2 Dark fleet emissions model
3 S1 Dark fleet emissions model
The current new version of the dark emissions model incorporates several improvements made since the version released in September. The model uses Sentinel-1 (S1) Synthetic Aperture Radar (SAR) data to estimate emissions from the ‘dark fleet.’ The ‘dark fleet’ term refers to vessels that do not broadcast AIS and, as a result, are not captured in AIS-based datasets (Rowlands et al. (2019)).
3.1 Methods
3.1.1 Model overview
To estimate emissions from the dark fleet, we spatiotemporally extrapolate our AIS-based emissions estimates to the dark fleet based on spatiotemporal vessel detections. For every S1 detection, GFW has determined whether or not the vessel is matched to an AIS vessel that was broadcasting at the same location and time, allowing us to determine the number of broadcasting and non-broadcasting vessels in a given location and time. We can also make this extrapolation disaggregated by vessel type and length, as the GFW S1 model can identify whether each dark fleet detection represents a fishing or non-fishing vessel and determine its length. See Paolo et al. (2024) for further information on the S1 detection model.
Our approach for estimating emissions for the dark fleet is as follows:
- Grid the ocean to 1x1 degree pixels.
- Use a monthly temporal aggregation unit.
- For each vessel type (fishing and non-fishing), use length percentiles to bin vessels into one of the 10 size classes. Set bins based on S1 unmatched detected length and assign same cutoffs to both the AIS-based vessel characteristics as well as the detected vessel length from S1. We will extrapolate AIS-based emissions to S1 emissions by vessel size class. We therefore make extrapolations from generally larger AIS-broadcasting vessels to generally larger dark fleet vessels, generally smaller AIS-broadcasting vessels to generally smaller dark fleet vessels, etc. Looking at Figure 3.1, the size distributions for fishing vessels across the length bins are very similar between S1 unmatched detections and AIS-broadcasting vessels.
For each 1x1 degree pixel, month, vessel type (fishing or non-fishing), and vessel size class, use our AIS-based emissions model to determine the amount of AIS-based emissions for each pollutant (CO2, CH4, N2O, NOX, SOX, CO, PM, PM10, PM2.5 and VOCs)
For each 1x1 degree pixel, month, vessel type, and vessel size class, use S1 SAR data to determine the following: 1) the number vessel detections that are matched to an AIS-broadcasting vessel; 2) the number vessel detections that are not matched to an AIS-broadcasting vessel (i.e., the number of dark vessels); and 3) the ratio of dark vessels to AIS-broadcasting vessels
Assigning missing ratios: S1 does not provide complete global spatial coverage. Its coverage is determined by the satellite’s orbital path and the presence of land, leading to limited coverage in some coastal areas and most of the open ocean. Within the S1 footprint, there may be pixels where, for a given time and vessel category, no detections are matched with any AIS-broadcasting vessels, resulting in
NULLratios. To estimate ratios for pixels that do not have a value, we calculate the K-nearest neighbors (KNN) for the 8 closest pixels having a ratio value from Step 4, and assign the mean value among these KNN ratios by month, vessel type, and vessel size class. 8 was chosen as the optimum value since it maximizesrsq_tradwhen trying to infer the ratio for pixels that have known ratios.Using the ratios generated in Steps 4-5, we multiply our AIS-based emissions estimates by the ratio of dark vessels to AIS-broadcasting vessels, for each pollutant by pixel, month and vessel category. We use a hierarchy for which ratio to use: 1) monthly pixel-level ratio if available; 2) KNN ratio if pixel-level ratio is not available.
This then gives us, for every month and location where AIS-based emissions estimates are available, corresponding dark fleet emissions estimates. Adding these two numbers together gives us yearly gridded total emissions estimates from across the AIS-broadcasting and dark fleets.
As an example, we can look at a hypothetical 1x1 degree pixel, in a hypothetical month, representing the S1 detections corresponding to a specific vessel size bin (Figure 2). We can see detections for fishing vessels and non-fishing vessels (e.g., cargos, tankers, etc). First, we consider fishing vessels. Here we can see that S1 had 2 detections that we were able to match to AIS-broadcasting vessels, and 8 detections which were unmatched (e.g., these are from non-broadcasting vessels). Therefore, the ratio of dark-to-broadcasting vessels is 8/2 = 4. For this pixel, month, vessel type, and size bin, we would therefore take our AIS-broadcasting emissions and multiply it by 4 to obtain our non-broadcasting emissions estimate. If our observed AIS-broadcasting emissions for fishing vessels was 5 metric tonnes of CO2, our non-broadcasting emissions would be 60*0.33 = 20 metric tonnes of CO2. This calculation is done for each of the nine pollutants.
Next, we consider non-fishing vessels. Here we can see that S1 had 15 detections that we were able to match to AIS-broadcasting vessels, and 5 detections which were unmatched (e.g., these are from non-broadcasting vessels). Therefore, the ratio of dark-to-broadcasting vessels is 5/15 = 0.33. For this pixel, month, size bin, we would therefore take our AIS-broadcasting emissions and multiply it by 0.33 to obtain our non-broadcasting emissions estimate. If our observed AIS-broadcasting emissions were 60 metric tonnes of CO2 for non-fishing vessels, our non-broadcasting emissions would be 60*0.33 = 20 metric tonnes of CO2. This calculation is done for each of the nine pollutants.
3.1.2 Data
3.1.2.1 S1 vessel detections
The S1 vessel detections table includes all the necessary variables to determine detection locations, timestamps, lengths and whether each detection is matched to an AIS-broadcasting vessel. It also contains the required parameters for filtering out noise. The latest table update corresponds to sentinel1_clean_v20250709.
Below is a summary of the variables included:
detect_id: Unique S1 vessel detection id.corrected_length_m: Inferred length from S1 detection, which has additionally been corrected for small vessels <22m using the relationship between known length and inferred length based on years of confident matches.presence: Presence of a vessel in the S1 detection. A value >0.7 indicates the reliable presence of a vessel.ssvid: ssvid (MMSI for AIS) of any AIS-broadcasting vessel spatiotemporally matched to the S1 detection.detect_lat: Latitude of vessel detection.detect_lon: Longitude of vessel detection.detect_timestamp: Timestamp of vessel detectionmatching_score: Threshold for determining whether the S1 detection is matched to an AIS-broadcasting vessel. Whenmatching_score > 4.23E-5 OR IFNULL(matching_score_secondary, 0) > 0.05, we classify the detection as being matched to an AIS-broadcasting vessel.confidence: Confidence in the match between the S1 detection and any AIS vessel, indicating the level of ambiguity in the match. If there are multiple detections and AIS vessels very close to each other, the matching score of each possible pair would be similar and the confidence would be low. In contrast, if there’s only one possible pair, the confidence would be high no matter what the score is. In cases where there are multiple potentialssvidmatches for a singledetect_id, we select the match with the highest confidence.repeated_detections: It is crucial to remove repeated S1 detections that may represent fixed infrastructure rather than actual vessels. Previously, we identified repeated detections by analyzing object positions, focusing on those that occurred within 100 meters of each other at least three times over a six-month period. However, the updated dataset already includes information for each detection using a similar method.at_known_anchorages: Vessels in anchorages are excluded to receive the same treatment as those within the 1nm coastal buffer, for which S1 currently lacks the capacity to properly classify as vessels. Although there is potential for improvement, as discussed and tested in GitHub issue #11, the impact on overall emissions is minimal.- Additional filtering is applied to reduce noise and exclude detected objects that may not be vessels, using the variables
likely_ambiguities,likely_infrastructure,likely_vehicles_on_roads, andpotential_ice.
3.1.2.2 S1 classification of whether each detection is fishing or not
To determine whether each S1 vessel detection is a fishing vessel, we use GFW’s model, which classifies each detection and defines its likelihood of being a fishing vessel (Paolo et al. (2024)). The variables of interest from these tables world-fishing-827.pipe_sar_v1_published.fishing_pred_even_cal and world-fishing-827.pipe_sar_v1_published.fishing_pred_odd_cal are:
detect_id: Unique S1 vessel detection id.fishing_50: Fishing score of the vessel. If this value is > 0.5, then we classify the vessel type as fishing. Otherwise, we classify the vessel type as non-fishing.
3.1.3 Latest model updates
The current version of the Dark Emissions model has been updated significantly since the initial release in September. The most notable improvements were:
New S1 dataset with better noise and erroneous detections removed. The new dataset contains less noise and incorporates features that allow for more reliable filtering of detections. Further, the matching score threshold for determining whether the S1 detection is matched to an AIS-broadcasting vessel has been updated to 2.84e-5. These updates on the S1 dataset increases our confidence in identifying vessels and preventing the assignment of emissions to non-vessel objects.
Correcting inferred vessel lengths for small vessels: Inferred length from S1 detection have been corrected for small vessels (< 22 meters) using the relationship between known length and inferred length based on years of confident matches.
Temporal extrapolation unit: Previously, we were extrapolating emissions using a time scale of one year. The current model version now calculates emissions on a monthly basis.
Vessels size bins: The initial model version used independent quantiles for each AIS and S1 dataset, grouping vessel sizes into two categories. In the current update, size bins are determined based on unmatched detected lengths from S1, with the same cutoffs applied to both AIS-based vessel characteristics and S1-detected vessel lengths. Additionally, the number of bins was increased, enabling more accurate assignment of emissions by vessel length. This adjustment ensured emissions were allocated appropriately, and led to an overall increase in the emissions estimate. Following a sensitivity analysis, we determined that dividing vessels into 10 size groups is optimal for better capturing emissions, as emissions by number of bins stabilized at this number.
The justification for using a discrete size categorization rather than a continuous approach lies in the nature of the AIS daily gridded emissions data, from which dark emissions are extrapolated. The pixel-level data include not only emissions associated with vessel-specific details but also trip characteristics, such as navigation time and distance within a specific grid cell on a given day. Thus, when extrapolating emissions from AIS to S1, we must account not only for vessel characteristics (e.g., whether it is a fishing or non-fishing vessel and its length) but also trip characteristics (e.g., time and distance contributing to those emissions) across all vessels within a size group. This level of detail cannot be directly inferred from S1 images alone. Therefore, using length bins allows us to capture these aggregated trip dynamics and emissions patterns more effectively.
KNN ratios: Previously, emissions from outside the S1 footprint were assigned using a global ratio by month and vessel group, based on the total number of dark vessel detections and the total number matched to AIS-broadcasting vessels across all pixels. While this approach captured some temporal variability, it overlooked significant spatial variability, as detection rates vary across different seas and distances from shore. Moreover, global ratios gave equal weight to pixels near the coast, which typically exhibit higher ratios, and those at the outer edges of the S1 footprint, more representative of open-ocean detections. This misrepresentation within global ratios falsely inflated emissions estimates.
To address this, we applied a KNN approach, assigning ratios based on the closest pixels with available data. This enhances the representativeness of emissions estimates and better captures spatial heterogeneity outside the S1 footprint. The same approach was used to fill gaps within the footprint.
3.1.4 Areas of potential model refinement
We have identified several areas for potential model refinement:
Ratios outside the footprint: The KNN approach for assigning ratios outperforms the use of global ratios and other tested methods, such as applying ratios from global trends on distance from shore. However, assigning outside-footprint ratios still presents certain challenges, leaving room for further improvement.
When analyzing the spatial distribution of KNN ratios, sharp changes in emissions can be observed at the footprint boundary. The outermost pixels inside the footprint often show smaller emissions compared to the immediate pixels just outside it. This discrepancy arises because ratio values for pixels outside the footprint rely on multiple neighbors, some of which are not immediately adjacent but come from farther, denser coastal areas. This can inflate ratios just outside the footprint compared to adjacent pixels within it, leading to an overestimation of emissions in open ocean areas—but to a lesser extent than previous approaches. While the aggregated AIS-based emissions and their spatial distribution help to balance and smooth dark emissions at this interface, there remains significant room for improvement to prevent this.
Additionally, unexpected spatial patterns in the dark emissions distribution, which do not align with expected changes in emissions, can be observed. This issue primarily stems from the ratio distribution of specific vessel groups, especially those overrepresented in AIS data (e.g., non-fishing larger vessels). For such groups, detections are unevenly distributed within the footprint, creating concentrated areas in some regions while leaving others empty. When calculating KNN ratios, these differences propagate into open ocean regions, resulting in high-contrast areas with sharp divisions. This issue becomes even more pronounced at the monthly level, where detections for even highly represented vessel groups within S1 detections are more heterogeneously scattered. These inconsistencies distort the spatial variability of dark emissions, making it less representative of the patterns we would expect in reality. Focal statistic techniques on ratios distribution have been applied to address the most evident anomalies, smoothering ratio estimates. However, this remains the aspect with the highest potential for improvement within the current model framework.
The natural evolution of the outside-footprint problem would involve employing more sophisticated methods, such as Kriging extrapolation or Machine Learning approaches. With the later we could assign ratios based not only on spatial proximity but also on additional factors, including temporal proximity, by oceanic basin, vessel characteristics, distance to shore, distance to ports, traffic routes, etc.
Model performance: Given the novelty of these methods, there is a need for alternative approaches to assessing model performance, as no comprehensive datasets exist for this purpose.
Currently, when comparing validation results for AIS and dark emissions estimates, it is important to consider that these validations were conducted using different datasets, and each method’s validation should be interpreted only within the context of its specific approach. For AIS estimates, we used the EU emissions dataset, whereas dark emissions estimates were validated against known AIS emissions for matching vessels in the dark dataset. This means the dark model’s performance is assessed relative to AIS emissions, validating the method rather than the emissions data itself. Thus validation metrics across models are not comparable. While simulated dark emissions performance is strong, any limitations in the AIS model inevitably affect the dark emissions estimates. These limitations are not captured in the dark validation metrics, which may lead to the impression that the dark model outperforms the AIS model.
Sentinel-2 data: Currently, we use S1 SAR data for our dark fleet emissions estimates. We tested extending this methodology to also incorporate Sentinel-2 (S2) optical imagery data. However, differences in detection capacity, coverage and noise, prevent a direct integration of S2 with S1. Future work should focus on combining both datasets to complement each other’s strengths.
3.2 Results
Below, we present the new results derived from incorporating the updates discussed above. Overall, emissions have increased compared to the previous version. For instance, dark emissions for CO2 now account for 12.2% of total emissions in 2024, compared to ~7% in the September delivery. The updates also enhance resolution by vessel category and provide higher spatial variability. The following sections explores these updated results.
3.2.1 Ratio of the dark fleet to the AIS-broadcasting fleet
First, we visualize the ratio of the dark fleet to the AIS-broadcasting fleet observed within the S1 spatiotemporal footprint (Figure 3.3). This ratio serves as the basis for extrapolating AIS-based emissions to dark fleet emissions. We can look at how the ratio changes over years, disaggregated by vessel type and size class. For each year, vessel type, and size class, this ratio represents the aggregate month ratio within the S1 spatial footprint (i.e., the total number of unmatched dark vessel detections to the total number of matched AIS-broadcasting detections). Matching intuition, fishing vessels tend to have a higher ratio of dark to AIS vessels than non-fishing vessels, and smaller vessels tend to have a higher ratio than larger vessels (Table 3.1).
With the new dataset and updates, we are observing lower ratios than those detected previously.
Another way to visualize these data is to look at the fraction of all detected vessels that are matched to AIS (i.e., the number of AIS-broadcasting detections divided by the total number of detections across the AIS-broadcast and dark fleets) (Figure 3.4). This is the metric reported by Paolo et al. (2024). Matching intuition, non-fishing vessels tend to have a higher fraction of S1 detections matched to AIS than fishing vessels, and larger vessels tend to have a higher fraction of detections matched to AIS than smaller vessels (Table 3.1). While the relative patterns align—non-fishing vessels are tracked at higher rates than fishing vessels—our estimates are higher compared to those of Paolo et al. (2024), who analyzed data from 2017-2021 and estimated that 70-79% of non-fishing vessels and 24-28% of fishing vessels are tracked with AIS.
| fleet | ratio_dark_to_ais_detections | fraction_tracked |
|---|---|---|
| Fishing | 0.614 | 0.620 |
| Non-fishing | 0.182 | 0.846 |
3.2.2 Emissions
Next, we look at the total annual global emissions for each of our pollutants over time (see Figure 3.5, Table 3.2), disaggregated by the AIS-broadcasting fleet and the dark fleet as detected by S1. We observe that dark emissions are higher than previously estimated due to improved assignment of emissions by vessel size Figure 3.1. While the use of an increased number of length bins led to a rise in dark emissions, updates to global ratios using instead the new KNN approach helped smooth this increase, which would have otherwise been substantially greater. PM10, PM2.5 and VOCs were estimated using OceanMind methods, as detailed in this section. Therefore, PM emissions estimates cannot be directly correlated with our results.
To summarize some additional key findings:
Industrial vessels emitted approximately 1.3 billion tons of CO2 in 2024, accounting for about 3.6% of global fossil fuel emissions ((IEA) (2023)).
Maritime emissions are likely growing at more than twice the rate of global CO2 emissions. Between 2016 and 2024, global fossil fuel emissions increased by 7% ((IEA) (2023)), while emissions from maritime vessels grew by 33%.
While 4% of maritime emissions in 2024 came from ‘high-information’ vessels that broadcast their positions via AIS and have IMO numbers, 83%% came from ‘low-information’ AIS-broadcasting vessels without IMO numbers, and 13% came from dark fleet vessels that do not broadcast AIS. Understanding low-information and dark fleet activity is crucial for estimating total emissions and tracking changes over time. Emissions from dark vessels decreased by 22%% between 2016 and 2024, whereas emissions from AIS-tracked vessels rose by 50%% over the same time period. This shift is largely due to increased AIS adoption and advancements in AIS technology.
From 2016 to 2024, increases in emissions from AIS-tracked vessels were primarily concentrated in the Pacific waters of East and Southeast Asia, which accounted for nearly half of the overall increase. Similarly, decreases in dark fleet emissions were also concentrated in this region, contributing to over half of the total reduction.
Our analysis also revealed seasonal and event-driven variations in emissions. Emissions consistently dipped during major holidays, including Christmas, New Year’s Day, and Chinese New Year, and also fell during the COVID-19 pandemic. Furthermore, our analysis also shows changes in the geographical distribution of emissions that are likely driven by new policies and infrastructure development.
Our AIS-based emissions estimates align closely with other published sources: we estimate approximately 1.3 B MT CO2 in 2024, whereas the latest IMO estimates are 1.06 B MT; the latest EDGAR estimates are 910 MMT; and the latest OECD estimates are 862 MMT. Other published estimates do not include dark fleet vessels, making GFW’s dark fleet estimates the first of their kind.
| year | Fleet | CO2 | NOX | SOX | CH4 | CO | N2O | PM | PM10 | PM2_5 | VOCS |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 2016 | Dark | 207,106,823 | 3,162,065 | 1,278,795 | 3,087 | 155,138 | 10,103 | 174,168 | 383,760 | 352,847 | 233,853 |
| 2017 | Dark | 228,301,101 | 3,037,970 | 1,406,507 | 3,004 | 153,159 | 11,038 | 177,384 | 420,214 | 386,364 | 253,025 |
| 2018 | Dark | 232,864,473 | 3,018,395 | 1,434,046 | 2,997 | 153,267 | 11,243 | 178,367 | 428,279 | 393,780 | 257,507 |
| 2019 | Dark | 242,076,693 | 3,121,568 | 1,490,599 | 3,128 | 159,308 | 11,702 | 185,288 | 446,304 | 410,353 | 269,549 |
| 2020 | Dark | 239,402,747 | 3,012,164 | 294,728 | 3,006 | 154,240 | 11,543 | 37,182 | 90,657 | 83,354 | 264,352 |
| 2021 | Dark | 223,812,978 | 2,831,148 | 275,550 | 2,838 | 145,275 | 10,804 | 34,907 | 84,904 | 78,065 | 248,378 |
| 2022 | Dark | 174,624,702 | 2,179,706 | 214,942 | 2,212 | 113,533 | 8,438 | 27,110 | 66,403 | 61,054 | 195,037 |
| 2023 | Dark | 156,400,187 | 1,997,715 | 192,558 | 2,062 | 104,966 | 7,593 | 24,707 | 59,872 | 55,050 | 177,946 |
| 2024 | Dark | 162,523,976 | 2,010,488 | 200,015 | 2,060 | 105,629 | 7,861 | 25,163 | 61,927 | 56,938 | 182,539 |
| 2016 | AIS | 780,889,103 | 13,573,763 | 4,833,909 | 12,720 | 635,914 | 38,215 | 704,985 | 1,442,282 | 1,326,103 | 874,026 |
| 2017 | AIS | 846,501,499 | 14,529,395 | 5,238,889 | 13,572 | 680,694 | 41,352 | 757,432 | 1,560,024 | 1,434,361 | 941,607 |
| 2018 | AIS | 876,809,809 | 14,897,047 | 5,425,214 | 14,011 | 701,399 | 42,854 | 780,849 | 1,618,553 | 1,488,175 | 979,887 |
| 2019 | AIS | 899,209,924 | 15,139,576 | 5,562,484 | 14,423 | 718,367 | 44,028 | 798,821 | 1,666,298 | 1,532,073 | 1,015,842 |
| 2020 | AIS | 889,302,629 | 14,816,298 | 1,099,983 | 14,223 | 706,586 | 43,569 | 161,981 | 340,080 | 312,686 | 1,009,692 |
| 2021 | AIS | 965,566,238 | 15,989,880 | 1,194,266 | 15,130 | 757,245 | 47,143 | 174,550 | 367,212 | 337,632 | 1,079,482 |
| 2022 | AIS | 1,038,855,368 | 17,138,169 | 1,284,752 | 16,388 | 817,252 | 50,812 | 187,859 | 396,442 | 364,507 | 1,172,586 |
| 2023 | AIS | 1,085,212,897 | 18,017,603 | 1,342,122 | 17,532 | 866,996 | 53,299 | 197,934 | 416,908 | 383,325 | 1,247,819 |
| 2024 | AIS | 1,167,605,666 | 19,156,424 | 1,443,761 | 18,503 | 920,275 | 57,199 | 210,961 | 446,935 | 410,933 | 1,329,062 |
The same results, disaggregated by month, show a similar trend with total emissions increasing over the years (see Figure 3.6).
Displaying a single pollutant, such as CO2 (see Figure 3.7), we can clearly see an increase in total emissions over the years, and certain seasonality in the total emissions. Furthermore, Figure 3.7 and Figure 3.8 clearly depict the increase in AIS emissions versus a slight decrease in dark emissions, potentially linked to an improved in AIS coverage over the years. Despite dark emissions being derived from AIS emissions, the model also captures independent temporal variability in dark emissions, as evidenced by distinct seasonal trends.
We can also explore how dark emissions are distributed by vessel type and across the time range. Figure 3.9 shows that emissions from fishing vessels, and the relationship between small and large fishing vessels, have been more consistent over the years, despite a significant decrease in the largest fraction since 2016, followed by stabilization and a slight increase in recent years.
For non-fishing vessels—which account for the majority of emissions (Table 3.3)—there are clearer changes in both the relationship between small and large vessels and their emissions over time. Notably, there has been an inversion in the proportion of emissions between small and large non-fishing vessels, with a more significant reduction in emissions from larger vessels—possibly due to higher AIS adoption of that group—than from smaller ones. In recent years, smaller vessels have accounted for a larger proportion of emissions.
| year | pollutant | dark | fishing | emissions_mt |
|---|---|---|---|---|
| 2016 | CO2 | FALSE | FALSE | 752904323 |
| 2016 | CO2 | FALSE | TRUE | 27984779 |
| 2016 | CO2 | TRUE | FALSE | 184340150 |
| 2016 | CO2 | TRUE | TRUE | 22766673 |
| 2017 | CO2 | FALSE | FALSE | 814659885 |
| 2017 | CO2 | FALSE | TRUE | 31841614 |
| 2017 | CO2 | TRUE | FALSE | 209931561 |
| 2017 | CO2 | TRUE | TRUE | 18369540 |
| 2018 | CO2 | FALSE | FALSE | 842226543 |
| 2018 | CO2 | FALSE | TRUE | 34583266 |
| 2018 | CO2 | TRUE | FALSE | 213898546 |
| 2018 | CO2 | TRUE | TRUE | 18965927 |
| 2019 | CO2 | FALSE | FALSE | 862223471 |
| 2019 | CO2 | FALSE | TRUE | 36986453 |
| 2019 | CO2 | TRUE | FALSE | 223252111 |
| 2019 | CO2 | TRUE | TRUE | 18824581 |
| 2020 | CO2 | FALSE | FALSE | 851150727 |
| 2020 | CO2 | FALSE | TRUE | 38151902 |
| 2020 | CO2 | TRUE | FALSE | 221494624 |
| 2020 | CO2 | TRUE | TRUE | 17908123 |
| 2021 | CO2 | FALSE | FALSE | 924270671 |
| 2021 | CO2 | FALSE | TRUE | 41295567 |
| 2021 | CO2 | TRUE | FALSE | 204280746 |
| 2021 | CO2 | TRUE | TRUE | 19532232 |
| 2022 | CO2 | FALSE | FALSE | 984549768 |
| 2022 | CO2 | FALSE | TRUE | 54305601 |
| 2022 | CO2 | TRUE | FALSE | 152332426 |
| 2022 | CO2 | TRUE | TRUE | 22292277 |
| 2023 | CO2 | FALSE | FALSE | 1017446617 |
| 2023 | CO2 | FALSE | TRUE | 67766280 |
| 2023 | CO2 | TRUE | FALSE | 131771886 |
| 2023 | CO2 | TRUE | TRUE | 24628301 |
| 2024 | CO2 | FALSE | FALSE | 1094315980 |
| 2024 | CO2 | FALSE | TRUE | 73289686 |
| 2024 | CO2 | TRUE | FALSE | 139790601 |
| 2024 | CO2 | TRUE | TRUE | 22733375 |
3.2.3 Spatial maps of emissions
Next we can look at spatial maps of global emissions by the AIS-broadcasting fleet and by the dark fleet. For each 1x1 degree pixel (the spatial resolution of the dark fleet model), we aggregate emissions separately for each pollutant for 2024.
The spatial distribution of dark emissions heavily depends on how areas outside the S1 footprint are handled. While the updated KNN approach improves upon the use of global ratios by capturing greater temporal and spatial variability, there is still room for refinement in this area, as detailed in the section about potential model refinements. The assignment of ratios affects not only the spatial distribution of emissions in areas not covered by S1 but also the total emission estimates. When examining the origin of emissions—whether pixel ratios were directly assigned based on S1 detections or determined using KNN—it becomes evident that most dark emissions originate from areas not covered by S1. This is expected, as S1 primarily covers coastal areas, leaving the majority of the ocean unobserved. As shown in Table 3.4, approximately 65–73% of emissions are attributed to areas outside the footprint. On the contrary, when considering relative values by area, we see higher emission densities within the S1-covered zones, as most detections occur closer to the coast.
Given the significant contribution of dark emissions from pixels outside the footprint—and from those with null values within the footprint—it is important to highlight how these contribute to the uncertainty in the total estimates, as they rely entirely on extrapolated ratios. Therefore, among dark emissions, it is essential to differentiate between emissions inside and outside the S1 footprint and assess uncertainty independently. This distinction underscores the importance of improving methodologies for handling emissions outside the footprint, given their substantial influence on total estimates and the associated uncertainties, which must be minimized.
| year | s1_footprint | emissions_co2_dark_mt | emissions_by_pixel | percentage_of_total |
|---|---|---|---|---|
| 2016 | within | 58435805 | 387.10232 | 28.21530 |
| 2016 | outside | 148671018 | 103.15980 | 71.78470 |
| 2017 | within | 76538113 | 292.72907 | 33.52507 |
| 2017 | outside | 151762988 | 108.68923 | 66.47493 |
| 2018 | within | 71532379 | 255.85199 | 30.71846 |
| 2018 | outside | 161332094 | 113.44981 | 69.28154 |
| 2019 | within | 75650557 | 258.47709 | 31.25066 |
| 2019 | outside | 166426136 | 116.42449 | 68.74934 |
| 2020 | within | 67635879 | 231.34925 | 28.25192 |
| 2020 | outside | 171766867 | 121.34379 | 71.74808 |
| 2021 | within | 69709449 | 229.90333 | 31.14629 |
| 2021 | outside | 154103529 | 105.49523 | 68.85371 |
| 2022 | within | 48976586 | 184.10520 | 28.04677 |
| 2022 | outside | 125648116 | 79.40160 | 71.95323 |
| 2023 | within | 49906497 | 188.19289 | 31.90949 |
| 2023 | outside | 106493690 | 65.15400 | 68.09051 |
| 2024 | within | 46878684 | 171.21694 | 28.84416 |
| 2024 | outside | 115645293 | 69.23454 | 71.15584 |
3.2.4 Regional trends
We can also look at broader regional trends in emissions estimates by looking at Major FAO Regions. There are 19 major FAO regions (Figure 3.20).
For each region, we summarize emissions for both the AIS-broadcasting and dark fleets. Below we do this focused on summarizing annual emissions of CO2, and focusing on 2016 and 2024 (the first and last years of our dark fleet time series data) (Figure 3.21, Figure 3.22). Here we can see the dominance of Pacific waters of east and southeast Asia (encompassed by the FAO regions of ‘Pacific, Northwest’ and ‘Pacific, Western Central’) in terms of emissions. These two regions accounted for 35% of global AIS-broadcasting emissions in 2024, and 48% of global dark fleet emissions in 2024.
Next we look at the change in CO2 emissions between 2016 and 2024 for each of these regions (Figure 3.23). Here we can see the Pacific waters of east and southeast Asia (encompassed by the FAO regions of ‘Pacific, Northwest’ and ‘Pacific, Western Central’) account for 45% of the increase in AIS-broadcasting emissions from 2016 to 2024. These areas also account for the largest decrease in dark fleet emissions, a 85% of the global decrease in dark fleet emissions. Both of these trends are consistent with the increase in AIS coverage in these areas.
3.3 Model Validation
There is not an existing validation dataset for measured emissions from the dark fleet (i.e,. those vessels that do not use AIS). However, we can use AIS data to assess the likely performance of the S1 dark fleet emissions model as follows:
- Our unit of prediction for extrapolating AIS-based emissions to dark fleet emissions is monthly, 1x1 degrees, by vessel class (fishing or non-fishing), and for 10 different size classes.
- We use the actual observed S1 detections to summarize the overall fraction of detections matched to AIS, categorized by vessel type and size class. Based on these fractions, we randomly assign a portion of each vessel type and size class to the simulated “training” dataset, with the remaining vessels allocated to the simulated “testing” dataset. The training dataset vessels will represent AIS-broadcasting vessels, while the testing dataset will represent simulated dark fleet vessels. So for example, for non-fishing largest vessels, the observed data show that 93.8% of S1 detections were matched to AIS-broadcasting vessels. To generate our simulated ‘testing’ and ‘training’ datasets for non-fishing large vessels, we therefore randomly assign 93.8% of the non-fishing large vessels from the AIS dataset to be in the simulated ‘training’ dataset and 6.2% of these vessels to be in the simulated ‘testing’ dataset. We do the same for the other vessel types and size classes based on the fractions detailed in
| fishing | length_size_class_percentile | fraction_s1_detections_matched |
|---|---|---|
| Fishing | Length bin: 1 | 0.215 |
| Fishing | Length bin: 2 | 0.265 |
| Fishing | Length bin: 3 | 0.296 |
| Fishing | Length bin: 4 | 0.323 |
| Fishing | Length bin: 5 | 0.352 |
| Fishing | Length bin: 6 | 0.388 |
| Fishing | Length bin: 7 | 0.429 |
| Fishing | Length bin: 8 | 0.482 |
| Fishing | Length bin: 9 | 0.547 |
| Fishing | Length bin: 10 | 0.704 |
| Non-fishing | Length bin: 1 | 0.124 |
| Non-fishing | Length bin: 2 | 0.223 |
| Non-fishing | Length bin: 3 | 0.451 |
| Non-fishing | Length bin: 4 | 0.605 |
| Non-fishing | Length bin: 5 | 0.727 |
| Non-fishing | Length bin: 6 | 0.788 |
| Non-fishing | Length bin: 7 | 0.829 |
| Non-fishing | Length bin: 8 | 0.866 |
| Non-fishing | Length bin: 9 | 0.918 |
| Non-fishing | Length bin: 10 | 0.938 |
- For each month and pixel, calculate the total emissions from the training dataset vessels. This will represent the simulated total emissions from the AIS-broadcasting fleet. We focus on CO2. However, performance of all other pollutants will be similar since the emissions factors are all linear scalars of each other
- For each analysis unit, calculate the ratio of simulated dark fleet vessels to AIS-broadcasting vessels, by vessel type and size class. Do this only for months and pixels that are covered by S1. This will represent the simulated dark-to-AIS vessel ratio that we would observe using the actual S1 data with the spatiotemporal S1 footprint.
- Since S1 does not cover all pixels globally, we additionally need a method for estimating the dark-to-AIS ratio for non-covered areas. To do so, we simply use the ratio observed across all covered pixels for a given month, vessel type, and size class (e.g., the total number of detected dark vessels divided by the total number of detected AIS vessels, by vessel type and size class).
- Multiply the emissions from the simulated AIS fleet (from Step 3) by the ratio of simulated dark vessels to simulated AIS vessels (from Step 4 for pixels within the S1 spatiotemporal footprint, and from Step 5 for all other areas) to get the predicted amount of emissions from the dark fleet.
- We can also calculate the actual observed emissions from the simulated dark fleet vessels, since in reality these vessels broadcast AIS and we can therefore calculate emissions from them.
- For both the observed and predicted emissions estimates, aggregate monthly pixel-level emissions across vessel types and sizes.
- For each month and pixel, we therefore have the predicted emissions from the simulated dark fleet and the observed emissions from the simulated dark fleet. We can then calculate the performance metrics for the model.
- We can calculate overall performance using all vessels (i.e., summing emissions across all vessels in each analysis unit), as well as disaggregated by vessel type (Table 3.6).
With this, looking at a scatterplot of observed CO2 emissions of the simulated dark fleet vessels vs estimated CO2 emissions for the simulated dark fleet vessels (Figure 3.24), summing emissions across all vessel sizes for each analysis unit and vessell type. Each point represents the emissions for a 1x1 degree monthly pixel, disaggregated by vessel type. Since there are many points across the entire time series. Visually inspecting this, it appears that the estimate emissions track closely to observed emissions for the simulated dark fleet vessels.
Using this approach, we examine the resulting scatterplot of observed CO2 emissions of the simulated dark fleet versus estimated CO₂ emissions for the simulated dark fleet vessels (Figure 3.24). Emissions are summed across all vessel sizes for each analysis unit and vessel type. Each point on the plot represents the emissions for a 1x1-degree monthly pixel. A visual inspection suggests that the estimated emissions closely align with the observed emissions for the simulated dark fleet vessels.
We can also evaluate the model’s performance metrics. From Table Table 3.6, it is evident that the model performs well across all metrics, regardless of whether the results are disaggregated by vessel type or not.
| .metric | .estimate | fishing | length_size_class | in_s1_footprint |
|---|---|---|---|---|
| rmse | 525.889 | All vessels | All sizes | All pixels |
| nrmse | 0.092 | All vessels | All sizes | All pixels |
| rsq | 0.992 | All vessels | All sizes | All pixels |
| rsq_trad | 0.992 | All vessels | All sizes | All pixels |
| rmse | 559.997 | FALSE | All sizes | All pixels |
| rmse | 89.638 | TRUE | All sizes | All pixels |
| nrmse | 0.095 | FALSE | All sizes | All pixels |
| nrmse | 0.085 | TRUE | All sizes | All pixels |
| rsq | 0.991 | FALSE | All sizes | All pixels |
| rsq | 0.993 | TRUE | All sizes | All pixels |
| rsq_trad | 0.991 | FALSE | All sizes | All pixels |
| rsq_trad | 0.993 | TRUE | All sizes | All pixels |