1  AIS-based emissions model

1.1 Methods

1.1.1 Model Overview

Several alternative models were tested - each using an engineering bottom-up approach, based on AIS data, vessel characteristic information, and various emission conversion factors from the literature. Here we describe the best and most comprehensive model, which closely follows the methodology described in the 2020 IMO “Fourth Greenhouse Gas Study” (Faber and Xing (2020)) and the 2017 ICCT “Greenhouse Gas Emissions From Global Shipping” study (Olmer et al. (2017)), and how we apply it to the GFW data, and any deviations made from the published methedologies. The validation section below describes several alternative model specifications.

From a high-level, we calculate emissions as follows:

  1. For each individual AIS message (ping), we calculate the main engine use, auxiliary engine use, and boiler use, each of which is a function of vessel characteristics, speed, and the time since the previous ping.

  2. Using emissions factors (EFs) for the main, auxiliary, and boiler engines for seven pollutants (CO2, CH4, NOX, SOX, CO, N2O, and PM), we calculate the emissions of each pollutant for each individual AIS ping for both the main, auxiliary, and boiler engines.

  3. For each pollutant and each AIS ping, we sum the emissions across the main, auxiliary, and boiler engines to get the ping-level emissions for each of these seven pollutants.

  4. For three additional pollutants (PM2.5, PM10, and VOCs), we multiply the CO2 emissions by a conversion factor to get the ping-level emissions for each of these three pollutants.

  5. With ping-level emissions, we are then able to aggregate emissions by vessel, by voyage, by port stay, by time, by space, etc.

1.1.1.1 Main engine

1.1.1.1.1 Main engine energy use

Based on Faber and Xing (2020) (page 64), main engine energy use (in killowatt-hours) is calculated as follows:

\[ \text{Main Engine Energy Use}_{kWh} = \text{Hours} \times \text{Load Factor} \times \text{Main Engine Power}_{kW} \] Where hours comes from each individual AIS message, main engine power (kW) comes from the vessel characteristics dataset, and the load factor comes from the product of several correction factors (CFs):

\[ \text{Load Factor} = \text{Speed-power CF} \times \text{Hull Fouling CF} \times \text{Weather CF} \times \text{Draft CF} \] Where: - hull fouling CF is 1.07, reflecting a 7% increase in resistance as described in Olmer et al. (2017) and Faber and Xing (2020) (see page 17 and Annexes page 270, respectively) - weather CF is a correction factor based on weather conditions, varying with the distance to shore. This factor is set at 1.1 for nearshore activity (≤5 nm from shore) to account for a 10% increase in resistance, and 1.15 for offshore activity (>5 nm), reflecting a 15% increase in resistance, as described in the Olmer et al. (2017) and Faber and Xing (2020) (see pages 18 and 270, respectively). - draft CF is extracted from the average draught by sector as reported in Olmer et al. (2017) (see Table 13 on page 20). Weights were applied by vessel type based on fuel consumption data from the Faber and Xing (2020) (see Annex 1, Figure 4), since fuel consumption values by type are proportionally related to emissions. This weighted average provided a final estimate of 0.85. Note: this factor could be refined using vessel class-specific averages. - speed-power CF is defined as \((\text{speed}_{knots} / {\text{design\_speed}_{knots}}) ^3\), with the additional stipulation that this ratio should not exceed 1:

\[ \text{Speed-power CF} = \begin{cases} 1 & \text{if } \frac{\text{speed}_{knots}}{\text{design\_speed}_{knots}} > 1, \\ \frac{\text{speed}_{knots}}{\text{design\_speed}_{knots}} & \text{otherwise} \end{cases} \]

In this equation, speed is derived from AIS-broadcasted speed measurements when there has been <= 1 hour since the previous AIS message (we therefore assume that all activity within the past hour is traveling at a similar speed). When it has been more than 1 hour since the previous message, the implied speed is used as a more accurate measurement of speed for that time period. This is calculated as distance from last position divided by hours since last position. Design speed for each vessels was estimated using a random forest regression trained using known registry design speed for a subset of vessels alongside other vessel characteristics including main engine power (\(ME\), kW) and gross tonnage (\(GT\)). This approach was applied to the entire GFW dataset of vessels.

\[ \text{design\_speed}_{knots} = (3.390 \times 10^{-4}) \cdot ME + (2.151 \times 10^{-5}) \cdot GT - (2.742 \times 10^{-9}) \cdot ME \cdot GT + 12.93 \]

As a last step, we ensure that the final load factor (the product of the above correction factors) does not exceed a value of 0.98, as recommended by Faber and Xing (2020) (see page 272).

\[ \text{Load Factor} = \begin{cases} 0.98 & \text{if (Load Factor)} > 0.98, \\ \text{Load Factor} & \text{otherwise} \end{cases} \] ###### Adjustments for fishing vessels

For fishing vessels, we adjust the main engine load factor based on the relationships published by Coello et al. (2015) (and later used by Sala et al. (2018)).

  • For fishing vessels of class trawlers and dredge_fishing, when they are actively fishing we assign a main engine load factor of 0.75. The intuition is that for these vessel types, even if they are moving slowly, their engines can be exerting tremdeous power while they are actually fishing with depolyed gear.
  • For all fishing vessels, we limit the main engine load factor so that it falls between 0.2 and 0.9
1.1.1.1.2 Main engine emissions

Main engine emissions for each pollutant (Table 1.1) is determined by multiplying each pollutant’s emissions factor (EF) by the main engine energy use. Main engine pollutant emission factors are derived from Appendix E in Olmer et al. (2017). For each pollutant, we use the average emissions factor for slow-speed, medium-speed, and high-speed diesel engines (SSD/MSD/HSD). SSD/MSD/HSD engines represent ~98% of vessels (Table 10 of Faber and Xing (2020)). The main engine emissions factors used for each pollutant are as follows:

Table 1.1: Main engine emissions factors, by pollutant
pollutant main_ef_g_per_kwh
CH4 0.010
CO 0.540
CO2 629.833
N2O 0.030
NOX 12.960
PM 0.605
SOX 3.917

Additionally, we apply a low-load correction factor based on the IMO Fourth GHG Report Table 20. Engines operating at very low loads below 20% operate inefficiently, and emit more of certain pollutants. Table 20 provides low-load correction factors which vary based on the exact engine load and for each pollutant. For each AIS ping, we therefore multiply the main engine emissions by this correction factor based on the engine load and pollutant. Main engine loads >20% do not get any low-load correction factor applied.

1.1.1.2 Auxilliary engine

1.1.1.2.1 Auxilliary engine energy use

Our initial modeling approach simplified the Faber and Xing (2020) recommendations by capturing the differences in auxiliary engine power consumption based on the vessel’s status, distinguishing between stationary and at-sea conditions. Speeds below 0.5 knots were considered stationary.

\[ \text{Aux engine Energy Use}_{kWh} = \text{hours} \times \begin{cases} \text{aux\_0sp}_{kW} & \text{if } \text{speed}_{knots} \leq 0.5, \\ \text{aux\_atsea}_{kW} & \text{otherwise.} \end{cases} \] Here, auxiliary engine power terms (\(\text{aux\_0sp}_{kW}\) and \(\text{aux\_atsea}_{kW}\)) were simplified from the ICCT and IMO models (which uses 4 operational phases — maneuvering, anchor, and berth, see Table 17 of the Faber and Xing (2020)). We did so by averaging auxiliary engine and boiler power for cruising and maneuvering into atsea and averaging auxiliary and boiler power for anchor and berth into 0sp. The rationale behind was that most vessels spend most of their time cruising and at ~0 speed.

While such simplification helped streamline the modeling process, it operated under the assumption of equal auxiliary emissions across vessel types and distinct operational phases, not fully capturing vessel behavior in terms of emissions. Therefore, after obtaining good validation results in the initial phase of model testing, we refined this approach to more closely follow the Faber and Xing (2020).

Below, we describe the methods followed to include disaggregate auxiliary engine and boiler power demands under four operational phase

1.1.1.2.2 Auxilliary engine and boiler energy use (4 phases)

The model described in Faber and Xing (2020), assumes that while in service, a ship is operating in one of four defined phases: at berth, at anchor, maneuvering, or at sea.

For small vessels, we follow the recommendations from the 4th IMO study (page 68, Faber and Xing (2020)), where auxiliary engine power and boiler power are relative to main engine power. For larger vessels, aux_engine_power_kw and boiler_power_kw are defined based on vessel class and operational phase.

\[ \text{Aux engine Energy Use}_{kWh} = \text{hours} \times \begin{cases} 0 & \text{if } \text{main\_engine\_power\_kw} \leq 150 \\ 0.05 \times \text{main\_engine\_power\_kw} & \text{if } \text{main\_engine\_power\_kw} \leq 500 \\ \text{aux\_engine\_power\_kw} & \text{otherwise} \end{cases} \]

\[ \text{Boiler Energy Use}_{kWh} = \text{hours} \times \begin{cases} 0 & \text{if } \text{main\_engine\_power\_kw} \leq 150 \\ \text{boiler\_power\_kw} & \text{otherwise} \end{cases} \]

The inclusion of the four phases for larger vessels requires the use of Table 17 from the Faber and Xing (2020), including energy demand for the auxiliary engine and the boiler. However, this table expresses power demand based on vessel tonnage in different units. Since GFW has vessel size in GT, we needed to convert some of the values represented in DWT, TEU, and CBM to GT. Here, we present the approach followed to establish a direct size units relationship by vessel category.

1.1.1.2.2.1 DWT conversion

To establish the GT-DWT relationship, we used data containing both GT and DWT for each vessel. By assessing the relationship between these units, which mostly present linear relationships by vessel type, we defined a simple regression allowing us to derive conversion expressions with sufficient confidence from one unit to the other.

Such data was obtained through web scraping from open online sources, containing information for 464799 vessels on variables such as type, gt, dwt, length_m, beam_m, through which we could draw the size units relationship by vessel class.

Out of all vessel types, we only need to evaluate the tonnage relationship for a few vessel types, the ones included in Table 17 from Faber and Xing (2020).

$ship_type
[1] "Bulk carrier"         "Chemical tanker"      "General cargo"       
[4] "Oil tanker"           "Other liquids tanker" "Refrigerated bulk"   
[7] "Ro-Ro"

In order to properly establish the size relationship, we need to group the categories from our dataset so they match the categories from Table 17. By doing so, we can evaluate each category’s relationship.

For instance, for Bulk carriers we have 7 categories which, according to Figure 1.1, present a linear relationship.

Figure 1.1: GT vs DWT relationship for Bulk Carriers.

The same occurs for chemical tankers with 5 categories, as shown in Figure 1.2.

Figure 1.2: GT vs DWT relationship for Chemical Tankers.

For general cargo, we have 3 categories. One of them, Passenger/General Cargo Ship, as seen in Figure 1.3, deviates from the linearity and may fall within the Ferry-pax only category from Table 17, so we will discart it.

Figure 1.3: GT vs DWT relationship for General Cargo vessels.

Related to oil tankers, several vessel categories contain the label oil. However, most of them actually belong to chemical or bulk carriers. In this grouping, we will exclusively include crude oil tankers and bitumen Tankers.

Figure 1.4: GT vs DWT relationship for Oil Tankers.

For the remaining liquid carriers, we will assign “Water Tanker”, “Wine Tanker” and “Molasses Tanker” from our table to the same category (Figure 1.5).

Figure 1.5: GT vs DWT relationship for Other Liquids Tankers category.

For refrigerated bulk, we have 2 categories, as shown in Figure 1.6.

Figure 1.6: GT vs DWT relationship for Refrigerated Bulk.

Lastly, for Ro ships, we have 3 categories, following distinct relationships as shown in Figure 1.7. Only Ro-Ro Cargo ships are the ones we are interested in, as the other two fall within the Ferry-RoPax category from Table 17.

Figure 1.7: GT vs DWT relationship for Ro-Ro.

With the defined equivalences between groups from Faber and Xing (2020) and groups from our dataset described above, we will update the original dataframe to adjust the regressions. By fitting gross tonnage (GT) based on deadweight tonnage (DWT) and grouped type, we obtain the expressions explaining the relationship between both size units per vessel type, along with the performance metrics summarized in Table Table 1.2. We’ll save this expression in a lm object and used it later to update table 17.

Table 1.2: Model summary of gross tonnage (GT) as a function of deadweight tonnage (DWT) and ship type.
r.squared adj.r.squared sigma statistic p.value df logLik AIC BIC deviance df.residual nobs
0.9935724 0.9935717 2496.954 1325496 0 7 -554797.7 1109613 1109694 374236342508 60024 60032
1.1.1.2.2.2 TEU conversion

TEU stands for Twenty-foot Equivalent Unit, which is a standard unit of measure used in the shipping industry to describe the capacity of container ships and terminals. One TEU represents the dimensions of a standard 20-foot long container.Therefore, TEU is used to quantify cargo capacity in terms of the number of 20-foot containers a vessel can carry. For example, a ship with a capacity of 10,000 TEU can carry 10,000 standard 20-foot containers.

For TEU, we will obtain GT equivalents based on the design formulas for the calculation of key design vessel characteristics from Abramowski, Cepowski, and Zvolenskỳ (2018) as detailed below:

\[ GT = -1097.4+11.049·TEU \]

1.1.1.2.2.3 CBM conversion

The size units of liquefied tankers represented as “CBM” refer to cubic meters (m³). This measurement indicates the volume capacity of the tankers, specifically how much liquefied gas (such as liquefied natural gas, LNG, or liquefied petroleum gas, LPG) they can carry.

For this unit conversion, we have not been able to find any large dataset to establish linear relationships, nor any publication defining expressions for unit conversion. The only available resource is the information from 23 vessels containing GT and CBM values, which allows to define a basic regression. This will establish the GT equivalence with intermediate to low confidence to update table 17 for gas tankers. This is a point for improvement, but for now, it will suffice.

Figure 1.8: GT vs CBM relationship for Gas Tankers.

After testing, we have seen how we can obtain a slightly better adjustment if we distinguish between LPG and LNG. However, Table 17 groups them together under the same category, so we will establish GT exclusively as a function of CBM (Table 1.3).

Table 1.3: Model summary of gross tonnage (GT) as a function of cubic meters (CBM).
r.squared adj.r.squared sigma statistic p.value df logLik AIC BIC deviance df.residual nobs
0.982426 0.9815891 7253.543 1173.945 0 1 -236.0421 478.0841 481.4906 1104891522 21 23
1.1.1.2.2.4 Updating table 17

Once we have defined the relationship between GT and the other size units for each vessel class, we will update Table 17 from (Table 1.4). This will allow us to incorporate the auxiliary engine and boiler power outputs per ship class and operational mode into our AIS-based model.

Table 1.4: Sample of Units Transformation from Table 17 of the IMO Report.
ship_type size_lower size_upper size_units size_lower_gt size_upper_gt
Bulk carrier 0 9999 dwt 0.000 7738.351
Bulk carrier 10000 34999 dwt 7738.864 20567.005
Bulk carrier 35000 59999 dwt 20567.518 33395.658
Bulk carrier 60000 99999 dwt 33396.171 53921.504
Bulk carrier 100000 199999 dwt 53922.017 105236.119
Bulk carrier 200000 NA dwt 105236.632 NA
1.1.1.2.2.5 Updating ship types to GFW classes

Our model runs using GFW vessel information tables. With the updated Table 17 including size ranges in GT, we now need to update the ship types in that table and group them according to the vessel classes available in GFW datasets. To do this, a preliminary analysis of IHS vessel types found in the GFW classes was conducted. While some could be directly assigned based on GFW classification criteria, others, such as some cargo and tankers, had to undergo a more thorough process. By evaluating the relationship between size and auxiliary power for three IHS types, we generated a regression best each types. This helped us disaggregate vessel class composition in the GFW dataset and obtain the approximate percentage composition of IHS vessel types represented within GFW vessel groups Table 1.5.

Table 1.5: IHS ship types and their equivalent GFW vessel classes. Weights have been modified based on Mark Powell’s preliminary analysis.
gfw_vessel_class tb17_ship_type weights
tug Service - tug 1.0000000
cargo General cargo 0.3200000
cargo Bulk carrier 0.4700000
cargo Container 0.1100000
cargo Ro-Ro 0.0500000
cargo Vehicle 0.0500000
cargo.bulk_carrier Bulk carrier 1.0000000
cargo.container Container 1.0000000
cargo.general General cargo 1.0000000
cargo.refrigerated Refrigerated bulk 1.0000000
cargo.ro_ro Ro-Ro 1.0000000
bunker Chemical tanker 1.0000000
reefer Refrigerated bulk 1.0000000
tanker Oil tanker 0.4500000
tanker Chemical tanker 0.3400000
tanker Other liquids tanker 0.0500000
tanker Liquefied gas tanker 0.1600000
tanker.chemical_oil Oil tanker 0.5696203
tanker.chemical_oil Chemical tanker 0.4303797
tanker.liquefied_gas Liquefied gas tanker 1.0000000
tanker.other Other liquids tanker 1.0000000
fishing Miscellaneous - fishing 1.0000000
seiners Miscellaneous - fishing 1.0000000
research Service - other 0.5000000
research Offshore 0.5000000
trawlers Miscellaneous - fishing 1.0000000
trollers Miscellaneous - fishing 1.0000000
passenger Ferry-RoPax 0.3300000
passenger Ferry-pax only 0.3300000
passenger Cruise 0.3300000
well_boat Miscellaneous - fishing 1.0000000
fixed_gear Miscellaneous - fishing 1.0000000
dive_vessel Miscellaneous - fishing 1.0000000
non_fishing Miscellaneous - other 0.5000000
non_fishing Yacht 0.5000000
fish_factory Miscellaneous - other 1.0000000
other_seines Miscellaneous - fishing 1.0000000
purse_seines Miscellaneous - fishing 1.0000000
set_gillnets Miscellaneous - fishing 1.0000000
squid_jigger Miscellaneous - fishing 1.0000000
patrol_vessel Service - other 0.5000000
patrol_vessel Offshore 0.5000000
pole_and_line Miscellaneous - fishing 1.0000000
set_longlines Miscellaneous - fishing 1.0000000
supply_vessel Service - other 0.5000000
supply_vessel Offshore 0.5000000
dredge_fishing Miscellaneous - fishing 1.0000000
pots_and_traps Miscellaneous - fishing 1.0000000
seismic_vessel Service - other 0.5000000
seismic_vessel Offshore 0.5000000
cargo_or_reefer General cargo 1.0000000
cargo_or_tanker General cargo 1.0000000
bunker_or_tanker Chemical tanker 1.0000000
container_reefer Refrigerated bulk 1.0000000
other_not_fishing Miscellaneous - other 1.0000000
tuna_purse_seines Miscellaneous - fishing 1.0000000
dredge_non_fishing Miscellaneous - fishing 1.0000000
drifting_longlines Miscellaneous - fishing 1.0000000
other_purse_seines Miscellaneous - fishing 1.0000000
specialized_reefer Refrigerated bulk 1.0000000
fish_tender Miscellaneous - fishing 1.0000000
driftnets Miscellaneous - fishing 1.0000000
other_fishing Miscellaneous - fishing 1.0000000

Using the resulting weights, we can more accurately combine IHS groups into GFW classes and obtain auxiliary and boiler power estimates for each operational phase, based on weighted means for each overlapping size range. The resulting table is stored in BQ under world-fishing-827.proj_ocean_ghg.aux_and_boil_power_by_operational_mode.

With it, we updated world-fishing-827.proj_ocean_ghg.vessel_info by including the corresponding boiler and auxiliary engine power demand considering vessel size and class for each of the four operational phases. This can then be fed into the model to provide more accurate emission estimates.

1.1.1.2.3 Auxilliary engine and boiler emissions

Auxiliary engine and boiler emissions (Table 1.6) are determined by multiplying each pollutants emissions factor by the auxiliary engine or boiler energy use. We calculate emissions for each pollutant using emissions factors derived from Appendix G and H from Olmer et al. (2017). For each pollutant, we use the average emissions factor for slow-speed, medium-speed, and high-speed diesel engines (SSD/MSD/HSD). SSD/MSD/HSD engines represent ~98% of vessels (Table 10 of Faber and Xing (2020)). The auxiliary engine and boiler emissions factors used for each pollutant are as follows:

Table 1.6: Auxilliary engine emissions factors, by pollutant
pollutant aux_ef_g_per_kwh
CH4 0.010
CO 0.540
CO2 699.667
N2O 0.033
NOX 12.116
PM 0.610
SOX 4.337

\[ \text{Aux engine emissions}_{g} = \text{hours} \times \begin{cases} \text{aux\_at\_berth}_{kW} \\ \text{aux\_at\_anchor}_{kW} \\ \text{aux\_maneuvering}_{kW} \\ \text{aux\_at\_sea}_{kW} \end{cases} \times \text{Aux emissions factor}_{g/kWh} \]

Table 1.7: Boiler emissions factors, by pollutant
pollutant boiler_ef_g_per_kwh
CH4 0.002
CO 0.200
CO2 958.000
N2O 0.043
NOX 2.033
PM 0.380
SOX 5.827

\[ \text{Boiler emissions}_{g} = \text{hours} \times \begin{cases} \text{boiler\_at\_berth}_{kW} \\ \text{boiler\_at\_anchor}_{kW} \\ \text{boiler\_maneuvering}_{kW} \\ \text{boiler\_at\_sea}_{kW} \end{cases} \times \text{Boiler emissions factor}_{g/kWh} \]

1.1.1.3 Total emissions

Finally, using the factors above, we estimate total emissions by multiplying Main Engine Energy Use, Aux Engine Energy Use and Boiler Energy Use by their respective emissions factors. We then sum the three values to get the total emission estimate.

\[ \text{Total Emissions}_{ CO_2, NO_X, ...} =\text{Main Engine Emissions}_{ CO_2, NO_X, ...} + \text{Aux Engine Emissions}_{ CO_2, NO_X, ...} + \text{Boiler Emissions}_{ CO_2, NO_X, ...} \]

1.1.1.4 Additional pollutants

For three additional pollutants (PM2.5, PM10, and VOCs), we multiply the CO2 emissions by the conversion factors in the following table. These conversion factors were provided by OceanMind, and thus our methodology is consistent wth their approach for these pollutants (which is detailed in Mayes et al. (2024)). To summarize, they are based off Table 27 from the 4th IMO report (Faber and Xing (2020)), which shows the emissions factors for each pollutant and fuel type (in terms of kg of pollutant per tonne of fuel consumed). The fuel types included are Heavy Fuel Oils (HFO), Liquefied Natural Gas (LNG), Marine Diesel Oil (MDO), and methenol. Using the 2018 factors, the emissions factors for PM2.5, PM10, and VOCs were each divided by the corresponding CO2 emissions factor, by fuel type, in order to convert between emissions of CO2 to emissions of these other pollutants, by fuel type. These emissions factor ratios were then used to calculate a a weighted average emissions factor for each pollutant, where the weighting was done by the average percentage of vessels that use each fuel type (which are taken as the 2018 values from Table 34 of the IMO report). This results in the following factors that we directly use:

gPM2.5/gCO2 gPM10/gCO2 gVOCS/gCO2
0.001598 0.001738 0.000933

1.1.2 Low sulfur fuel emissions correction factors for post-2020 data

Starting on January 1, 2020, the IMO required lower sulfur content fuel (0.5%, instead of the higher 2.5% that was typical for HDO prior to 2020). For data >= 2020-01-01, we therefore apply a new correction factor for SOX and PM to account for this lower sulfur content fuel.

For SOX: based on equation 15 (p. 74) of the 4th IMO report, the SOX emissions factor scales linearly with sulfur percentage content. For the pre-2020 SOX EF, we use Table E from the 2017 ICCT study, and take the average of the HFO value which is based on a 2.5% sulfur content (which is consistent the HFO row in Table 22 from the 4th IMO report). Assuming the sulfur content drops from 2.5% to 0.5% starting in 2020, this means an 80% drop in sulfur content, and an 80% drop in the sulfur EF. This 80% is consistent with several studies that looked the the impact this new requirement had on global sulfur emissions (@yuan2024abrupt, @yoshioka2024warming). So for the >= 2020-01-01 SOX EF, we can simply multiply our current < 2020-01-01 EF by 0.2. We can do this across the three main, aux, and boiler EFs.

For PM, PM2.5, and PM10: Equation 16 (p.74) of the 4th IMO report defines the PM10 EF relationship to sulfur content for HFO. This equation isn’t a perfect scalar factor like SOX, and depends on SFCi and has a constant. Based on Table 19 (p. 70) from the 4th IMO report, the average SFCi of HDO is 185 g/kWH (175 for SSD, 185 for MSD, and 195 for HSD). If we plug this into Equation 16, we get a < 2020-01-01 EF of 73.38 assuming a sulfur content of 2.5% (1.35+185*7*.02247*(2.5-.0246)) Assuming a post-2020 sulfur content of 0.5%, we get a >= 2020-01-01 EF of 15.18 (1.35+185*7*.02247*(0.5-.0246) ). The ratio of these two is 15.18/73.38 = 0.206, which is almost exactly the ratio for sulfur. So for the >= 2020-01-01 EFs for PM PM10 and PM2.5, we simply multiply our current < 2020-01-01 EF by 0.206. We can do this across the three main, aux, and boiler EFs of each pollutant.

1.1.3 Data

1.1.3.1 Individual AIS messages

For our individual AIS messages (pings) dataset, we leverage the latest-and-greatest version of the GFW’s AIS pipeline, Version 3. This is one of the GFW’s core internal datasets. This process automates the parsing, cleaning, augmenting, and publishing of raw AIS data (Kroodsma et al. (2018)). This table provides data from 2012 to present. Using this table as our starting point, we are able to estimate emissions from all analyzed pollutants for every single AIS message. These ping-level emissions can then later be aggregated however desired (e.g., by vessel, by voyage, by destination or arrival port, by time, by space, etc.)

Variables of interest within this table include the following:

  • ssvid: source specific vessel id; MMSI for AIS
  • hours: time since the previous position in the segment
  • speed_knots: speed (knots) from AIS message
  • implied_speed_knots: distance from last position divided by hours since last position
  • meters_to_prev: distance (meters) to the previous point in the segment
  • distance_from_shore_m: distance from shore (meters)
  • distance_from_port_m: distance from port (meters)
  • neural net score: The score is 1 if the neural net thinks this is a fishing position.
  • night_loitering: 1 if the seg_id of every message of a squid_jigger that is at night and not moving, 0 if not.

In order to minimize noisy data, we only include AIS messages that occur within valid segments (i.e., select seg_id from pipe_ais_v3_published.segs_activity where good_seg), and only also within daily segments that are not overlapping with each other (i.e., those that do not occur in overlap_segs_daily_v20241202).

1.1.3.2 Vessel characteristics

Vessel characteristics also represent another one of the core GFW datasets. These tables provides metadata for all vessels contained within GFW, organized by MMSI. The information for each vessel includes: 1) official registry information, when available (Park et al. (2023)); or 2) algorithm-derived vessel characteristics such as vessel class, engine power, and gross tonnage, when registry data are not available (Kroodsma et al. (2018)). The GFW vessel characteristics database leverages extensive work that has been done to scrape and aggregate many publicly available vessel registries (Park et al. (2023)). Note that we are currently using a cutting edge version of this database, which differs from Version 3 of the pipeline in that it uses a new experimental random forest algorithm for inferring certain vessel characteristics when they are not available in official vessel registries (vessel type, main engine power, length, gross tonnage, and max speed).

We are also leveraging two brand new cargo and tanker vessel type classification sub-models developed by GFW that build off the general vessel classification algorithm for low information vessels (i.e., those that don’t have known registry information). We can now differentiate many vessels that were previously lumped together as an undifferentiated cargo vessel type into the specific categories of bulk_carrier, container, general, refrigerated, and ro_ro. We can also now differentiate many vessels that were previously lumped together as an undifferentiated tanker vessel type into the specific categories of oil or chemical, liquefied gas, and other liquids. These classes now align with the IMO cargo vessel class types, allowing us to more accurately assign auxiliary engine power and boiler power for these vessel classes using the IMO methodology. These new models each leverage a random forest that is trained on information on port visit patterns vessels with known IMO cargo or tanker vessel class types. Sequences of port visits are converted into usable model features by implementing a Word2Vec model that transforms them into numeric arrays called embeddings. Additionally, each model also uses model features based on vessel activity including port hours, average distance from shore, and average speed. This allows us to more accurately classify cargo or tanker vessel types by looking at the most common IMO known vessel types that use the same ports. The cargo sub-model achieves am F1 weighted average score of 91%, while the tanker sub-model achieves am F1 weighted average score of 95%.

Variables of interest from pipe_ais_v3_published.vi_ssvid_v20250201 (the core GFW pipeline 3 vessel characteristics table) include the following:

  • ssvid: source specific vessel id; MMSI for AIS
  • best.flag: best flag state (ISO3) for the vessel
  • activity.active_hours: hours the vessel was broadcasting AIS and moving more than 0.1 knots. If desired, we can use this as a filter; vessels with < 24 hours of active hours have very limited data from which to calculate emissions from.
  • registry_info.registries_listed: vessel registries the vessel is listed on
  • registry_info.best_known_shipname: best known shipname for the vessel from registries
  • ais_identity.n_shipname_mostcommon.value: the most common normalized shipname broadcasted by this vessel
  • registry_info.best_known_callsign: best known callsign for vessel from registries
  • ais_identity.n_callsign_mostcommon.value: the most common normalized callsign broadcasted by this vessel
  • registry_info.best_known_imo imo_registry: best known IMO number for the vessel from registries
  • ais_identity.n_imo_mostcommon.value imo_ais: the most common normalized IMO number broadcasted by this vessel
  • offsetting: true if this vessel has been seen with an offset position at some point between 2012 and 2019 (this should be FALSE; if it is TRUE, it can be used as a filter to remove potentially erroneous/noisy vessels)
  • overlap_hours_multinames: the total numbers of hours of overlap between two segments where, over the time period of the two segments that overlap (including the non-overlapping time of the segments), the vessel broadcast two or more normalized name, where each normalized name was broadcast at least 10 or more times. That is a bit complicated, but the goal is to identify overlapping segments where there were likely more than one identity. (this should be 0; if it is > 0, it can be used as a filter to remove potentially erroneous/noisy vessels)

Variables of interest from proj_ocean_ghg.rf_predictions_v20250613 (the new experimental random forest vessel characteristics table developed for this project) include the following:

  • ssvid: source specific vessel id; MMSI for AIS
  • rf_best_vessel_class: best vessel class for the vessel (using official registry information where available; or the random forest vessel characteristics algorithm where registry information is not available. For cargo or tanker vessels where registry information is not available, we use the new cargo and tanker vessel specific classification sub-model.)
  • rf_best_engine_power_kw: best engine power (kilowatts) for the vessel (using official registry information where available, or the random forest vessel characteristics algorithm where not available)
  • rf_best_tonnage_gt: best tonnage (gross tons) for the vessel (using official registry information where available, or the random forest vessel characteristics algorithm where not available)
  • rf_best_length_m best length (meters) for the vessel (using official registry information where available, or the random forest vessel characteristics algorithm where not available)
  • rf_best_max_speed_knot best maximum speed (i.e., design speed) (knots) for the vessel (using official registry information where available, or the random forest vessel characteristics algorithm where not available)
  • on_fishing_list_rf_best: GFW determination of whether the vessel is a fishing vessel, using information from vessel registries and the random forest model

Variables of interest from proj_ocean_ghg.cargo_subtypes_v20250613 (the new experimental cargo and tanker sub-model) include the following:

  • ssvid: source specific vessel id; MMSI for AIS
  • vessel_subclass: Cargo or tanker sub-class, where available

There are currently 1,067,985 AIS-broadcasting vessels in the GFW dataset (this number excludes any AIS transponders which are labeled as fishing gear, helicopters, or submarines). Of these, we are able to estimate emissions 925,682 unique active vessels over our entire time period. Of these, 901,214 (97%) are ‘low-information’ vessels without an IMO number.

1.1.3.3 Voyages

Again leveraging Version 3 of the pipeline, GFW’s voyages table contains information for port-to-port voyages made by vessels. This table leverages extensive work done by the GFW team to: 1) define ports, 2) determine when vessels arrive at or depart from a port, and 3) determine voyages that are define by a port departure and a port arrival (Watch (2021)).

To define ports, specific anchorages are first identified by using the AIS data to find S2 cell locations where at least 20 unique vessels remained stationary at some point since 2012 (where ‘stationary’ is defined as moving less than 0.5km within a 12-hour period). Once these initial anchorage locations have been identified, anchorages within 4km of each other are grouped into ports. In this way, a single port may contain multiple anchorages within that port. Port names are then assigned to each of these locations spatially according to the following heirarchy:

  1. World Port Index
  2. GeoNames 1000 database that describes all settlements globally that have a population of at least 1,000 people
  3. The top destination reported in AIS messages of stationary vessels that defined that anchorage
  4. Contributed names and regional port databases

Once ports have been identified, heuristics are used to identify port entries and exits:

A vessel enters port when it comes within 3 kilometers of an anchorage point and exits port when it is outside 4 kilometers of the anchorage point. We use different threshold distances to avoid situations where a vessel continuously enters and exits port. This situation is still common, however, as vessels travel along coastlines and repeatedly come within close proximity to numerous anchorages. To distinguish actual port visits from coastal transits, we further identify when a vessel appears to stop at a given port. The vessel is considered to have “stopped” at port if its speed drops below 0.2 knots, and this port stop ends when the speed rises above 0.5 knots. AIS is often switched off when a vessel enters port, and it is turned back on when it leaves. As a result, we track port “gaps,” where a vessel that has entered port does not broadcast on AIS for at least four hours. Port stops and port gaps are behaviors indicating that a vessel visited a port for a specific reason and/or engaged in some activity while at the port, such as landing catch or exchanging supplies and crew. We can then allocate the at-sea activity of vessels to individual voyages between port visits.

Once these port entries and exits have been identified, voyages can simply be defined as all activity between those port events. In this way, individual AIS message (and its associated emissions) can be assigned to a voyage.

Variables of interest within this table include the following:

  • ssvid: source specific vessel id; MMSI for AIS
  • trip_id: A unique identifier for the trip generated by the ssvid, vesssel-id and the exit time of starting visit
  • trip_start: The initial timestamp of the voyage, when the vessel leaves port
  • trip_end: The final timestamp of the voyage, when the vessel reaches port
  • trip_start_anchorage_id: The id of the anchorage where the voyage starts
  • trip_end_anchorage_id: The id of the anchorage where the voyage ends

Further information of the starting and ending anchorages can be obtained by joining this table to GFW’s anchorage table, which includes the following:

  • anchorage_id: The id of the anchorage
  • label: Port name
  • iso3: Port ISO3 code
  • lat: latitude of the anchorage
  • lon: longitude of the anchorage

In summary, from 2015-01-01 to 2025-06-30, we have 145,747,116 unique voyages across 805,021 unique vessels. These trips visited 14,682 unique ports across 209 unique countries.

1.1.3.4 Port visits

Again leveraging Version 3 of the pipeline, we use GFW’s port visits table. Port visits are determinede using the same methods as describe above for assigning voyages. Variables of interest within this table include the following:

  • ssvid: source specific vessel id; MMSI for AIS
  • visit_id: Unique ID for this visit
  • start_timestamp: timestamp at which vessel crossed into the anchorage
  • end_timestamp: timestamp at which vessel crossed out the anchorage
  • start_anchorage_id: anchorage_id of anchorage where vessel entered port
  • end_anchorage_id: anchorage_id of anchorage where vessel exited port
  • confidence: How confident are we that this is a real visit based on components of the visits: 1 -> no stop or gap; only an entry and/or exit 2 -> only stop and/or gap; no entry or exit 3 -> port entry or exit with stop and/or gap 4 -> port entry and exit with stop and/or gap

For quality control purposes, we filter this dataset to just those port visits with the highest confidence level, 4. We also only filter to those port visits where the starting and ending port label are the same.

As with the voyages dataset, for information of the port starting and ending anchorages can be obtained by joining this table to GFW’s anchorage table. Variables of interest within this table include the following:

  • anchorage_id: The id of the anchorage
  • label: Port name
  • iso3: Port ISO3 code
  • lat: latitude of the anchorage
  • lon: longitude of the anchorage

Note again that since a single port can have multiple anchorages, it is possible that a single port visit has different starting and ending anchorages, and therefore lat/long locations.

In summary, from 2015-01-01 to 2025-06-30, we have 109,617,280 unique port visits across 835,196 unique vessels. These port visits trips occurred in 14,431 unique ports across 209 unique countries.

1.1.4 Areas of potential model refinement

We have identified a number of areas for potential model refinement. They are all related to the need for improved vessel characteristics metadata:

  1. Vessel classification: Continue to align GFW vessel classes with IHS vessel classes: Some GFW and IHS vessel classes are currently categorized slightly differently (for example, those related to tankers), meaning that we need to translate and aggregate certain information that is provided by the ICCT and IMO for IHS vessel classes (i.e., auxiliary engine power by vessel type) into the GFW vessel classes.
  2. Draft correction factor: Currently, we use the same draft correction factor for all vessels. This single draft correction factor is currently an average of vessel class-specific correction factors, weighted by the total emissions by each vessel class. Future model iterations may want to use vessel class-specific draft factors.
  3. Size units conversion: The inclusion of the four operational phases requires the use of auxiliary engine and boiler energy demand values by vessel size. As described earlier, this entails setting unit conversion expressions that can be refined to better capture energy demand, especially those for CBM conversion.

1.2 Results

In this section, we provide some high-level results from our emissions model.

1.2.2 Emissions by vessel class

Next, for 2024, we summarize the global annual emissions by vessel class (Figure 1.13). These are summarized using the GFW vessel class categories. We first plot simply CO2 emissions, then look at a table of all pollutants.

Figure 1.13: Summary of global 2024 CO2 emissions by vessel class

We next can look at a table of all pollutants for 2024 (Table 1.12):

Table 1.12: Summary of global 2024 emissions by vessel class and pollutant
vessel_class CO2 NOX SOX VOCS CO PM10 PM2_5 PM N2O CH4
tanker.chemical_oil 240,047,644 3,377,729 296,161 242,988 158,184 88,232 81,125 38,552 11,415 2,948
passenger 228,838,747 2,036,916 280,603 230,835 109,218 84,011 77,243 29,319 10,730 1,933
cargo.container 196,436,153 4,137,712 243,953 272,227 201,667 81,037 74,509 43,233 10,161 4,469
cargo.bulk_carrier 169,353,456 3,366,512 210,302 186,232 151,505 64,038 58,880 33,987 8,308 3,015
cargo.general 73,742,032 1,356,534 91,364 91,964 65,538 29,184 26,833 14,610 3,707 1,404
tanker.liquefied_gas 73,270,269 1,388,193 90,924 74,948 61,229 27,026 24,849 14,055 3,539 1,180
trawlers 48,463,405 989,333 60,218 54,239 47,185 18,466 16,979 9,954 2,394 901
cargo.ro_ro 42,925,936 807,143 53,251 45,564 36,209 16,032 14,741 8,268 2,088 712
tug 18,769,929 417,420 23,297 34,802 22,384 8,793 8,085 4,649 1,053 574
bunker 15,275,039 85,166 18,658 14,743 5,262 5,528 5,083 1,606 703 80
other_fishing 12,951,008 272,055 16,082 18,464 15,381 5,425 4,988 2,883 678 304
supply_vessel 7,578,618 156,834 9,405 11,480 7,971 3,243 2,981 1,699 402 187
specialized_reefer 7,313,700 119,952 9,052 7,314 5,496 2,678 2,462 1,286 351 104
dredge_non_fishing 4,124,723 82,553 5,120 5,179 3,908 1,636 1,505 862 209 84
patrol_vessel 3,827,480 75,893 4,748 5,274 3,750 1,575 1,448 816 198 85
drifting_longlines 3,820,624 79,854 4,745 5,200 4,367 1,570 1,443 837 198 86
container_reefer 3,698,132 70,144 4,591 3,632 3,066 1,346 1,237 705 178 58
seismic_vessel 2,776,295 60,538 3,447 4,644 3,126 1,240 1,140 659 151 77
other_not_fishing 2,580,742 35,281 3,178 3,384 1,900 1,041 957 441 130 42
squid_jigger 2,382,356 50,498 2,959 3,363 2,820 994 914 531 124 56
tuna_purse_seines 1,817,801 39,707 2,259 2,576 2,189 760 698 412 95 43
cargo.refrigerated 1,467,227 26,757 1,820 1,413 1,175 531 488 272 70 22
set_gillnets 1,145,545 24,347 1,423 1,671 1,393 485 446 258 60 28
dive_vessel 931,186 20,111 1,156 1,505 1,039 410 377 218 50 25
set_longlines 917,084 19,652 1,139 1,331 1,115 387 356 207 48 22
pots_and_traps 850,512 18,611 1,057 1,308 1,091 368 338 198 45 22
pole_and_line 444,109 9,336 551 630 526 186 171 99 23 10
well_boat 411,784 8,011 511 434 358 153 141 81 20 7
research 306,239 5,905 380 364 278 119 109 62 15 6
reefer 289,423 3,601 357 284 175 105 97 43 14 3
tanker.other 286,701 2,681 352 293 143 106 97 38 13 3
dredge_fishing 249,167 5,146 309 319 270 100 92 53 13 5
other_seines 142,725 3,041 177 208 174 60 55 32 8 3
other_purse_seines 55,792 1,219 69 82 69 24 22 13 3 1
fish_factory 49,560 860 61 73 46 21 19 10 3 1
trollers 48,789 1,090 61 79 66 22 20 12 3 1
bunker_or_tanker 14,965 74 18 14 5 5 5 2 1 0
driftnets 769 16 1 1 1 0 0 0 0 0

1.2.3 Spatial maps of emissions

Next we can look at spatial maps of emissions by pollutant, aggregated across all vessel types. These maps are shown at a spatial resolution of 0.1x0.1 degrees (Figure 1.14 - Figure 1.20).

Reading layer `World_Countries_Generalized' from data source 
  `/Users/gmcdonald/github/ocean-ghg/data/raw/World_Countries_Generalized_Shapefile/World_Countries_Generalized.shp' 
  using driver `ESRI Shapefile'
Simple feature collection with 251 features and 4 fields
Geometry type: MULTIPOLYGON
Dimension:     XY
Bounding box:  xmin: -20037510 ymin: -30240970 xmax: 20037510 ymax: 18418390
Projected CRS: WGS 84 / Pseudo-Mercator
Figure 1.14: Map of 2024 CO2 emissions, aggregated across all vessel classes, at 0.1x0.1 degree spatial resolution.
Figure 1.15: Map of 2024 NOX emissions, aggregated across all vessel classes, at 0.1x0.1 degree spatial resolution.
Figure 1.16: Map of 2024 SOX emissions, aggregated across all vessel classes, at 0.1x0.1 degree spatial resolution.
Figure 1.17: Map of 2024 CH4 emissions, aggregated across all vessel classes, at 0.1x0.1 degree spatial resolution.
Figure 1.18: Map of 2024 CO emissions, aggregated across all vessel classes, at 0.1x0.1 degree spatial resolution.
Figure 1.19: Map of 2024 N2O emissions, aggregated across all vessel classes, at 0.1x0.1 degree spatial resolution.
Figure 1.20: Map of 2024 PM emissions, aggregated across all vessel classes, at 0.1x0.1 degree spatial resolution.
Figure 1.21: Map of 2024 PM2.5 emissions, aggregated across all vessel classes, at 0.1x0.1 degree spatial resolution.
Figure 1.22: Map of 2024 PM10 emissions, aggregated across all vessel classes, at 0.1x0.1 degree spatial resolution.
Figure 1.23: Map of 2024 VOCS emissions, aggregated across all vessel classes, at 0.1x0.1 degree spatial resolution.

1.3 Model Validation

The initial stages of model development involved multiple rounds of preliminary validation against testing model versions to identify the best-performing one among those evaluated. Following this testing phase, and with the AIS-based model built as described above, we validated it against measured absolute emission values from a dataset of a European Union (EU) monitoring program.

1.3.1 EU emissions data

We used the \(CO_2\) emissions data from maritime transport provided by the European Maritime Safety Agency, which is part of the monitoring, reporting, and verification program of carbon emissions from maritime transport, set by the Regulation (EU) 2015/757. As detailed below, vessels with certain characteristics and certain trips made by such vessels operating in EEA sea ports, must report their emissions on an annual basis.

1.3.1.1 What’s in an out of EU 2015/757

Here, we summarize the relevant provisions of this legislation that define the vessels and trip characteristics that need to be filtered from our data for validation purposes.

Vessel characteristics:

  • Ships above 5000 gross tonnage.

  • Excludes warships, naval auxiliaries, fish-catching or fish-processing ships, wooden ships of a primitive build, ships not propelled by mechanical means, or government ships used for non-commercial purposes.

Trips characteristics:

  • Trips from their last port of call to a port of call under the jurisdiction of a Member State and from a port of call under the jurisdiction of a Member State to their next port of call, as well as within ports of call under the jurisdiction of a Member State.

  • Vessel activities covered by the law are those when the ships are at sea as well as at berth.

  • Ship at berth includes any ship which is securely moored or anchored in a port falling under the jurisdiction of a Member State while it is loading, unloading or hotelling, including the time spent when not engaged in cargo operations.

  • Port of call means the port where a ship stops to load or unload cargo or to embark or disembark passengers; consequently, stops for the sole purposes of refueling, obtaining supplies, relieving the crew, going into dry-dock or making repairs to the ship and/or its equipment, stops in port because the ship is in need of assistance or in distress, ship-to-ship transfers carried out outside ports, and stops for the sole purpose of taking shelter from adverse weather or rendered necessary by search and rescue activities are excluded.

  • Ship to ship transfers carried out outside ports are covered by the Regulation when these transfers take place as part of a voyage starting and/or ending with a port of call under the jurisdiction of a Member State.

  • A ship to ship transfer carried out outside ports is not considered as a port of call. As a consequence, if for example, a vessel leaves an EEA port, arrives to a US harbor performs a ship to ship operation outside the port limits, and then goes to South Corea the emissions falling within MRV scope will be the emissions released during the whole voyage from the EEA port of call until the port of call in South Korea. However, if the ship to ship transfer can be carried out within the port limits, that operation would constitute a port of call. The voyage covered by the MRV Maritime Regulation would then be an EEA port of call to a US port of call.

  • If a ship performs more than 300 voyages during the reporting period and all of these voyages either start from or end in a port within a Member State, the company can be exempted from monitoring the detailed parameters for each voyage.

Time ranges:

  • The reporting period covers one calendar year during which CO2 emissions have to be monitored. For voyages starting and ending in two different calendar years, the monitoring and reporting data shall be accounted under the first calendar year concerned.

1.3.2 Data filtering

Given the law’s exceptions, in order to validate emission estimates, we need to filter our results to match the EU data contents and aggregation. In this regard, the vessel characteristics in terms of type and gross tonnage can be filtered by simply assessing which vessel IDs are included in the EU dataset. As for the time range, since the data is aggregated yearly by the starting date of a trip, the filtering is straightforward.

The most challenging aspect of this data filtering corresponds to the trip characteristics, specifically when defining the “ports of call.” The main obstacle is the inability to determine the type of activity a vessel is engaged in while in port, which prevents us from identifying which trips are included within those aggregated yearly emissions. To work around this, and establish which vessels’ emissions can be included in the validation, we have compared the total distance and total hours at sea reported in the EU data to the total distance traveled across all trips in our trip-level emissions for that vessel and year. If these numbers were within ±5%, ±10%, or ±15% difference, then we could reasonably assumed that aggregated trips in the GFW dataset correspond to trips included in the EU dataset.

The EU data has been compiled and stored in proj_ocean_ghg.eu_validation_data, while the selection and part of the filtering from our data has been conducted through the queries in eu_validation_trip.sql and eu_validation_port.sql, differentiating between emissions by trip and port visits, and stored in proj_ocean_ghg.eu_validation_trip and proj_ocean_ghg.eu_validation_port respectively.

1.3.3 Validation results

Emissions estimates are divided between emissions at sea and emissions at port. Here, we present the differences between our results and the EU emissions data.

1.3.3.1 Trip emissions

As detailed in the Methods section, we have defined trip emissions as those occurring between ports. In the EU dataset, these emissions are categorized under three variables: emissions from EEA-EEA, EEA-NonEEA, and NonEEA-EEA seaports. We have run our model, selected the data by trips involving at least one EEA port, aggregated the results by year, and selected vessels listed in the EU dataset for a specific year.

After discarding 1077 observations with no time at sea, and assessing potential duplicates due to different ssvid for the same imo_number (0 duplicates in this dataset), we ended up with a total of 217103 observations—year and vessel—selected from our emission estimates. This is 87.7% of the EU data observations. By applying different margin values to the annual hours at sea we get the following performance metrics detailed in Table 1.13. As we can see, higher performance is achieved applying a 10% margin, with which we obtained a selection of 1599 observations for validation.

Table 1.13: Performance metrics for Model 6.3 validated against the EU dataset using the hours at sea selection method and different margin values.
mae rmse nrmse rsq rsq_trad mape mpe threshold n_observations
2231.621 5068.543 0.456 0.801 0.792 29.600 -3.216 0.05 926
2353.508 5099.621 0.454 0.798 0.794 29.948 -3.458 0.10 1599
2410.567 5028.879 0.441 0.807 0.805 Inf -Inf 0.15 2277
2579.295 5612.569 0.501 0.766 0.749 Inf -Inf 0.20 2919
2735.795 5854.677 0.515 0.760 0.734 Inf -Inf 0.25 3621
3103.941 7982.631 0.690 0.655 0.523 Inf -Inf 0.30 4396
3386.662 8373.700 0.724 0.670 0.476 Inf -Inf 0.35 5289
3862.253 9200.055 0.775 0.681 0.399 Inf -Inf 0.40 6341

Comparing against the EU dataset with that 10% margin, the model demonstrates a good fit with an R-squared value of 0.8. While the MAE and RMSE indicate a considerable average magnitude of prediction errors, we need to consider that the average emission values are quite large since we are estimating absolute emissions per year. In fact, the normalized RMSE suggests that the errors, relative to the range of observed values, are moderate. Further, the model exhibits a tendency to underestimate emissions, as evidenced by the MPE. Despite some limitations, this validation framework offers a useful alternative to the one derived from the ML results, avoiding the related outlier inconveniences from emissions expressed in distance units. We can visually explore the relationship between our results and the EU data, observing certain linearity between both emission values with some spreading as the emission values increase (Figure 1.24).

Figure 1.24: Relationship between EU data and our emission estimates using hours at sea and a 10% margin.

We can double-check these results by applying the data filtering margin on distance traveled instead of hours. In this regard, while the EU dataset does not contain information on the total nautical miles (nm) navigated, we can extract this from the annual average \(CO_2\) emissions per distance. If we apply distinct margins to the total nautical miles (nm) navigated, we achieve the highest performance values for the 25% margin, with which we select 3551 observations. Overall, the performance metrics are slightly worse than those obtained by using hours at sea, which was a direct measure available in the EU dataset, rather than an indirect estimate like total nm navigated (Figure 1.25 and Table 1.14).

Table 1.14: Performance metrics for Model 6.3 validated against the EU dataset using total nm selection method and different margin values.
mae rmse nrmse rsq rsq_trad mape mpe threshold n_observations
1788.215 3866.017 0.360 0.872 0.870 23.392 1.76 0.05 1090
1811.906 3707.641 0.347 0.881 0.879 Inf -Inf 0.10 1738
1933.623 3964.672 0.346 0.881 0.880 Inf -Inf 0.15 2346
2111.669 7541.384 0.660 0.654 0.564 Inf -Inf 0.20 2934
2260.113 7599.934 0.656 0.675 0.570 Inf -Inf 0.25 3551
2435.037 7424.169 0.645 0.696 0.584 Inf -Inf 0.30 4286
2701.152 7510.823 0.655 0.707 0.571 Inf -Inf 0.35 5124
3087.718 7947.216 0.677 0.724 0.542 Inf -Inf 0.40 6130
Figure 1.25: Relationship between EU data and our emission estimates using navigated nm and a 25% margin.

1.3.3.2 Port emissions

As for emissions at ports, these consist of emissions during port stays. Here, we have filtered those emission results by port visits within the EEA and followed the same aggregation procedure as for trips. However, some assumptions have been made due to the impossibility of filtering out stays that may not be included in the EU dataset. As mentioned earlier, one of the inconveniences related to using this dataset is the definition of “ports of call”, which establishes whether a vessel trip and port visit is considered or not under the regulation and, by extension, whether its emissions are reported and available in the EU dataset. Unable to define the activities performed by a vessel in a port, we assume that we will be overestimating the emissions at port since the EU does not include them all. With it, and applying the same procedure, we see that the performance values are quite poor given that overestimation, leaving us with the need to find an alternative way to validate our port emission results while improving our models estimates (Table 1.15 and Figure 1.26).

Table 1.15: Performance metrics for Model 6.3 validated against the EU dataset port emissions.
model mae rmse nrmse rsq rsq_trad mape mpe
Model 6.3 7369.646 17146.17 12.431 0.137 -153.529 Inf -Inf
Figure 1.26: Relationship between EU data and our port emission estimates.

By applying an additional filtering step, selecting those observations from the trips under the 10% hours at sea margin approach, we would expect to narrow down our observations to those actually happening under “ports of call”. However, as the performance values in Table 1.16 and the emissions relationship shown in Figure 1.27 indicate, our model still does not capture emissions at port as reflected in the EU dataset, highlighting the need for improvement in this aspect.

Table 1.16: Performance metrics for Model 6.3 validated against the EU dataset port emissions for a subset of observations.
model mae rmse nrmse rsq rsq_trad mape mpe
Model 6.3 1382.316 6236.424 4.196 0.079 -16.616 Inf -Inf
Figure 1.27: Relationship between EU data and a subset of our port emission estimates.
1.3.3.2.1 2 vs 4 operational phases

As described in the methods section, we tested the inclusion of 2 operational phases and 4 operational phases against the entire GFW dataset. The inconvenience of the first approach was the oversimplification of such phases—with different energy demands—when vessel movement is minimal, presumably failing to quantify emissions accurately, especially when vessels are at port. Additionally, the inclusion of all 4 operational phases incorporated the emissions from boilers, potentially reducing the underestimation of emissions from the 2 operational phases model.

The validation results for both approaches showed a slightly better performance for trip emissions (0.02 superior in R2) for the 2 operational phases model, although the 4 operational phases model underestimated the actual emissions less. As for port emissions, as expected, they considerably improved under the 4 operational phases approach, despite still presenting low performance. Therefore, while the inclusion of 4 operational phases allows us to closely follow the ICCT/IMO methodology, the validation results also show a clear advantage of implementing such refinement, with a slight drawback in trip emissions performance. However, we must consider that the inclusion of these additional phases comes with several assumptions in vessel class grouping, giving room for additional improvements. Better vessel class resolution in the GFW datasets could potentially translate to improved performance. Furthermore, alternative validation datasets, yet to be defined, could help establish clearer confidence intervals.

1.3.4 Comparison to other global emissions estimates

Several other emissions estimates have been done by the International Maritime Organization (IMO) (most recently for 2018), Emissions Database for Global Atmospheric Research (EDGAR) in (most recently for 2023), Organization for Economic Co-operation and Development (OECD) (most recently for 2024 using their experimental database). Below we compare our GHG emissions estimates with the findings of these studies to validate our results

The emissions inventories used in this comparison rely on a range of methods and have some differences in sector definitions and underlying data sources. Here we compare the most recent year of data from each inventory.

Highlights include:

  • GFW’s total estimated emissions (1.35 billion MT CO2 for both broadcasting vessels and non-broadcasting vessels, and 1.17 billion MT CO2 for just AIS-broadcasting vessels) are similar to OECD, EDGAR, and IMO, indicating high reliability and utility of our methodology and results.

  • IMO’s, OECD’s, and EDGAR’s emissions estimates do not include “dark” fleets, making GFW’s dark fleet estimates the first of their kind.

Comparison of our global annual emissions estimates with other inventories (the most recent year available for each inventory is shwon).
CO2 emissions (billion MT) Data source
1.35 GFW (2024, AIS + S1)
1.17 GFW (2024, AIS)
0.97 OECD (2024)
0.91 EDGAR (2023)
1.06 IMO (2018)