contents and pH values >4.5 (Figure 2). Such conditions are favoring microbial N turnover processes and especially N2O production via denitrification [Granli and Bøckman, 1994]. The same finding holds true for elevated N2O emissions simulated in central Africa, and Central as well as South America (Figure 2).

Figure 8.

Ten-year (1991–2000) average of annual N2O emissions from tropical forest soils.

Figure 9.

Interannual variability of N2O emissions. Given is the standard deviation (SD) of annual N2O emissions for the simulated years 1991–2000.

[47] This spatial pattern can also be supported by annual N2O emission extrapolations from monthly field measurements. Annual N2O emissions of up to 6.1 kg N ha−1 yr−1 were reported for primary tropical rainforests of Costa Rica [Keller et al., 1993; Keller and Reiners, 1994]. Werner et al. [2007] also reported high N2O emissions for the onset of the wet season in a tropical rainforest in Kenya (annual estimate: 2.6 ± 1.2 kg N ha−1 yr−1). Annual N2O emission estimates for tropical rainforests of the Amazon region were reported to range between 1.4 and 2.6 kg N ha−1 yr−1 [Melillo et al., 2001]. N2O emissions of 1–3 kg N ha−1 yr−1 for rainforest in Australia [Kiese et al., 2003, 2005] and tropical southwest China [Werner et al., 2007] were reported.

[48] Simulated regional N2O emissions do not necessarily match measured N2O emissions on a site scale, as can be shown for Southeast Asia. For this region, Ishizuka et al. [2002] as well as Purbopuspito et al. [2006] reported annual N2O emissions <1.5 kg N ha−1 yr−1, whereas simulated N2O emissions for the respective grid cells yielded N2O emissions >1.5 kg N ha−1 yr−1 (Figure 8). Since a comparison on site scale returned comparable results between measured and simulated fluxes for the location (Figure 6), this difference is related to differences in soil properties as observed on site scale and used on the regional scale in our GIS. For the respective grid cells in Indonesia we found significantly higher soil clay and organic carbon contents as reported by the field studies [Ishizuka et al., 2002; Purbopuspito et al., 2006], which support the higher N2O emission in the grid-based model simulations.

[49] On average, the mean annual N2O emission source strength from rainforests worldwide was 1.2 ± 0.3 kg N ha−1 yr−1, corresponding to a total source strength of 1.34 Tg N yr−1 (assuming a total rainforest area of 10.9 × 106 km2, Table 4). Neglecting possible uncertainties in model parameters, this estimate still has a significant uncertainty induced by the uncertainty of soil and vegetation initialization. Figure 10 shows the frequency distribution of the 106 simulated N2O emissions of the LHS exercise. The graph shows that the distribution of simulated N2O emissions is skewed toward lower values, with a median of 0.8 kg N ha−1 yr−1. Using the distribution curve and assuming that this curve as derived from the 1000 grid cells is representative for the entire data set, we defined the input data induced uncertainty of our global estimate to be the percentage deviation of the first quartile (lower boundary) and the third quartile (upper boundary) as compared to the median value (see inset of Figure 10). This means that the lower boundary of the global N2O source strength of tropical rainforests equals:

equation image
Figure 10.

Frequency distribution of simulated N2O emissions by Latin hypercube sampling of 1000 randomly selected locations of the data set. The inset shows the progression of the median, first quartile (Q1), and third quartile of all obtained N2O fluxes with the increase of sampled grid cells. The quartiles are used to describe the lower and upper bounds of data-induced model uncertainty (lower boundary: −34.6%; upper boundary: +76.5%).

Table 4. Global N2O Emission From Tropical Rainforest Soils (Mean N2O Emission 1991–2000) Separated by the Contribution of Different Continentsa
 Area, 106 km2N2O Emission, kg N ha−1 yr−1N2O Source Strength, Tg yr−1
  • a

    Standard deviation expresses the intra-annual variability observed for the 10 simulation years.

South America6.0261.11 ± 0.260.671 ± 0.154
Africa3.0551.13 ± 0.280.344 ± 0.084
Asia1.4321.80 ± 0.440.258 ± 0.063
Central America0.3101.63 ± 0.350.051 ± 0.011
Oceania/Australia0.1041.09 ± 0.270.011 ± 0.003
Total10.9261.22 ± 0.291.335 ± 0.315

[50] The upper boundary is:

equation image

[51] In view of the rather conservative approach of boundary conditions we are convinced that the global source strength of tropical rainforest is indeed in the range of 0.88–2.37 TG N yr−1. Our global source strength of 1.34 Tg N yr−1, as well as the data induced uncertainty range, are in agreement with previous approaches applied for estimating the global N2O source strength of tropical rainforest ecosystems. Matson and Vitousek [1990] estimated the global N2O source strength of tropical forest soils to be 2.4 Tg N yr−1 (= 1.8 Tg N yr−1 using our area estimate for tropical rainforests). Their estimate was based on upscaling of a limited number of publications (<10) on N2O fluxes from tropical rainforest soils using a simplified soil and vegetation classification scheme. A few years later, considering newly published data, Breuer et al. [2000] provided a new estimate of 3.55 Tg N yr−1 (= 2.6 Tg N yr−1 with our area), which is noticeably higher than our mean estimate. However, these authors pointed out that such a simple upscaling approach may not be reliable for estimating the global source strength of tropical rainforests for atmospheric N2O. On the basis of a simple global model which reflects the spatial and temporal variability of the major controlling factors of N2O production in soils, Bouwman et al. [1995] reported the global N2O source strength of closed tropical forest ecosystems to be 2.3 Tg N yr−1 ± = 1.5 Tg N yr−1 for our area). Using a comprehensive statistical analysis of published measurement data, Stehfest and Bouwman [2006] recently estimated the N2O source strength of closed tropical forests to be 1.17 Tg N yr−1 (= 1.5 Tg for our area). A more mechanistic approach for estimating N2O emissions from rainforest soils was applied by Potter et al. [1996]. By linking the biogeochemical CASA model to a global GIS they calculated that N2O emissions from tropical rain forest ecosystems to be approximately 1.3 Tg N yr−1 (the area extend of rainforest was approximately the same as in our study), and thus only slightly lower than our figure. Contrasting results have been obtained by Melillo et al. [2001] in their study on N2O emissions from tropical forest ecosystems of the Amazon. They used the Terrestrial Ecosystem Model (TEM) for calculation of net nitrogen mineralization rates (Nmin). N2O emissions were than calculated using a loss ratio of 1.4% (value based on field studies) of calculated Nmin. For the period 1978–1995, Melillo et al. derived an estimate of 1.3 Tg N yr−1 for tropical forests in the Amazon basin (5.4 × 106 km2) which is significantly higher than our estimate of 0.8 Tg N yr−1 (mean for 1991–2000) for South America. This discrepancy may be attributed to three major differences.

[52] 1. There are different model approaches. Melillo et al. [2001] simulated N2O emissions as a static fraction of Nmin on monthly time step. In our study N2O emissions were mechanistically simulated in daily time steps as direct or indirect products of N-turnover via nitrification and denitrification.

[53] 2. There are differences in the GIS database on soil and vegetation properties. Melillo et al. [2001] used the general FAO-UNESCO database from 1971, whereas we used soil profile data from the ISRIC-WISE soil database [Batjes, 2002a] and linked this information to the global FAO soil map. Comparable differences also exist with regard to climate data (different years and different origin) and vegetation properties, for the latter one we used output of the LPJ-DGVM for initialization of the ForestDNDC-tropica model.

[54] 3. There is different spatial extent of tropical forests in both studies.

[55] Since Melillo et al. [2001] do not provide a detailed evaluation of their model against field data a final judgment on the model performance is not possible. Furthermore, in the framework of inventorying national N2O emissions from arable soils of the USA, Del Grosso et al. [2006] pointed out that daily simulation steps are a prerequisite for reliable annual estimates of N2O emissions from soils.

[56] Besides the spatial variability in N2O emissions, we could also demonstrate a pronounced interannual variability of N2O emissions across the simulated years (1991–2000). Highest interannual variability of N2O emissions as driven by the variability in rainfall was predominantly predicted for those regions having annual N2O emissions >1.5 kg N ha−1 yr−1. This is most evident for the eastern Congo and Papua New Guinea as well as for mountain rainforest areas in the Andes (Figures 8 and 9). In these areas, the interannual variability of N2O emissions given as standard deviation for the 10 years simulation period was as high as 2 kg N ha−1 yr−1, whereas in regions with lower mean annual N2O emissions the interannual variability is not exceeding 0.5 kg N ha−1 yr−1.

[57] In this study we present an average global N2O emission for a 10 years time span, thereby acknowledging significant interannual (e.g., Africa 1993: 0.21 Tg N yr−1 versus 1994: 0.42 Tg N yr−1) and seasonal variations as shown in Figure 11 on a continental scale. One can postulate that this variability of the N2O source strength should be visible in the atmospheric N2O signal. However, this will require a detailed comparison of our simulation with approaches using inverse modeling of global N2O emissions sources as recently published by Hirsch et al. [2006].

Figure 11.

Interannual variations of N2O emissions from tropical rainforest soils of South America (S-America), Africa, Asia, Central America (C-America), and Oceania/Australia for the years 1991–2000.

4. Conclusions

[58] This study provides a revised estimate of N2O emissions from tropical forest soils worldwide using the mechanistic biogeochemical ForestDNDC-tropica model. Our estimate of 1.34 Tg N yr−1 (0.88–2.37 Tg N yr−1) is at the lower end of previously reported estimates, but still shows that tropical rainforests are indeed one of the major sources of atmospheric N2O. For several reasons: (1) comprehensive recalibration and testing of the ForestDNDC-tropica, (2) development and use of a detailed global GIS database which considers the variability of soil, vegetation properties and regional climate, (3) assessment of input parameter uncertainty and model sensitivity and (4) applying multiyear simulations, we are convinced that our estimate provides the most profound estimate of the global source strength of tropical rainforest for atmospheric N2O. On the basis of daily modeling time steps and a spatial resolution of 0.25° × 0.25° our study shows a pronounced spatial and temporal variation of N2O emissions on a global scale. We hypothesize that this variability should be visible in the atmospheric N2O signal, thus offering a chance for independently validating our inventory, for example, by using inverse modeling techniques.

[59] However, our approach still has a high degree of uncertainty and a narrowing of the uncertainty range will strongly depend on (1) the availability of more detailed N2O emission measurement with corresponding information on soil and vegetation properties and weather records and (2) improved spatial information on soil and vegetation data.

Appendix A

[60] Table A1 provides an overview of the recalibrated model parameters prior to and after the calibration procedure, as well as their lower and upper bounds.

Table A1. Overview of the Recalibrated Model Parameters Prior to and After the Calibration Procedure and Their Lower and Upper Boundsa
ParameterIndexBeforeAfterMinimum BoundMaximum BoundDescription
  • a

    SP denotes scaling parameter.

Carbon Partitioning
Frac.HumusP010.90.9600.850.999fraction of passive humus of total carbon
Frac.R.LitterP020.70.7650.30.99fraction of resistant litter pool of total litter
Ratio.L2VL.LitterP032.02.0880.110.0ratio labile Litter: very labile litter
Frac.Bio.acriveP040.20.2060.10.3fraction of active humads
Frac.Bio.active.LP050.90.9180.50.99fraction of labile humads in active
 
Nitrification
KNITP061.00.4030.051.0SP gross nitrification rate
KNOP070.0050.0110.0010.1SP nitrification induced NO
KN2OP080.0010.000530.00050.01SP nitrification induced N2O
 
Denitrification
Threshold.DDRP090.60.6990.40.8threshold of drought induced activity reduction [WFPS, 0…1]
Min. DDRP100.150.1740.00010.5minimum of denitrification activity
Decr. DDRP110.80.7490.50.9daily reduction of denitrification activity
Incr. DDRP121.11.0881.051.5daily recovery of denitrification activity
NO3toNO2P130.10.1190.00010.2SP NO3 to NO2 conversion
N2OtoN2P140.20.1920.00010.4SP N2O to N2 conversion
 
Decomposition
DDRFP1520.013.6740.000130.0SP decomposition rate
Mois.DC1P160.5950.7380.30.9moisture control of decomposition (parameter 1)
Mois.DC2P178.09.9474.018.0moisture control of decomposition (parameter 2)
Temp.DC1P1835.1534.27030.040.0temperature control of decomposition (parameter 1)
Temp.DC2P195.395.3473.015.0temperature control of decomposition (parameter 2)

Acknowledgments

[61] We would like to thank Almut Arneth of Lund University, Sweden for making the LPJ-DGVM model available to us. This work was partly funded by the NitroEurope IP.

Ancillary