From the Cover: Satellites can reveal global extent of forced labor in the world’s fishing fleet

While forced labor in the world’s fishing fleet has been widely documented, its extent remains unknown. No methods previously existed for remotely identifying individual fishing vessels potentially engaged in these abuses on a global scale. By combining expertise from human rights practitioners and satellite vessel monitoring data, we show that vessels reported to use forced labor behave in systematically different ways from other vessels. We exploit this insight by using machine learning to identify high-risk vessels from among 16,000 industrial longliner, squid jigger, and trawler fishing vessels. Our model reveals that between 14% and 26% of vessels were high-risk, and also reveals patterns of where these vessels fished and which ports they visited. Between 57,000 and 100,000 individuals worked on these vessels, many of whom may have been forced labor victims. This information provides unprecedented opportunities for novel interventions to combat this humanitarian tragedy. More broadly, this research demonstrates a proof of concept for using remote sensing to detect forced labor abuses.

Forced labor in fisheries, a type of modern slavery, is increasingly recognized as a human rights crisis. The International Labor Organization (ILO) defines forced labor as “all work or service which is exacted from any person under the menace of any penalty and for which the said person has not offered himself voluntarily” (1). The ILO provides a framework of 11 forced labor risk indicators (2) that have all been documented within the fisheries sector, including indicators representative of debt-bonded labor, as well as indicators representative of servitude or slave labor such as abusive working and living conditions. In 2015, reports emerged on forced labor in Thai fisheries (3) and the role of forced labor in producing seafood imported to the United States (4). More recent reports have described the global nature of the problem (5), and there has been a call to integrate social responsibility into ocean science (6). Despite widespread condemnation and ambitious commitments, forced labor remains poorly understood in the fisheries sector. Here we show that recently available high-frequency vessel monitoring of the global industrial fishing fleet can shed new light on forced labor at a much finer resolution. We combine expertise from on-the-ground human rights practitioners and satellite vessel monitoring data for over 16,000 industrial fishing vessels to estimate 1) the number of high-risk vessels and the number of crew who may be victims working on those vessels, 2) where these vessels fish, and 3) what ports these vessels visit. This information can inform new market, policy, and enforcement interventions to combat forced labor in global fisheries. This research more generally demonstrates how remote sensing can detect forced labor abuses by observing dynamic behavior.Current estimates of forced labor in fisheries are coarse and are based on country-level statistics. Using country-level household surveys, the ILO estimated that 16 million people were victims of forced labor in 2016, with 11% of these in agriculture, forestry, or fisheries (7). The Global Slavery Index reports that the seven countries with highest slavery risk in 2018 generated 39% of global fisheries catch (3, 8), and Tickler et al. found that the United States has slavery risks of 0.2 kg per metric ton for domestic seafood and 3.1 kg per metric ton for imported seafood (9). While these studies are important for broadly understanding which countries have risk, current methods are unable to detect this problem at the level of individual fishing vessels, which will be essential for targeted interventions.We empirically examine whether vessels reported to exhibit any of the ILO indicators of forced labor behave in ways that are systematically different from other vessels, and then exploit this information using machine learning to discriminate between vessels that use forced labor from those that do not. We do so by measuring a suite of features that can be observed using satellite Automatic Identification System (AIS) vessel monitoring data made available by Global Fishing Watch (GFW) (10). There may be many behavioral correlates with forced labor that could help to differentiate between high-risk and low-risk vessels. To determine which model features to include, we first conducted a literature review of investigative journalism reports and looked for instances of forced labor case accounts that detailed specific behaviors that could be observed using vessel monitoring data. We next conducted informal phone interviews with experts from several nongovernmental organizations (NGOs) working in this field, during which we asked interviewees what observable vessel behaviors they would look for if they wanted to identify suspicious activity. The machine-learning approach we use does not assume that vessels behave in any particular way; rather, it merely uses the features identified by literature review and expert insight to exploit any observed empirical differences between vessels that use forced labor and other vessels. NGO experts and investigative journalism suggest that gaps in AIS transmission, port avoidance, transshipment, and extended time at sea may indicate the presence of forced labor (11). Certain features, like information on catch and the species being targeted, could also be helpful in discriminating between high- and low-risk behavior by providing more context on the fishing taking place. However, these data are not currently available at the vessel level on a global scale. Data on recruitment practices and vessel ownership and information on from where the crew originates could also be helpful, but, again, these data are not widely available. We arrived at a list of 27 vessel behavior and characteristic features for which we have globally available data at the vessel level (SI Appendix, Table S1, and SI Appendix).To build a predictive model for identifying high-risk vessels, we developed a training dataset that includes the behavior and characteristics of known forced-labor vessels, as well as the behavior and characteristics of other vessels. We compiled a comprehensive database of vessels that were reported to display one or more of the ILO forced labor indicators (2); these vessels are labeled as “positives.” We do not, however, know which vessels do not use forced labor (“negatives”). Rather, any vessel that we do not label as positive is “unlabeled,” and may in fact be a positive vessel that has not yet been identified or may truly be a negative vessel. This is an example of “positive-unlabeled (PU)” learning, a less straightforward problem than traditional supervised machine learning (12). We use PU learning to predict whether or not 16,261 longliner, trawler, and squid jigger fishing vessels were high-risk during each year they operated between 2012 and 2018 (“vessel-years”). We focus on this subset of vessels because they broadcasted sufficient and reliable AIS positions and because these are the only fishing gear types with documented cases of forced labor aboard vessels that broadcasted sufficient AIS data. These vessels represent 33% of the total time at sea spent by all fishing vessels operating in this time period tracked by GFW. Our PU approach leverages information from all positively labeled vessels (n = 22 unique vessels across 22 vessel-years using our baseline model assumption), but places less emphasis on unlabeled vessels given their uncertain nature (n = 16,257 unique vessels across 66,314 vessel-years using our baseline model assumption).