Revealing global risks of labor abuse and illegal, unreported, and unregulated fishing
This research complies with all relevant ethical regulations. The study protocol was evaluated and approved by the Stanford Institutional Review Board (Stanford IRB #49308).
Port risk assessment
Port risks associated with labor abuse and IUU fishing were evaluated through an online global survey in English, Spanish and French, deployed using the platform Qualtrics. Within the survey instrument, all participants were provided with information about the study and advised that their participation was entirely voluntary and anonymous (i.e. that responses could not be tracked to an individual). They were also advised how they could contact Stanford IRB with any concerns. Informed consent was documented by tick-box to ensure anonymity, because signature would have provided identifiable information. Participants were also advised of withdrawal procedures, and that they could opt out of taking the survey at any time by closing their browser.
Over 95 experts from seafood companies, research institutions, human rights organizations, and governments addressed questions in the survey that evaluated the risk levels for labor abuse and IUU fishing associated with ports, and answered questions about other known risk factors including vessel flag20,32,62, vessel gear type19,63, and time-at-sea22. Based on respondents’ selected countries of expertise, the survey provided a list of ports for those countries. For each port, respondents were instructed to indicate whether, based on their experience, the port was known for labor abuse or IUU fishing, and if so, their level of certainty about that association. Within the survey, we defined low certainty as the port being known to be poorly monitored or regulated, medium certainty as the port having suspected occurences of labor abuse or IUU fishing, and high certainty as the port having documented cases of labor abuse or IUU fishing. When the port was not associated with risk of either IUU fishing or labor abuse, respondents were asked to select not associated.
Port risk scores
To describe fishing grounds associated with risk of labor abuse and IUU fishing, we focused on the ports used by fishing vessels to likely land their catches (i.e., the arrival port). For each port, survey responses were converted to a value of −1 for not associated, and 1/3, 2/3 and 1 for low, medium and high certainty of association with labor abuse or IUU fishing, respectively. We assigned each fishing trip a port risk score by summing all responses for each arrival port (Supplementary Fig. 9), although we also explored departure port risk for labor abuse because movement of people into situations of labor abuse could be related to conditions at the departure port (Supplementary Fig. 2). Consequently, port risk scores took into account both the number of responses and the level of agreement among the respondents on their level of certainty for each port (Supplementary Fig. 9). Low scores indicate that several respondents evaluated the ports used during the fishing trip as not associated with labor abuse or IUU fishing risks, whereas high scores indicate high certainty among multiple respondents that the port is exposed to high risks of labor abuse or IUU fishing. We removed all data where respondents indicated two different levels of certainty when assessing one dimension of risk, e.g. selected both not associated and low for IUU fishing.
Modeling approach
The port risk score of the fishing trip was then linked to information about vessel movement and fishing activity from Global Fishing Watch (GFW) to model at-sea and transshipment risks. Data were extracted from GFW databases on September 15, 2021, and reflect updates to that date. From the onboard Automatic Identification System (AIS) data, GFW identified over 8.7 million fishing trips from 2012 to 2019 that include information on the vessel flag and fishing gear type. GFW also determines when fishing occurs on a trip and estimates the number fishing hours at each observed location through machine learning based on vessel movement patterns64. We matched the end coordinates of the fishing trips reported by GFW with the nearest port in our survey by searching within a 3-km radius from each port. We were then able to assign port risk scores for 1.8 million trips based on 412 ports for the labor abuse analysis and 3.0 million fishing trips based on 455 ports for the IUU fishing analysis.
Using these data, we developed a model to identify traits of fishing vessels that contribute to port risk. Predictors used in the model were flag groups (5 levels: Flag group 1, Flag group 2, Flag group 3, China, Other) (Supplementary Table 2), gear types (9 levels: set longline, drifting longline, squid jigger, pots and traps, trawlers, set gillnet, pole and line, trollers, purse seine), and time-at-sea (5 levels: less than 1 month, 1–3 months, 3–6 months, 6–12 months, 12 months and more). Categorization of gears is based on available data for training the models that GFW uses in their classification. Gears for which less data were available may have greater classification error.
These predictors were chosen because they represent known risk factors for labor abuse or IUU fishing19,20,22,32,62,63 and were vessel characteristics that were standardized and consistent within the GFW database. We categorized vessel flag groups based on Ford and Wilcox20, who used globally consistent indicators of foreign ownership (ratio of nationally flagged to nationally owned vessels), control of corruption, and fidelity to the flag state’s exclusive economic zone (EEZ), including territorial and archipelagic waters (i.e. remaining within the flag state’s EEZ), to better understand what aspects of vessel flag may be driving risk (Supplementary Table 2). Flag group 1 (6 states for this study) is characterized by a high ownership by countries other than the flag state (high foreign ownership ratio), high proportion of vessels operating outside their EEZ (Supplementary Table 2) and higher control of corruption. Flag group 2 (20 states for this study) is mainly distinguished by poor control of corruption, as well as low fidelity to the flag state EEZ, and intermediate levels of ownership by countries other than the flag state (Supplementary Table 2). Flag group 3 (91 states for this study) represents flag states with high ownership by the flag state, a high proportion of vessels operating within their flag state EEZ, and intermediate control of corruption (Supplementary Table 2). Although China was clustered in Flag group 3 by Ford and Wilcox20, we separated China as its own flag group because of its dominance in the data (41{aa306df364483ed8c06b6842f2b7c3ab56b70d0f5156cbd2df60de6b4288a84f} of 8.7 million fishing trips with all predictors), which could otherwise mask the predictive power of other flag groups. The flags that did not fall into the above three categories were grouped as ‘Other’ (39 states for this study) (Supplementary Table 2).
Port risk and cumulative fishing vessel and carrier hold capacities by port
To determine whether experts simply perceived port risk as a reflection of port size, we investigated the relationship between port risk score and cumulative fishing and carrier vessel hold capacity estimates for major ports65 (n = 99), as proxies for port size. We matched these ports with our data (39 ports for labor abuse risk and 45 ports for IUU fishing risk) and assigned port risk scores from our survey. We tested the correlation between port risk score and cumulative fishing vessel hold capacity and separately, for port risk score and carrier vessel hold capacity, using Kendall’s rank correlation.
At-sea risk model
To estimate at-sea risk, we created decision trees for labor abuse and IUU fishing, each trained with over 100 trees with gradient boosting66, with a maximum depth of 10, shrinkage of 0.05, a predictor subsampling rate of 0.6, and a root-mean-square error (RMSE) as a loss function. We evaluated the importance and effect of each predictor to the model through SHAP (SHapley Additive exPlanations) values30. Grounded in cooperative game theory67, the SHAP value measures the contribution of each predictor to the prediction with respect to a model baseline. In our case, it measures a change in the predicted port risk score attributed to each predictor for each fishing trip. Predictor importance was calculated as \(I_j=\frac1n\mathop\sum \nolimits_i=1^n|\phi _j^i|\), where j is the predictor, n is the number of observations, and \(\phi _j^i\) is the SHAP value of predictor j in observation i. Larger values indicate greater contribution to the model prediction. For each predictor with high contribution to the model, we estimated the main effect to the predicted port risk score by computing SHAP values when the predictor was present in the trip. We obtained the predicted port risk score solely attributed to the predictor by adding the main effect to the model baseline without accounting for interaction effects. To explore the potential interactions between predictors of interest, we also obtained the predicted port risk score by adding the main effect and the interactions to the model baseline. The model was fitted using ‘XGBoost’ 1.0.0 under Python 3.7.1. SHAP values were computed using ‘shap’ 0.35.0 under Python 3.7.1.
We summarized spatial risk at-sea based on arrival port risk score. Based on the distribution of the observed risk score of both labor abuse and IUU fishing, port risk score was binned into three groups (low: risk score < 0, medium: 0 ≤ risk score < 2, high: risk score ≥ 2) using a univariate k-means method68 (Supplementary Fig. 9). Subsequently, each fishing trip was categorized into one of three classes based on the expert-assessed port risk group of the arrival ports. For fishing trips that did not have an expert-assessed port risk (7.0 million trips for labor abuse, 5.8 million trips for IUU fishing), we predicted the port risk score using the model and categorized these trips into one of three classes in the same way. For each class of fishing trips (observed and predicted combined), we obtained cumulative fishing hours over all fishing vessels in a 1 × 1-degree grid. Fishing hours were then scaled to the corresponding areas in km2. The maps identified the top 5{aa306df364483ed8c06b6842f2b7c3ab56b70d0f5156cbd2df60de6b4288a84f} of values and were created using bivariate Gaussian kernel density estimation, weighted by the scaled cumulative number of fishing hours in each grid, with a band width of five degrees.
Transshipment risk model
Transshipment risk areas were modeled using trips taken by carrier vessels that had encounters with fishing vessels during the course of their trip. We also used single-vessel loitering events69 as a predictor in the model. Two-vessel encounters are defined as two vessels remaining within 500 m of each other for longer than 2 h, traveling at less than 2 knots while at least 10 km from an anchorage69,70. Single-vessel loitering events occurred when carrier vessels stayed at least 20 nautical miles from shore, traveling at less than 2 knots for 8 h or more, vessel behavior consistent with transshipment, but with no other vessel observed through AIS in the immediate vicinity69. Loitering can be indicative of transshipment with a vessel that has an AIS transponder that has been turned off or the transshipment vessel waiting until its next task. Under these criteria, GFW listed 5,811 trips by carrier vessels that encountered at least one fishing vessel from 2012 to 2019. By matching AIS information of these trips with coordinates of expert-assessed ports in our survey, we assigned port risk scores to 3,229 trips based on 76 ports for labor abuse risk and 3,386 trips by carrier vessels based on 82 ports for IUU fishing risk.
Using these data, we developed decision tree models for labor abuse and IUU fishing to identify areas of higher transshipment risk by estimating arrival port risk of trips by carrier vessels. The models were each trained with over 300 trees with gradient boosting, with a maximum depth of 10, shrinkage of 0.05, a subsampling rate of 0.6, and RMSE as a loss function. Predictors in the model were flag type of a carrier vessel, time-at-sea, flag type of encountered fishing vessels and their gear types, and the occurrence of a loitering event. Flag type (of both the transshipping fishing vessel and carrier vessel), gear type of the transshipping fishing vessel, and time-at-sea were categorized in the same way as in the model of at-sea risk and converted to numeric with one hot encoding. Predictor importance was evaluated using SHAP values.
We visualized spatial risk of transshipment using port risk score in a similar way to at-sea risk. Specifically, we assigned the port risk class of the trip to the coordinates of an encounter event during the trip. When trips by carrier vessels did not have an expert-assessed port risk (2,582 trips for labor abuse and 2,425 trips for IUU fishing), we predicted port risk score using the model and categorized these trips into one of three classes. For each class, we summed the number of all encounter events in a 1 × 1-degree grid and scaled these to the corresponding areas in km2 and presented the top 5{aa306df364483ed8c06b6842f2b7c3ab56b70d0f5156cbd2df60de6b4288a84f} of values. The maps were created using bivariate Gaussian kernel density estimation, weighted by the scaled density of transshipment encounter events in each grid, with a band width of five degrees.
Model performance and robustness analyses
We evaluated the performance of our models using root-mean-square error (RMSE) with 10 × 5-fold cross-validations (Fig. 2). We also evaluated model performance using Matthews Correlation Coefficient (MCC) by 10 × 5-fold cross-validations after categorizing predicted port risk score into three risk classes (Fig. 2).
To evaluate the robustness of our modeling results, we simulated cases where we had received fewer responses from experts in the port risk survey by randomly dropping 10{aa306df364483ed8c06b6842f2b7c3ab56b70d0f5156cbd2df60de6b4288a84f} or 20{aa306df364483ed8c06b6842f2b7c3ab56b70d0f5156cbd2df60de6b4288a84f} of the experts’ responses from the survey and re-assigning a port risk score to each port. Using the subsampled data, we re-fit the model with the same hyperparameters and calculated predictor importance and effects on port risk score through SHAP values. The procedure was repeated 10 times for 10{aa306df364483ed8c06b6842f2b7c3ab56b70d0f5156cbd2df60de6b4288a84f} and 20{aa306df364483ed8c06b6842f2b7c3ab56b70d0f5156cbd2df60de6b4288a84f} drops (Supplementary Fig. 4; Supplementary Note 1). The similarity of predictor importance and effects among subsamples was measured as intraclass correlation (ICC), and agreement was tested using F-test (Supplementary Note 1). We also assessed agreement among experts using Shannon entropy and found that it was medium to high among experts assessing risk at port (Supplementary Fig. 9), indicating that our survey captured shared perceptions of port risk.
Port stop duration
We obtained the duration of port stops by fishing vessels using the database curated by GFW. From AIS signals, GFW records the location and time of three events when vessels are at port: (1) the beginning of a port stop event, defined as when a vessel travels at less than 0.2 knots while in port (within 4 km of an anchorage point), (2) the ending of a port stop event, defined as when a vessel travels at greater than 0.5 knots while in port, and (3) port gap, when a gap between AIS signals is greater than four hours while in port. We calculated the duration of a port stop as the time difference between the beginning and ending of the port stop event that took place at dock. We removed port stop events where the beginning of a port stop event did not immediately follow the port entry or the ending of the previous port stop, or where the ending of a port stop event was not immediately followed by port exit or the beginning of the next port stop. We merged two consecutive port stops when they took place at the same anchorage point and the gap was less than 30 minutes. We also removed port stops of less than one hour, which may be vessels moving slowly at or between anchorage points. After removing potentially spurious events, we identified 6 million vessel port stop events between 2012 and 2019, among which were over 183,000 visits by foreign vessels that we used in the analysis.
We then compared port stop duration of foreign fishing vessels across flag groups. In a linear mixed-effects model, we specified logarithm of port stop duration (hours) as a response variable and flag group as an explanatory variable. We also included port state and fishing gear type as random effects to account for potential differences in port infrastructure between countries and known differences in unloading and fueling times for different vessel gear types due to the vessel and catch size. We used a similar approach to model port stop duration by gear type using port states as a random effect. To be consistent with the risk mapping, we limited the analysis to the vessels with specified flag states and gear types in the GFW database. The model was run using ‘lme4’ ver. 1.1–2671 in R 4.0.4, and the confidence intervals were estimated using 1000 bootstraps.
Risk relationships
To understand correlations between labor abuse and IUU fishing risks, we used Kendall’s rank correlation. We hypothesized that risk may change with transshipment, so we explored whether there were differences in the risk of ports that fishing vessels visited, between those that did and did not transship using χ2 test of independence.
Regional differences
To explore regional differences in fishing patterns in high-risk areas, we selected 6 high-risk area subsets for labor abuse and IUU fishing: Southwest Atlantic, Humboldt Current, around the Maldives (for labor abuse) or the Western Indian Ocean (for IUU fishing), West Africa, around the Azores, and around the Galapagos Islands (Supplementary Fig. 5). We identified the proportion of flag states and gear types in these high-risk regions for both labor abuse and IUU fishing based on fishing hours (Supplementary Tables 7, 8, 11, and 12). We also identified the top destination ports in the identified high-risk regions based on number of trips (Supplementary Tables 9 and 10).
Port State Measures analysis
The Port State Measures Agreement (PSMA) focuses on due diligence around foreign-flagged fishing and support vessels (e.g., carrier and bunker vessels). Territories often maintain separate flag identities from their sovereign states (e.g., Greenland and Denmark, Anguilla and the United Kingdom, etc.). For this analysis, we consider these flags to be ‘domestic’ with respect to the sovereign state, because many sovereign states enter into PSMA on behalf of their territories72. As ‘domestic’ vessels, they are not subject to the additional measures stipulated by PSMA. Because of the size of the fleets involved, and the legal status of Chinese Taipei with respect to PSMA, we treat vessels flagged to the fishing entity of Taiwan and Chinese-flagged vessels as separate entities13 (10.5{aa306df364483ed8c06b6842f2b7c3ab56b70d0f5156cbd2df60de6b4288a84f} of visits flagged to the fishing entity of Taiwan were observed in mainland China, and 0.2{aa306df364483ed8c06b6842f2b7c3ab56b70d0f5156cbd2df60de6b4288a84f} of visits by Chinese-flagged vessels were observed in the Chinese Taipei EEZ in 2012–2019).
To evaluate the initial impact of PSMA ratification and entry into force in 2016 on vessel dynamics, we analyzed how the number of vessel visits to each port state changed between 2015 (pre-PSMA) and 2017 (post entry into force) by 65 port states72. In the analysis, we used port states that constitute the top 95{aa306df364483ed8c06b6842f2b7c3ab56b70d0f5156cbd2df60de6b4288a84f} of the cumulative number of port visits to remove port states that received few visits by foreign fishing vessels for each flag group. We also excluded port states that ratified PSMA in 2017 from the analysis to keep the calendar comparisons consistent (i.e., Albania, Denmark on behalf of Greenland and the Faroe Islands, Djibouti, Japan, Kenya, Madagascar, Maldives, Mauritania, Montenegro, Namibia, Peru, Senegal, Togo). For each port state, we counted the number of visits by ‘foreign’ fishing vessels (i.e., vessels flagged to states other than the port state or its territories) for each flag group. For consistency, we limited the analysis to the vessels we used in the risk mapping analysis, and to vessels that were found to be active both in 2015 and 2017 in the GFW database.
We performed Bayesian hierarchical models under the before-after-control-impact design with a log-normal error distribution. In each model, we specified the number of port visits by foreign vessels as a response variable, sampling year (2 levels: 2015 and 2017), PSMA ratification in 2016 (2 levels: yes or no) and the interaction as explanatory variables, and port state as a random intercept. The response variable was linearly scaled by dividing by the maximum value in each model. The coefficient of the interaction indicates the proportional change of visits by foreign vessels to PSMA ratifiers between 2015 and 2017, after controlling for the change in non-PSMA ratifiers. For priors, we used a normal distribution with μ = 0 and σ = 10 for the mean of each fixed effect and a half-Cauchy distribution with β = 5 for the error term. For the random intercept, we used a normal distribution with μ = 0 and σ specified as a half-Cauchy distribution with β = 5. We sampled 2 chains of 5000 samples with 2000 burn-ins each. The analysis was performed using PyMC3 ver. 3.11.273 in Python 3.7.8.
Reporting summary
Further information on research design is available in the Nature Research Reporting Summary linked to this article.