Marine vessel inventories demonstrate that ship emissions cannot be neglected in assessing environmental impacts of air pollution, although significant uncertainty in these inventories remains. We address this uncertainty by employing a bottom-up estimate of fuel consumption and vessel activity for internationally registered fleets, including cargo vessels, other commercial vessels, and military vessels. We identify model bias in previous work, which assumed internationally registered ships primarily consume international marine fuels. Updated results suggest fuel consumption is ∼289 million metric tons per year, more than twice the quantity reported as international fuel. According to our analysis, fuel used by internationally registered fleets is apparently allocated to both international and domestic fuel statistics; this implies either that ships operate along domestic routes much of the time or that marine fuel sales to these ships may be misassigned. If the former is true, then allocation of emissions to international shipping routes may underestimate near-coastal emissions from ships. Our updated inventories increases previous ship emissions inventories for all pollutants; for example, global NOx emissions (∼6.87 Tg N) are more than doubled. This work also produces detailed sensitivity analyses of inputs to these estimates, identifying uncertainty in vessel duty-cycle as critical to overall emissions estimates. We discuss implications for assessing ship emissions impacts.
If you can't find a tool you're looking for, please click the link at the top of the page to "Go to old article view". Alternatively, view our Knowledge Base articles for additional help. Your feedback is important to us, so please let us know if you have comments or ideas for improvement.
 Emissions inventories provide important information to atmospheric scientists, pollution modelers, and policy makers. They are a fundamental input to evaluate potential impacts of pollution on the environment and human health. Moreover, regulations to control pollution sources are directly aimed at reducing total emissions, typically on a source-by-source basis. In setting regulations, policy makers tend to focus either on sources causing greatest impact (i.e., the largest sources) or on the most cost-effective sources to control (i.e., the least regulated sources). Recently, non-road sources in general, and ship emissions in particular, have come under increasing domestic and international regulation because inventories have shown them to be both larger than previously considered and mostly unregulated [Environmental Protection Agency, 1998a, 1998b, 1999, 2002a, 2002b, 2003; European Commission, 2002a, 2002b; International Maritime Organization, 1998].
 Three general elements apply to all emissions inventories of combustion sources, whether stationary or mobile. First, an estimate or measure must be made of the combustion source activity level (e.g., power and/or fuel consumption). Second, emissions resulting from this activity must be computed. Third, these results must be assigned to a location so that air quality impacts (and sometimes jurisdictional authority) can be determined. Without an accurate representation of these elements, scientists cannot effectively inform policy decisions regarding human health and environment. All non-road mobile sources have been less well characterized than stationary or on-road mobile sources, primarily because of uncertainties in locations of activity and in activity levels. In international shipping, these uncertainties are compounded by limited information about the emissions factors assigned to marine engines.
2. Previous Ship Emissions Inventories
 Progress has been made in assigning international shipping traffic geographically. The first geographically resolved, global inventory of ship emissions derived international shipping traffic densities from voluntarily reported, ship-based weather observations [Corbett and Fischbeck, 1997; Corbett et al., 1999]. Other research assigned global emissions estimates to major shipping lanes [Lawrence and Crutzen, 1999; Olivier et al., 1996]. These approaches showed qualitatively good agreement with regionally resolved ship inventories [Carlton et al., 1995; Kesgin and Vardar, 2001; Streets et al., 1997, 2000], but resolution differences or inconsistent assumptions prevent reconciling geographic assignment of ship activity between spatial scales. Recent comparison of four approaches for assigning location to international ship activity on a global scale suggests that the Automated Mutual-assistance Vessel Rescue system (AMVER) may provide more representative international shipping distribution than other approaches [Endresen et al., 2003]. Endresen et al. claim that emissions can be allocated more accurately across ship traffic profile derived from the AMVER reporting frequency by using vessel size to weight the emissions allocation to each grid location. That paper presents a new and perhaps more representative methodology for assigning emissions from international marine bunkers to a global geographic grid, which is a significant contribution.
 Improvements also have occurred in test data for in-service marine engine exhaust emissions. These shipboard and manufacturer test data are used to establish average emission factors for various pollutants. In earlier work [Corbett and Fischbeck, 1997], average emission factors primarily relied on data from the Lloyd's Marine Exhaust Emissions Programme [Carlton et al., 1995]. Later research included additional in-service and manufacturer data and attempted to quantify the uncertainty in average emission factors [Skjølsvik et al., 2000]. Most recently, a study conducted by ENTEC and IVL for the European Commission updated these sources to compile and include the most comprehensive published test data for commercial marine vessels [European Commission and ENTEC UK Limited, 2002]. This report derived emission factors (NOx, SOx, HC, PM, and CO2) for five different engine types and three different fuel types. Fuel sulfur contents are based on data provided to the International Maritime Organization's Marine Environment Protection Committee [European Commission and ENTEC UK Limited, 2002; International Maritime Organization and Marine Environment Protection Committee, 2001]; ENTEC used average values of 2.7% for residual oil, 1.0% for marine distillate oil, and 0.5% for marine gas oil. Moreover, power-based emission factors and specific fuel consumption values were estimated for each engine-fuel combination using three different activities or operating modes of the ships: (1) “at sea” (or cruising); (2) “in port” (includes time spent hotelling, loading and unloading); and (3) “maneuvering.” ENTEC also developed a methodology to assign these factors to vessel type according to the weighted average of engine and fuel combinations within each vessel category. This represents the best current summary of emission factors available for large-scale emission inventories.
 An alternative “bottom-up” approach, often used at local and regional scales, estimates fuel consumption from engineering assumptions based on installed power, hours of operation, and operating profile. This has been the method of choice for port-based, local, and regional inventories [ARCADIS Geraghty & Miller, Inc., 1999a, 1999b; Booz Allen & Hamilton, 1991; Carlton et al., 1995, 1994; European Commission and ENTEC UK Limited, 2002; Lloyd's Register, 1993]. One major disconnect in ship emission inventories has been reconciling activity-based inventories with fuel-based inventories. Only recently has this disconnect been reconciled in regional inventories by relying on an explicit fuel-consumption model used for fuel-tax allocation purposes, based on specific vessel characteristics and transit information and validated with tax receipts on a national level [Corbett, 2002]. However, reconciling these two approaches is toughest at the global level.
 All large-scale inventories that included shipping emissions have relied upon the accuracy of international marine fuel statistics (see Table 1). Recent work by Endresen et al.  comes closer to a “bottom-up” approach by applying a statistical fuel consumption model to estimate world fleet fuel consumption and emissions. By using statistical correlations between deadweight tonnage and power, they indirectly estimate power from ship registry statistics and then estimate fuel consumption. Their model results are segregated into international cargo ship activity, domestic cargo ship activity, and non-cargo ship activity for all internationally registered ships above 100 gross registered tons. However, even that work applies assumptions designed to achieve agreement with international marine fuel statistics.
Table 1. Summary of Emissions Estimates and Related Fuel Consumption Dataa
Fuel Consumption, Million Tonnes
NOx, Tg N
SOx, Tg S
Inventory Years Represented
Where previous studies presented multiple years or results from separate methodologies, we report an average of their results for simplicity. While fuel consumption and related emissions are expected to vary from year to year (particularly within a given study's set of assumptions), the differences among these various estimates relates more to input assumptions and data than to interannual variation.
To convert from CO2 data reported by EDGAR to fuel consumption, the CO2 emissions were converted to elemental carbon and divided by the average percent by weight of carbon in marine fuel (86%).
 One problem is that not all statistical sources have defined international marine fuels the same way [Olivier and Peters, 1999]. Original International Energy Agency (IEA) definitions [International Energy Agency, 1987] have been reworded to be more consistent with reporting guidance under IPCC [Houghton et al., 1997]. The IEA defines “international marine bunkers (fuel) [to] cover those quantities delivered to sea-going ships of all flags, including warships. Consumption by ships engaged in transport in inland and coastal waters is not included.” The IEA defines national navigation to be “internal and coastal navigation (including small craft and coastal vessels not purchasing their bunker requirements under international marine bunker contracts). Fuel used for ocean, coastal and inland fishing should be included in agriculture.”
 We update global fuel estimates for internationally registered ships (i.e., greater than 100 gross registered tons) by adopting a more rigorous methodology for estimating energy consumption that is independent of fuel sales statistics. Rather than estimating ship engine sizes based on vessel tonnage, we directly obtain engine power and apply vessel activity data to compute fuel consumption. Current emissions rates are then applied to this “bottom-up” fuel-use model to estimate emissions from internationally registered ships. Detailed discussion of methodology and assumptions follow, but our calculations can be summarized in the following equation.
where PMW is accumulated installed engine power for each subgroup, F% MCR is engine load factor based on duty cycle profile, thrs/yr is average engine running hours for each subgroup, SFOCg/kWh is the power-based specific fuel oil consumption, and Eg/kWh is the power-based emissions factor for each pollutant. Figure 1 presents a general schematic of the updated fuel-estimation methodology. (See Corbett et al.  for illustration of the previous model.) We estimated vessel activity, fuel consumption, and emissions using the following general methodology and assumptions.
3.1. Identify the Number of Registered Ships and Engines in Service
 We consider first the structure of the world's merchant fleet, using current ship registry data [Lloyd's Maritime Information System (LMIS), 2002] for oceangoing ships greater than 100 gross tonnes (GT). Engine-specific data were supplemented by industry statistics provided by MAN B&W (accessible by one of the authors). We identified a total of 88,660 ships equipped with 116,280 main engines totaling some 280,100 MW (323,000 MW including auxiliary engines). Table 2 classifies these ships by general type and presents main engine and power totals.
Table 2. Profile of World Fleet, Number of Main Engines, and Main Engine Powera
Number of Ships
Percent of Fleet
Number of Main Engines
Percent of Main Engines
Installed Power (MW)
Percent of Total Power
Percent of Energy Demand
The world fleet represents internationally registered vessels greater than 100 gross tons; the cargo fleet represents those vessels whose main purpose is transporting cargo for trade. Percent of energy demand mainly adjusts for reduced activity (in loads and hours) by military vessels under typical operations.
General cargo vessels
Other (research, supply)
Registered fleet total
World fleet total
 About 67% of these ships are powered by four-stroke compression-ignition engines (operating on the compression-ignition, or diesel cycle, and therefore referred to as diesel engines). Some 26% are powered by two-stroke diesel engines. Six percent of the ships have “unknown” diesel engines (i.e., either two- or four-stroke) and only one percent are turbine-driven. Most turbine-driven vessels (80%) are steam turbines with oil-fired boilers; the number of aero-derivative gas turbine engines in the commercial fleet is very low.
 When estimating fuel consumption (and related emissions) from ships, the installed engine power rather than the number of engines or vessels is of major importance. The fleet has approximately 84,000 four-stroke engines with total installed power of 109,000 MW and some 27,000 two-stroke engines with total installed power of 164,000 MW. Engines with “unknown” cycle types and “turbines” together make up only about 2.5% of total installed power for main engines. This suggests that 27,000 two-stroke marine prime movers account for almost 60% of the fleet's total energy output and fuel consumption. The majority of these engines are large-bore, low-speed diesel engines above 10 MW rated output. Two-stroke engines are the main consumers of bunker fuel and therefore the major sources of oceangoing ship emissions, followed by four-stroke engines.
 Approximately 50% of the power produced by marine two-stroke diesel engines comes from some 9800 low-speed engines manufactured by MAN B&W. Data used in this analysis were cross-checked with technical data for engines manufactured by the other major manufacturers. This enables us to ensure that technical engine data used for this study and the assumptions applied for this analysis closely reflect actual industry conditions.
 On the basis of prior work [Adcock and Stitt, 1995; Corbett and Fischbeck, 1997], we include the nearly 20,000 military vessels worldwide, acknowledging that these engines have very different activity profiles. The naval ship is designed for sustained speeds in excess of their endurance (cruise) speed [Markle and Brown, 1996]; as a result, the installed power in the world Navy is nearly 40% of the total installed power of the entire commercial fleet. However, the world's Navy ships use less energy than their installed power would suggest, because most of the time they operate in peaceful conditions at endurance speeds, partly to achieve fuel/cost savings. Average engine power characteristics for the U.S. Navy show that military ships operate below 50% power for 90% of the time that they are underway [Naval Sea Systems Command (NAVSEA), 1994]. Moreover, military vessels spend more time in port than modern commercial ships. Where commercial ships cannot effectively earn profits unless underway, Navy studies claim that military ships typically spend as much as 60% to 70% of their time in port [NAVSEA, 1992]; this is confirmed by public records of Naval ship activity, which showed that typical deployment rates range between 40% and 55% of the fleet. This means that in practice, military ships would demand some 14% of total energy required by the World Fleet (commercial plus military fleets), much less than their installed power implies (see Table 2).
 Our best estimate of fuel consumption used a more detailed version of this general methodology, based on industry vessel and engine data typically not available in the ship registries. For example, we subdivided each of the main engine groups to define five main groups of propulsion engines onboard vessels: (1) two-stroke low-speed engines; (2) four-stroke medium-speed engines; (3) four-stroke high-speed engines; (4) turbines; and (5) others. Each main group was also subcategorized as: (1) large-bore engines (high engine power); (2) medium-bore engines (medium engine power); (3) small-bore engines (lower engine power); and (4) turbines (steam and gas). This resulted in a total of 132 subgroups (with an average of only 900 engines in each). Auxiliary engines were treated as their own subgroup. We adjusted our activity estimates to include only engines on active vessels (e.g., consuming fuel); on average only about 1% of all ships are laid-up or under repair [LMIS, 2002].
3.2. Estimate the Number of Engine Service Hours
 For typical engine models within these subgroups, we surveyed ship operators and naval engineers, project engineers, and marine engines sales personnel for average yearly running hours, annual average engine loads, and fuel type. Table 3 summarizes the information obtained from a review of manufacturer engine files and consultation with operators of larger vessels (typical of most of the installed power on ships). On the basis of operator and manufacturer survey data, marine engines operate about 6500 hours per year on average, or about 74% of the year.
Table 3. Summary of Engine Profiles From Manufacturer and Operator Survey
3.3. Determine Engine Load Profiles, Including Full-Cruise Power and Duty Cycle
 We assume that typical maximum power in service is 80% of rated engine power, per commonly accepted guidelines for heavy-duty engines available in various engine specifications. Most studies have applied the International Organization for Standardization (ISO) standard duty cycles for marine engines [ISO, 1996]; however, our survey data suggest that, on average, main engines may not operate according to these default profiles. We apply survey data and our own industry experience to establish an appropriate duty cycle for oceangoing transport ships (cargo vessels and passenger ships), and we apply the standard E3 duty cycle for oceangoing non-transport ships (fishing and factory vessels, research and supply ships, tugboats). For auxiliary engines, an estimated 50% average load and 3500 operating hours per year was used. (Fuel estimates are very dependent upon duty cycle assumptions, as our later discussion of uncertainty shows.)
3.4. Apply Average Fuel Consumption Rates for Each Engine-Fuel Combination
 For main engines, we matched engine types in each subgroup with the fuels consumed by these engines [ISO, 1986, 1987]. Ninety-five percent of two-stroke, low-speed engines use heavy fuel oil (HFO or residual fuel), and 5% are powered by marine distillate oil (MDO). Fuel consumed by 70% of the four-stroke, medium-speed engines is HFO, with the remainder burning either MDO or marine gas oil (MGO). Four-stroke, high-speed engines all operate on MDO or MGO. We assume that the unknown diesel engine types are small, high-speed engines all operating on MDO or MGO, steam turbines all have boilers fueled by HFO, and gas turbines are powered by MGO.
 The average fuel consumption is a composite of the fuel-usage rates at various engine loads (dependent on duty cycle). Marine engine data on actual fuel consumption rates were used to establish a typical average fuel consumption rate, applied to the average engine load across the duty cycle. In general, cargo ships have more fuel-efficient, larger engines than non-transport ships. Typical fleet-average fuel consumption rates were 206 g/kWh for transport ships and 221 g/kWh for non-transport ships, although this is a simplified summary of our subgroupings for the best estimate. There can be transport ships with specific fuel consumption rates of >221 g/kWh and non-cargo ships with consumption rates of <206 g/kWh; more decisive criteria are the size of the engines, the fuel they burn, and the number of years they have been in operation (wear conditions). By using actual fuel consumption data, we include these factors.
3.5. Calculate the Annual Fuel Consumption for All Engine Sub-Groups
 Emission factors depend upon many parameters, but engine make and type, size, speed, and load are of paramount importance. Other factors are type and specification of the fuel, engine design (in-line or V-type), mode of operation (constant speed or propeller law), and number of cylinders. In our own first calculations, we used measured emissions and measured fuel consumption rates from more than 50 curves for different engine types to validate our methodology.
 To illustrate how emissions factors may vary by engine type and load, we plotted a selection of four typical emission factor curves for (state-of-the-art) NOx-optimized engines [Koehler, 2002a, 2002b, 2003]. These engines included: (1) a large (1-m bore) two-stroke, low-speed diesel engine rated at 5700 kW/cylinder; (2) a large (0.58-m bore) four-stroke, medium-speed diesel engine rated at 1300 kW/cylinder; (3) a smaller (0.32-m bore) medium-speed diesel generator set rated at 480 kW/cylinder; and (4) a typical (0.185-m bore) high-speed diesel engine rated at 192 kW/cylinder. The first two engines represent typical main engine ratings and the last two represent typical auxiliary engine ratings. NOx emission factors are plotted by load in Figure 2. (Note that existing even larger low-speed, two stroke diesel engines can have emission factors in excess of 110 g NOx/kg fuel. These curves are only typical examples.) For the most interesting load points between 50% and 100%, there is not much variation of the emission profile for each given engine type. This limits the NOx calculation error, should the real average load for a certain group or sub-group of engines deviate from the average value. At 75% load, all the low-speed two-stroke engines for which measured data were reviewed have emissions factors ranging from about 80 to 110 kg NOx/tonne fuel, similar to all previous studies. When the composite average emission factor is calculated by using the duty cycle weighting, these fuel-based results are very similar to recent analyses by ENTEC, which reported both fuel-based and power-based emissions factors.
 Some literature suggests that fuel-based factors are inferior to power-based factors because fuel-based factors do not show the same increase in emissions factors at low loads [Energy and Environmental Analysis, Inc., and Sierra Research, 2000]. The ENTEC report clearly reconciles these differences by stating “emission factors in kg/tonne fuel can be obtained by taking the g/kWh factor and dividing by the specific fuel consumption.” When this fuel-based conversion is done, the relatively higher fuel consumption at low loads and the relatively higher emission rates (in g/kWh) at low loads tend to cancel out; the result is that fuel-based emission factors are more flat over the load range. Overall, the inventory results are nearly identical using either approach, if the fuel consumption is directly based on engine activity, which our bottom-up analysis is.
 Starting with the average power per engine, based on the duty cycle weightings in Step 3 above, our methodology is similar to the ENTEC approach [European Commission and ENTEC UK Limited, 2002]. These emission factors represent in-service engines in the current fleet. We estimate the weighted average emissions for each vessel type, based on the engine-fuel combinations for each type. For the purposes of this study, we focused on the “at-sea emissions factors” presented by ENTEC, since most of the fleet's main engine power is at full-cruise or slow-cruise loads. We apply ENTEC assumptions for average fuel-sulfur levels, discussed above. For auxiliaries, we also used ENTEC's approach for emissions factors. Fleet-average emission factors are summarized in Table 4.
Table 4. Fleet-Average Summary of In-Service Emissions Factors (g/kWh)
On the basis of an average fuel-sulfur content of 2.5% for heavy fuel oil.
PM factors for marine diesels remain very uncertain, are still difficult to measure, and can be defined differently. We apply the latest published average values for in-service marine engines [European Commission and ENTEC UK Limited, 2002].
3.7. Estimate Emissions Using Power-Based Emissions Factors for Engine Sub-Groups
 Aggregate these calculations by subgroup and vessel type.
 We estimate that world fleet fuel consumption, calculated for all main and auxiliary marine engines in the internationally registered oceangoing fleet (including military vessels), is ∼289 million metric tonnes annually. Heavy fuel oil (HFO) represents nearly 80% of the fuel consumed by these engines. This is considerably greater than (more than double) the amount of fuel (and corresponding CO2) reported for international marine bunkers. Furthermore, the annual emissions from ships (particularly NOx, SOx, and CO2) are significantly greater than previously considered (in our earlier work and in others' published estimates; see Table 1). NOx emissions are 6.87 Tg N per year, also more than twice as large as earlier estimates. Sulfur emissions are 6.49 Tg S annually, approximately 53% more than previously estimated. (Closer agreement for sulfur results from better data about actual fuel-sulfur levels in the fleet.) CO2 emissions are 249 Tg C per year. Particulate matter (PMtotal) and hydrocarbon (HC) are estimated to be 1.64 Tg PM10 and 0.769 Tg HC, respectively.
 These results raise at least one important question: Why is there such a large discrepancy between fuel statistics and actual fuel usage by internationally registered ships? The question can be posed in three parts.
4.1. Are the Previous Fuel-Based Models Systematically Biased by Assuming That Internationally Registered Ships Would Exclusively Use Fuel Identified as Sold for International Vessel Activity?
 This work and the recent study by Endresen et al.  both support the possibility that international fuel statistics do not describe the total fuel consumed by the world fleet of ships. Additional evidence is provided by Olivier and Peters . Research has shown that there often can be a tendency to underestimate systematic error in models and experiments [Morgan and Henrion, 1990].
4.2. Could the Bottom-Up Model Presented Here be in Error?
 Our model is essentially the same bottom-up approach that fleet operators and designers would use to evaluate fuel consumption. It is also very similar to the methodologies used in local and regional non-road and on-road inventories. However, applying this methodology globally requires a number of inputs be applied to many similar vessels and engines, or that very high resolution be used in classifying vessel types and engine characteristics into subgroups. In our best estimate, we use 132 subgroups. Scaling-up of the methodology for fuel estimation may itself insert bias or error. To evaluate the potential for our model to significantly overestimate fleet fuel use, we performed an analysis of the sensitivity of our model to uncertain inputs and assumptions.
 For the sensitivity analysis, we simplified the subgroups primarily by vessel type and defined the uncertain inputs to include the ranges of characteristics that we aggregated into these broader subgroups (Table 5). We also applied only two duty cycles to the commercial fleet, for simplicity and because they are in close agreement with the fleet operator responses used in our best estimate. These duty cycle profiles (Tables 6 and 7) are applied to the adjusted full-cruise power of 80% full load, and speed relationships follow the propeller law. We varied 10 general inputs (shown in Table 3) between lower, best estimate, and upper bounds using triangle distributions. Use of triangle distributions is an accepted way to perform analyses of uncertainty and sensitivity where the actual shape of the probabilistic distribution is not known [Morgan and Henrion, 1990]. Using Monte-Carlo simulation, we evaluated the impact of simultaneously varying each parameter according to the triangle distributions on the estimates for fuel consumption and emissions for each pollutant. Figure 3 presents the output sensitivity to uncertain average input variables for fuel consumption of the World Fleet and Cargo Fleet. Also shown is a summary of international fuel statistics reported in the World Energy Database (available from Energy Information Administration, World Energy Database and International Energy Annual 2001, available at http://www.eia.doe.gov/emeu/world/main1.html), with error bars representing differences between various published sources for international marine fuel statistics [Skjølsvik et al., 2000].
Table 5. Variable Ranges for Uncertainty Analysis
Auxiliary engines were all assumed to have a composite duty cycle of 50%, with sensitivity to auxiliary fuel estimate evaluated independent of duty cycle. Auxiliary fuel estimate ∼15 million metric tonnes fuel per year.
T, transport ship; NT, non-transport ship. Duty cycle changes were defined with correlations to “full cruise” power, because the total weighting cannot exceed 1.00.
SOx emissions from diesel engines are strictly related to the sulphur content of the fuel. These factors are based on the weighted-average fuel sulfur levels, per the discussion. On average, they represent an average fuel-sulfur content of 2.5%.
PM factors for marine diesels remain very uncertain, are still difficult to measure and can be defined differently. We apply the latest published average values for in-service marine engines.
Table 6. Representative Duty Cycle Applied to Transport Ships Sensitivity Analysis
93% of rated
85% of rated
46% of rated
23% of rated
80% of rated
61% of rated
10% of rated
1% of rated
Table 7. Duty Cycle (ISO E3) Applied to Non-Transport Ships Sensitivity Analysis
93% of rated
85% of rated
46% of rated
23% of rated
80% of rated
61% of rated
10% of rated
1% of rated
 Note that our best estimates are the result of setting input parameters deterministically to values we believe may be most representative; these produce an estimate that does not cross the CDFs at the fiftieth percentile. This is partly because our effort to characterize uncertainty attempted to include reasonable input ranges that might best agree with published statistics. For example, we would need to reduce engine activity hours by some 20% to 63% to get agreement between our fuel estimates and previous research, depending on which estimates we try to match and whether we include only cargo ships, all commercial ships, or all international ships including military vessels. It appears unlikely that published fuel statistics could provide enough fuel for the world fleet, and only if vessel activity assumptions are significantly reduced.
Figure 4 presents uncertain input factors that contribute the greatest uncertainty in the overall estimate for fuel consumption in the World Fleet. These inputs necessarily all relate to the methodology by which we estimate engine activity, power, and specific fuel consumption. The most important uncertain inputs relate to vessel duty cycle and hours of operation, which are data that may be obtained directly from ship operators. Currently, regional inventories of marine activity in metropolitan port cities are focusing on the quality and resolution of these data. This should enable better accuracy of large-scale inventories over time.
Figure 5 presents the output sensitivity distribution for NOx emissions from the World Fleet and the Cargo Fleet, along with the previously reported uncertainty analysis [Corbett and Fischbeck, 1997; Corbett et al., 1999]. Uncertain energy intensity in the fleet dominates the uncertainty in pollutant inventories. Figure 6 presents uncertain input factors that contribute the greatest uncertainty in the overall estimate for NOx emissions in the World Fleet. All uncertain inputs except the NOx emissions factor relate to the model's estimates of engine activity for commercial and military vessels. In other words, uncertainty in the NOx emission factor is less important to our estimate than uncertainty in the vessel activity. Similar results apply to other pollutants; in fact, only the poorly known PM factor ranks higher than duty cycle in contribution of uncertainty.
 The fifth and ninety-fifth percentile values for annual NOx estimates in the World Fleet using our updated methodology are shown in Table 8. It is worth noting that even for the subset of the world fleet most likely to have been included in fuel-based model assumptions, the prior best estimate of 3.08 Tg N per year is lower than the updated fifth percentile estimate using engine activity to estimate fuel and emissions.
Table 8. Comparisons of Updated Emissions (Including 5th and 95th Percentiles) With Previous Best Estimates
 Similarly, we performed an analysis of uncertainty for other pollutants (Table 8). The updated estimates for SOx emitted by the World Fleet are only 53% higher than prior estimates, and for the Cargo Fleet the estimate for SOx emissions is within 12% of previous estimates. Of course, since SOx emissions are directly dependent upon fuel sulfur, other assumptions about fuel sulfur content would significantly modify these bounds. As with sulfur emissions, improved data about PM from large marine engines could modify PM upper and lower bounds. Moreover, this analysis adopts the uncertainty ranges on emission factors reported by ENTEC, which may be based on more complete test data for pollutants other than PM.
Figure 7 presents uncertain input factors that contribute the greatest uncertainty in the overall estimate for SOx emissions in the World Fleet. Figure 8 presents uncertain input factors that contribute the greatest uncertainty in the overall estimate for PM emissions in the World Fleet. Because of the relatively good fleet statistics on fuel sulfur, the fuel sulfur factor (a direct function of fuel sulfur) is less important than uncertain inputs for vessel activity. However, emission factors for PM remain the most important uncertain input.
4.3. If the Current Model is More Accurate in Describing the Fuel Used Annually by International Vessels, Can World Fuel Inventories Account for the Discrepancy Between International Marine Fuel Statistics and Actual Ship Fuel Consumption?
 Clearly, internationally registered vessels must be consuming domestic supplies of marine fuels. However, fuel statistics are generally well understood, especially for the OECD nations that sell most of the world's marine fuels (World Energy Database 2001). We might suggest that international marine fuels are only used by cargo transport vessels, and that non-transport vessels are generally operating domestically. With this correction, fuel identified as international marine bunkers would supply only 49% of the ships in the fleet, and only 46% of marine engines on cargo vessels.
 Unfortunately, reconciling fleet fuel consumption between international and domestic fuel supplies is not simple. To get agreement with international fuel statistics, we must assume that about 31% of the fuel consumed by transport vessels occurs while operating in short sea or coastwise service using domestic fuels and that all non-transport ships (fishing, research, tugs, etc) operate on fuel included in domestic inventories.
 To test whether this could be reasonable, we compared our marine fuel estimates with domestic residual fuel consumption and international residual bunkers statistics (World Energy Database 2001). Evidence that some of the 600+ million metric tons of fuel used domestically could be used by the ships in our “bottom-up” inventory would suggest at least that the World Energy Database is consistent with our results. We reconstructed a time series of fuel usage by oceangoing ships using historical records for installed power and a simplified inventory calculation for past years. We then compared our fuel consumption estimate for marine heavy fuel oil with published consumption statistics for residual fuel. Only about 13% of world domestic residual consumption would be necessary to account for the discrepancy between international fuel statistics and our estimates (Figure 9). Considering strictly the OECD nations, between 28% and 29% of domestic residual fuel would account for this difference.
 This comparison is suggestive but not conclusive without knowing how much residual fuel was actually sold to stationary versus mobile domestic sources. We turned to the Combined State Energy Data System (CSED) statistics for the United States [Energy Information Administration, 1999], the world's largest provider of marine fuels over the past ten years at least – accounting for some 21% of marine residual sold by OECD nations (World Energy Database 2001). The Energy Information Administration publishes state-by-state consumption statistics by fuel type and by end-use sector. We identified residual fuel consumption in the transportation sector in the United States domestic inventory as marine residual fuels, primarily because no other mobile source category consumes residual fuel. We compared states that recorded the largest consumption of residual fuels by ships with the top states for waterborne commerce [Corbett and Fischbeck, 2000; U.S. Army Corps of Engineers, 2001].
 Results (included in Figure 9) provide four important insights. First, total residual fuel consumption reported in the CSEDS agrees exactly with domestic residual consumption statistics in the World Energy Database, providing a consistency check. Second, the 10-year average residual consumption by marine transportation in the United States equals more than 40% of total residual consumption, more than enough to accommodate the difference between our estimates and international residual fuel statistics. Third, the top states for waterborne commerce (measured by total cargo volume) are also the top states for marine fuel sales. This suggests that most of domestic marine fuel consumption is going to cargo transport ships, and to vessels that directly support commerce. Finally, while domestic land-based consumption of residual fuel has decreased significantly over the past 2 decades, marine transportation fuel consumption has not changed much, either internationally or domestically (Figure 9).
5. Scientific and Policy Implications
 Scientists need an accurate and complete inventory, not one limited by the reported international fuel sales that ignore “domestic fuel” usage by ocean shipping. Our inventory may provide better estimates than previous work, but it cannot directly suggest how large ship emissions impacts may be at the local and regional levels. Clearly, more research is needed to assign geographically the “domestic emissions” from internationally registered ships, perhaps by reconciling inventories conducted at different geographic scales.
 Our results and those of Endresen et al.  suggest the need to revisit the geographic distribution of ships. Endresen et al. consider AMVER as potentially better than COADS, but they report important bias toward larger cargo vessels less likely to be operating within domestic routes. A profile of oceangoing ship traffic and emissions using AMVER's distribution would be inconsistent with our results that suggest more ship emissions may be occurring closer to land than currently characterized. Perhaps only emissions resulting from the use of international marine fuels (as defined) should be allocated to AMVER-derived (or COADS-derived) vessel traffic profiles. In any case, a methodology to conduct and present analysis of uncertainty in geographic representation using proxy data sets is needed. Recent studies have suggested this need in other geographically resolved global inventories [Olivier et al., 1999].
 Our results also suggest that recent research to account for in-plume decay and post-exhaust chemistry in models will need to continue to evaluate the potential impact from ship emissions. A number of studies have attempted to reconcile the apparent discrepancies between predicted pollution concentrations using large-scale chemical transport models and field observations [Corbett et al., 2002; Davis et al., 2001; Ferek et al., 1998; Kasibhatla et al., 2000; Russell et al., 1999]. If our larger inventory of ship emissions is more accurate, then why do model predictions using smaller inventory estimates over predict observations of ambient pollutant concentrations? This research area may be more important in light of these results.
 Finally, we suggest several policy implications. Ship emissions are larger than previously considered, and therefore continue to merit policy attention. Oceangoing ships produce about twice as much NOx as previously estimated, and a significant fraction of these emissions may fall under domestic accounting (if not jurisdiction). Moreover, given that nearshore emissions from ships may be much larger than depicted in global inventories, regional and local policy jurisdictions may have additional reasons to consider stronger action than the global standard set by the International Maritime Organization (IMO). International treaty and domestic regulatory development will need to consider environmental implications of both domestically assigned and internationally assigned energy use by ships. This will make policy efforts both more important and more complicated.