Predicting Illegal Fishing on the Patagonia Shelf from Oceanographic Seascapes

Abstract

Many of the world's most important fisheries are experiencing increases in illegal fishing, undermining efforts to sustainably conserve and manage fish stocks. A major challenge to ending illegal, unreported, and unregulated (IUU) fishing is improving our ability to identify whether a vessel is fishing illegally and where illegal fishing is likely to occur in the ocean. However, monitoring the oceans is costly, time-consuming, and logistically challenging for maritime authorities to patrol. To address this problem, we use vessel tracking data and machine learning to predict illegal fishing on the Patagonian Shelf, one of the world's most productive regions for fisheries. Specifically, we focus on Chinese fishing vessels, which have consistently fished illegally in this region. We combine vessel location data with oceanographic seascapes -- classes of oceanic areas based on oceanographic variables -- as well as other remotely sensed oceanographic variables to train a series of machine learning models of varying levels of complexity. These models are able to predict whether a Chinese vessel is operating illegally with 69-96% confidence, depending on the year and predictor variables used. These results offer a promising step towards preempting illegal activities, rather than reacting to them forensically.

Full text

Predicting Illegal Fishing on the Patagonia Shelf from
Oceanographic Seascapes
A. John Woodill, Maria Kavanaugh, Michael Harte, and James R. Watson†
July 13, 2020
Abstract
Many of the world’s most important fisheries are experiencing increases in illegal fishing, under-
mining efforts to sustainably conserve and manage fish stocks. A major challenge to ending illegal,
unreported, and unregulated (IUU) fishing is improving our ability to identify whether a vessel is
fishing illegally and where illegal fishing is likely to occur in the ocean. However, monitoring the
oceans is costly, time-consuming, and logistically challenging for maritime authorities to patrol. To
address this problem, we use vessel tracking data and machine learning to predict illegal fishing
on the Patagonian Shelf, one of the world’s most productive regions for fisheries. Specifically, we
focus on Chinese fishing vessels, which have consistently fished illegally in this region. We combine
vessel location data with oceanographic seascapes – classes of oceanic areas based on oceanographic
variables – as well as other remotely sensed oceanographic variables to train a series of machine
learning models of varying levels of complexity. These models are able to predict whether a Chinese
vessel is operating illegally with 69-96% confidence, depending on the year and predictor variables
used. These results offer a promising step towards preempting illegal activities, rather than reacting
to them forensically.
Keywords: geospatial intelligence; illegal, unreported, and unregulated fishing; IUU; machine
learning; prediction; seascapes; sustainable fisheries.
†Correspondence: A. John Woodill, College of Earth, Ocean, and Atmospheric Sciences, Oregon State University,
johnwoodill@gmail.com. We would like to thank Global Fish Watch for providing the vessel location data for our
analysis. This work was supported by NASA under the award "Securing Sustainable Seas: Near real-time monitoring
and predicting of global fishing fleet behavior" (Award No. 80NSSC19K0203). Additional support for developing
seascapes was provided from NASA under the award " Dynamic Seascapes to Support a Biogeographic Framework
for a Global Marine Biodiversity Observing Network " (Award No. 80NSSC18K0412).
arXiv:2007.05470v1 [cs.LG] 10 Jul 2020

1 INTRODUCTION
Illegal fishing adversely impacts the health of our oceans and will continue to challenge our attempts
to achieve sustainable ocean use unless it is addressed. Recent estimates show illegal fishing repre-
sents as much as 20% of the global fish catch and is valued at $36.4 billion annually (Sumaila et al.
2006; Shaver and Yozell 2018). These negative effects reach into local fishing communities that
rely on sustainably managed fish stocks for sustenance, to support the local economy, and can even
cause geopolitical instability that further undermines economic development (Österblom and Bodin
2012; Mendenhall et al. 2020). To address this important problem we must improve our ability to
detect and predict illegal fishing (Pramod et al. 2014; UN DESA 2016). Illegal fishing undermines
sustainability fisheries by reducing the accuracy of official fish catch and stock estimates, making it
difficult for policymakers and regulators to manage local ecosystems (Worm and Branch 2012). As
a result, illegal fishing pushes total fishing effort beyond maximum sustainable levels thereby creat-
ing negative fish population growth rates and undermining our ability to monitor marine protected
areas and sustain the health of the oceans (Lubchenco and Grorud-Colvert 2015).
One of the main challenges with the monitoring of illegal fishing is that the oceans are costly,
time consuming, and logistically challenging for maritime authorities to patrol. This challenge
becomes even more difficult because vessels will turn off their GPS transponders and go undetected
or falsify their location – this is known as "spoofing" (de Souza et al. 2016). Those vessels that spoof
their location comprise a global "dark fleet" (Sumaila et al. 2006). To detect illegal fishing, recent
research has focused on detecting vessels by detecting spoofing in their Automatic Identification
System (AIS) data (Ford et al. 2018; Miller et al. 2018; Petrossian 2018). However, while detection
is key, it will only ever provide information about past events. Recent advances have targeted the
prediction of illegal fishing, not just detection, by identifying anomalous spatial behaviors of fleets as
they respond to nearby illegal activities (Watson and Woodill 2019). While the method can provide
an early warning signal that illegal activities are soon to happen, precursors of illegal activities are
unable to identify specific vessels as acting legally or not. An important next step is to improve our
ability to predict in near real-time whether a fishing vessel is fishing legally or not, based solely on
the ocean conditions of a given location.
Here, we demonstrate that illegal fishing can be predicted from oceanographic conditions, using
2

the Patagonia Shelf as a case-study. The Patagonian Shelf is a region off the coast of Argentina
that is known to be one of the most productive areas in the world for fisheries. Important ma-
rine ecosystems combined with year-round convective mixing and seasonal fronts provide abundant
pelagic and demersal resources (Csirke 1987; Arkhipkin et al. 2013). This productive fishing region
supports fishing trawlers, squid jiggers, and drifting longline fleets; these fleets have been extracting
up to 1 million tonnes of squid annually, for example, consisting mostly of Argentinean and over 1
million tonnes of fish annually, including Hakes (Merluccius hubbsi), Hoki (Macruronus magellan-
icus), Rock Cod (Patagonotothen ramsayi) and Southern Blue Whiting (Micromesistius australis)
(Arkhipkin et al. 2013; Laptikhovsky et al. 2013). However, the region is heavily overfished. The
UN Food and Agriculture Organization (UNFAO) provided a report in 2015 that showed 42% of
the Southwest Atlantic area was maintaining unsustainable stock levels, the third-worst among 15
major statistical areas (FAO 2018). The report also highlights that the most important species
(in terms of landings), Argentine shortfin squid (Illex argentinus) and is "maximally sustainably
fished." While the region provides productive ocean fishing conditions, overfishing combined with
illegal fishing will have long term negative effects on the sustainability of the region.
Illegal fishing is rampant in the Patagonian Shelf. The former Argentine undersecretary of state
for fisheries, Cesar Lerena, says the country is losing as much as $5 billion US dollars and thousands
of jobs annually (Profeta 2018). Monitoring and enforcement in the region are difficult because
legally and illegally operating fishing vessels operate alongside one another. Further, some of the
most productive areas for fishing are located adjacent to the EEZ boundary. Therefore, fishing
vessels can choose to operate legally just outside the EEZ or enter the EEZ to fish illegally, mostly
undetected. Taking advantage of this spatial feature of the Patagonia shelf, Chinese fishing vessels
have been shown to amass near Argentine waters at the EEZ in search of the shortfin squid (Harkell
2018). Due to this, the sustainability of the squid stock is increasingly threatened as the difficulty
to detect and predict illegal activity increases.
In this paper, we have developed a method for predicting whether a given vessel is fishing
illegally or not, based on its location and the oceanographic conditions it is experiencing. We focus
on Chinese fishing vessels operating on the Patagonia shelf. China has the largest distant water
fishing fleet of any nation and its vessels have engaged in numerous instances of illegal fishing in
locations around the world (Sala et al. 2018; Profeta 2018). Importantly, Chinese fishing vessels do
3

not have the right to fish in the Argentine Exclusive Economic Zone (EEZ). To identify whether
a Chinese fishing vessel is fishing illegally in the Argentine EEZ, we developed a random forest
classifier using oceanographic variables in combination with Automatic Identification System (AIS)
vessel location tracking data. We then used the learning mechanism of the random forest model to
identify the most important oceanographic predictors. Lastly, we also explored results when only
an extremely limited set of oceanographic predictors were used – simulating a situation where only
real-time data are available. There is a growing emphasis on the use of machine learning to improve
the detection and prediction of illegal fishing, and for quantifying total fishing effort in general
(Belhabib et al. 2019). A key advance in the use of machine learning is to use these methods to
accurately distinguish legal from illegal fishing effort. Not only will this provide an estimate of
a key yet hidden source of fish mortality (Agnew et al. 2009), critical for developing accurate and
sustainable fisheries management policies, but this will provide a route to more efficient enforcement
of the laws of the sea.
2 MATERIALS AND METHODS
2.1 Data
Gridded AIS Vessel location data were obtained from Global Fish Watch (GFW) for the Patagonian
Shelf region of South America for the period 2012-2016 (Kroodsma et al. 2018). This publically
available data provide daily fishing vessel locations and fishing effort (in hours) at a 10th-degree
spatial resolution, as well as a unique AIS identifier (MMSI). We also utilized vessel-specific char-
acteristics by MMSI from GFW, such as flag (country) and gear type. For this analysis, we defined
a vessel operating illegally as: (1) a Chinese fishing vessel (denoted by flag), (2) located inside the
Argentine EEZ, with (3) positive fishing effort (i.e. in terms of the GFW data, this is identified in
terms of non-zero fishing hours). In these data, the majority of illegal Chinese vessels were squid
jiggers (81%) and fishing trawlers (15%) with most illegal fishing occurring in February and March
(see Figure 2). Our dependent variable is imbalanced where the number of illegal fishing vessels
makes up only 1.5% of the total observed vessels. This subset provided 114,103 observations, 781
unique vessels operating legally, and 186 unique fishing vessels operating illegally (see Figure 1).
Sea surface temperature (SST), chlorophyll (CHL), and seascape (SEA) observations were mea-
4

sured, using nearest neighbor interpolation, for each vessel location data point. Seascapes are a
relatively new way to categorize ocean features in a way that preserves underlying data distribu-
tions, even when non-linear (Kavanaugh et al. 2016). The marine environment can be viewed as
a mosaic of distinct seascapes, with unique combinations of biological, chemical, geological, and
physical processes that define habitats that change over time (See Table 1 for a list of Seascapes in
the Patagonia Shelf; Kavanaugh et al. 2014). In addition to SST and CHL, which are typically two
main oceanographic variables used to predict the spatial distribution of fishes (Kleisner et al. 2010;
Roberts et al. 2016), apex predators (Scales et al. 2017) and fishing (Belhabib et al. 2019; Crespo
et al. 2018; Soykan et al. 2014), here we make use of seascapes to help summarize the numerous
dimensions of variability in ocean conditions. SST and CHL are available from NASA Ocean Color
and SEA are available from NOAA ERDDAP in 8-day increments (NASA Ocean Biology Process-
ing Group 2014, 2017; ERDDAP 2016; Kavanaugh et al. 2018). Due to cloud cover, observations
for every surface location were not available so we recorded this as "not available". We also used
latitude and longitude coordinates of each vessel and whether they are inside the EEZ or not as
predictor variables (see Table 2 for a list of all variables).
2.2 Predicting Illegal Fishing
The overarching assumption with our data is that fishing vessels operate in locations where the
ocean conditions are favorable for fishing. As such, vessels that are currently operating legally will
evaluate their risk to operate illegally if ocean conditions change in such a way that incentivizes
movement (Watson et al. 2018) to regions where they are not allowed to operate (i.e. inside the
Argentine EEZ). When these conditions change and vessels choose to operate illegally, they receive
a benefit from the change in ocean conditions but at a cost (risk) of getting caught through local
enforcement (Sumaila et al. 2006). Our definition of what constitutes illegal fishing – i.e. any
Chinese vessel fishing in the Argentine EEZ – is strict because we do not know if a particular vessel
has a license to operate in the Argentine EEZ. Due to data limitations, it is not possible to know
which Chinese vessels have licenses to fish in the Argentine EEZ, but we do know that this is only a
small fraction of the total number of the Chinese fleet operating in the region (if any). So while our
assumption that Chinese fishing vessels operating inside the EEZ, are operating illegally, is strict it
is also conservative.
5

To predict whether a Chinese vessel is operating illegally we developed a random forest classifier
where the binary choice is defined as operating legally (y= 0) or illegally (y= 1). The classifier
was fit with explanatory variables discussed above as well as a dummy variable, specifically the
month of the year due to the seasonality of fishing in the region. Our modeling strategy is similar
to species distribution modeling where (illegal) fishing effort is linked to environmental variables
to predict species habitat preferences (Elith and Leathwick 2009; Russo et al. 2011; Watson and
Haynie 2016; O’Farrell et al. 2017). We diverged from these standard models to develop a binary
classifier that can predict whether the vessel operated illegally. Our approach allows us to target
individual vessels and identify where and when illegal activity is likely to occur.
Random forest classification is popular because it can handle complex relationships between
explanatory variables (e.g. multicollinearity, non-linearity) by "bagging" the data into different
subsets across different trees. These subset trees reduce the correlation across the data, thus reducing
the variance and improving predictions. Another benefit of a random forest classifier is that it
identifies the importance of each variable in terms of their contribution to prediction. This allowed
us to include a variety of variables, identify those most important, and reduce our model down to
a relatively simple model while maintaining classification skill. This "kitchen sink" approach does
not just provide a useful prediction tool, but it improves our understanding of the drivers of illegal
fishing.
To validate our model, we utilized a block cross-validation technique where the test set was
a single year and the training set was the remaining years. This technique produced five-folds,
so validation is made on each fold (year). The reason for validating our model in this way is
to prevent overfitting our model due to spatial autocorrelation (Roberts et al. 2017). When the
training and test sets contain locations close (spatially) to each other and from the same year
(temporally), cross-validation will predict those observations correctly while being over optimistic
about observations further away. As a result, the trained model will not generalize to out-of-sample
(test set) predictions, thus leading to overconfidence in the performance of the model. Therefore, we
utilize a block cross-validation technique to ensure observations in the training and test set contain
a similar spatial and temporal structure and are independent.
Since our data is imbalanced (more legal vessels than illegal vessels), we are interested in how
well our model can predict illegal fishing (y= 1) and not legal fishing (y= 0) so our accuracy
6

metrics should address the imbalance. We used five metrics to test the accuracy of our model:
(1) Precision and Recall; (2) F-Score; (3) Area Under the Curve (AUC); (4) Average Precision
(AP); and (5) Average Recall (AR). Precision is the fraction of true illegal activity relative to the
number of false hits of illegal activities (the lower the precision, the more false positives included
in predictions). Recall is the fraction of true illegal activity relative to the number of false misses
of illegal activity (the lower the recall, the more false negatives). Precision describes the ability to
predict true positive values whereas recall describes the ability to predict false positives. We utilize
an F-1 score as our primary accuracy metric because it reflects the accuracy of a classifier with
imbalanced classes (Chawla 2009). An F-1 score measures the accuracy of the test set by using the
harmonic mean of precision and recall where 100% equals an accurate prediction. We also consider
additional metrics to evaluate model performance. Precision and recall can be plotted against each
other over different thresholds of predicted probabilities within the random forest model. From
this comparison, AUC is computed by integrating under the curve for each year to find the level of
accuracy. Finally, AP and AR summarize the precision-recall plot by using the weighted mean of
precision (AP) and recall (AR) for each class (legal/illegal) at each threshold. If AP or AR equals
one then this indicates zero false positives or false negatives in the predictions. For each fold, we
estimated each of these metrics to validate the accuracy of the model.
Our first step to modeling illegal vessels relies on the "kitchen sink" approach where we used a
large number of variables as predictors. Next, we utilized the learning mechanism of a random forest
to identify predictors that are most important to prediction accuracy. We then created a separate
model that included only the top five most important variables (SST, CHL, I-EEZ, LAT, LON).
This simple model allowed us to reduce the complexity of the model to a subset of predictors. In
addition, we also created a model with only oceanographic variables (SST, CHL, SEA). Together,
these reduced models were used to identify the specific ocean conditions that are related to illegal
fishing on the Patagonia shelf. These reduced models were also created as an attempt to develop
a methodology that could be operationalized with real-time data feeds, for which there is only a
limited number of variables relative to the full suite of variables in historical datasets.
7

3 RESULTS
The main results for predicting illegal vessels with all variables are provided in Figure 3 and Table
3 column 1. Figure 3 provides a precision-recall plot where predicted probabilities are varied across
different thresholds to test the sensitivity of the results. Table 3 column 1 reports the F-1 score
for each year that reflects how well the model can accurately predict illegal Chinese fishing vessels.
Results show that the random forest classification algorithm can identify instances of illegal fishing
– as defined as a Chinese fishing vessel in the Argentine EEZ with non-zero fishing effort – with
an accuracy between 69-89% depending on the year (i.e. between 2012-2016). Additional accuracy
metrics are provided in Table S1 where Average-Precision (AP) ranges from 0.79-0.96, Average-
Recall (AR) ranges from 0.69-0.99, and area-under-curve (AUC) ranges from 0.79-0.96. These results
show we can confidently predict whether a vessel is operating illegally using oceanic biophysical
characteristics and measures of geospatial distance.
Next, we used the random forest classifier to identify the best predictors in our model. Figure
4 Panel A provides a list of the top five predictors and importance (variable importance is scaled
to one). The top five variables that are important for predictions are longitude, latitude, sea
surface temperature, chlorophyll, and whether the vessel is inside the Argentine EEZ. Using this
information, we next created a reduced model that includes only these five variables, to again
detect illegal fishing. The results from this analysis are provided in Table 3 column 2, and Table S1
column 2. The F-1 score accuracy ranges from 69-96%, AP values range from 0.75-0.99, AR values
range from 0.76-0.97, and AUC values range from 0.75-0.99. These results highlight that, somewhat
counterintuitively, the reduced model is more accurate compared to our first âĂIJkitchen-sink"
model.
We next used only oceanographic variables: Seascape (SEA), sea-surface temperature (SST)
and chlorophyll (CHL), and the month of the year; simulating a real-time situation where only
these variables are available, and critically when there is an absence of vessel location data (See
Figure 4 Panel B for top five oceanographic predictors). While learning in the random forest model
improved our prediction accuracy, we also wanted to know how changes in oceanographic conditions
can predict illegal fishing. Table 3 column 3 and Table S1 column 3 provides the results from this
analysis. This ocean-only model showed that even with only oceanic biophysical variables, we can
8

still classify illegal fishing with an accuracy of between 71-89%. AP values range from 0.67-0.85,
AR values range from 0.79-0.97, and AUC values range from 0.71-0.88. While this model does not
perform as well as the previous two models it can inform us of when the ocean conditions are ripe
for illegal fishing.
Our model can predict whether a Chinese vessel is operating illegally on the Patagonia shelf; it
can also provide insight into the behavior of fishing vessels in the region. If Chinese vessels expect
better conditions for fishing tomorrow, will they decide to enter the EEZ and fish illegally? The best
seascape predictor from the random forest model was Seascape 14: "Temperate Bloom Upwelling"
(Kavanaugh et al. 2014). The temperate bloom upwelling seascape is common to subpolar and
temperate waters and indicates upwelling which is almost always associated with productive fish-
eries. To explore the behavior of fishing vessels we tracked their movement from operating legally
to illegally and the change in temperate bloom upwelling seascape. We estimated the number of
legal fishing vessels that are not located in temperate upwell blooming seascape, but then move to
this seascape the next day to operate illegally. Figure 5 shows this transition for temperate bloom
upwelling as well as the other seascape transitions. In total, 83 Chinese fishing vessels decided to
enter the EEZ and begin fishing illegally when they noticed these favorable oceanographic/fishing
conditions. The âĂIJTemperate Transition" seascape was the next important seascape, with 80
vessels transitioning to this seascape to operate illegally. The temperate transition zone is typified
by temperatures ranging from 11-15 C, moderate to high salinities (>34 psu), low mesotrophic
conditions ([chl-a] 0.2 to 0.5 mg m-3 with positive sea level anomalies indicative of convergent zones
(see Table 1 for a list of seascape). These results suggest fishing vessels adjust their risk preferences
to move to favorable areas of the ocean, which can be characterized by specific seascapes, to begin
operating illegally at the risk of being caught to capture more fishing effort when conditions are
favorable.
Finally, to gain an understanding of the regional oceanographic conditions associated with il-
legal fishing, we explored the spatial behavior of Chinese fishing vessels on the Patagonia shelf by
examining the spatial distribution of the temperate bloom upwelling seascape when: (1) a day when
high levels of illegal activity occur (n=76); and (2) when low levels of illegal activity occur (n=3).
Figure 6 provides a map of legal fishing vessels (black dots), illegal fishing vessels (red dots), and the
spatial distribution of the Temperate Bloom Upwelling seascape (blue). Note that the area covered
9

by the temperate bloom upwelling seascape when there is a high level of illegal activity (panel A),
is less than when there were low levels of illegal activity (panel B). This result suggests that when
as upwelling/bloom seascape conditions diminish, this constricts the available fishing area so that
vessels are more likely to move into the Argentine EEZ and operate illegally. The comparison made
in Figure 6 provides further (qualitative) evidence of spatial behavioral responses to changes in
ocean conditions.
4 DISCUSSION
To identify when illegal fishing occurs on the Patagonia shelf, we developed a random forest classifier
trained on vessel location data and oceanographic information. We focused on Chinese fishing vessels
operating illegally inside the Argentine EEZ, and used the spatial distribution of different seascapes
as predictors, amongst others. We performed cross-validation by individual years to measure the
skill of our predictions and found that the model can predict illegal fishing with an accuracy of 69-
89%. We also created a reduced model, one with a restricted set of important predictor variables,
and showed that the random forest classifier can still accurately predict with 69-96% confidence.
Further still, using oceanographic predictor variables only, we still achieved a level of accuracy from
71-89%. Finally, we explored how fishers’ spatial behaviors respond to changes in a specific seascape
– the temperate blooms upwelling seascape, which we find is the seascape to which fishers move to
fish (illegally). A recent paper provided a new method for measuring illegal fishing effort based on
the spatial behavior of vessels located in an illegal fishing area (Belhabib et al. 2019). Our method
improves upon this method by providing the critical first step in classifying whether the fishing
vessel is, in fact, operating illegally, based on oceanic conditions. These results provide important
insights into the limits to our predictive skill and the (spatial) behavioral choices of fishing vessels
that commit these crimes.
The methodology introduced here is a necessary first step to making (near) real-time predictions
of illegal activities at sea, potentially for the whole globe. To make subsequent improvements, an
important next step is to expand the scope of our analysis to include other important regions
where illegal fishing occurs. One challenge will be that the Patagonian Shelf is a unique case study
because the Argentine EEZ border essentially runs straight through a major fishing hotspot, leading
10

to foreign fishing vessels essentially âĂIJfishing along the line" of the EEZ. As a consequence, foreign
vessels are often tempted (by changing ocean conditions or by changing spatial distributions of target
species) to make incursions into the Argentine EEZ. This specific spatial nature of the Patagonia
shelf system will not be represented in other systems around the world, and how and why illegal
fishing will occur will be nuanced in ways that new machine learning approaches might be required
in other regions.
To improve the utility of our machine learning approach, it can (and will) be combined with other
data and methods. For example, recent advances in the detection of anomalous multiscale patterns
in fleet spatial behaviors have shown that (spatial) early-warning signals proceed IUU activities
((Watson and Woodill 2019). Methods like these can be combined with the machine learning
method presented here, making it possible to refine the situational awareness of IUU ongoing in a
particular region. No single approach offers a panacea. Essentially, a suite of IUU detection and
prediction techniques can (and should) be employed in concert to identify potential "hot spots"
of illegal activity. Doing so will greatly increase the efficiency by which enforcement organizations
monitor/patrol areas of our oceans.
A major caveat to note is that we have defined illegal fishing specifically in terms of Chinese
vessels operating in the Argentine EEZ, which we have assumed they have no right to do. In general,
this should be true, as most Chinese vessels that do not have the rights to fish in Argentine waters
(Mallory 2013). But we cannot say with complete confidence that this is so. As a consequence,
there is a risk that our analysis includes false positives. Further, our analysis shows that vessels (of
all nations except Argentina) operate extremely close to the EEZ line. Vessels may not be aware
of the exact location of the EEZ line, and perhaps even "a little" incursion into Argentine waters
might be the norm. To explore the impact of the proximity of fishing vessels to the EEZ line, we
repeated our analysis but with an increased buffer zone around the EEZ line; specifically, illegal
fishing was redefined as a Chinese vessel fishing within 2, 5 and 10 kilometers (km) of the EEZ
line. Table 3 columns 4-6 and Table S1 columns 4-6 provides the results of this sensitivity analysis.
The 2km and 5km buffer results show the model is still able to predict illegal fishing with an F-1
accuracy score between 45-92%. However, at a 10km buffer, the results are less stable across all
accuracy metrics. In 2012, there were no fishing vessels operating illegally inside the EEZ so we
are unable to evaluate accuracy metrics. For 2013-2016, the F-1 scores range between 0 and 92%.
11

While the model does break down with a buffer at 10km, it still performs rather well with a 2-5km
buffer. This suggests that it is possible to predict the short-distance but frequent incursions into
Argentine waters, but not the long-distance incursions that are rare.
An important result from this analysis is the need for finer data resolution to predict illegal
activity. While our model does well at predicting illegal activity when vessels incur <5km into
the EEZ, the model accuracy drops when vessel incursions are >10km from the EEZ border. The
important point to note is that the resolution of the oceanographic data used is 10km, meaning the
model results are confined to this spatial scale also. Therefore, to make accurate determinations
of illegal fishing, higher spatial and temporal resolution oceanographic data is required. Higher
spatial resolution (300 m to 1 km) ocean color is available from moderate resolution spectrometers
on NASA and ESA satellites. For example, the proposed PACE (Phytoplankton Aerosol Carbon
and Ecosystems) mission of NASA will provide hyperspectral remote sensing reflectances which
will increase the capacity of imagery to provide information on phytoplankton size distribution and
functional type, and thus bloom succession and lower trophic level food quality. However, all polar
orbiting satellites tend to have repeats on the order of 1-2 days, which may minimize the utility
of a near real-time monitoring scheme, given travel times and speeds of industrialized fishing ves-
sels. Geostationary satellites, however, provide near continuous monitoring of a region, potentially
providing hourly information on the location of features. Currently, the only geostationary satellite
with operational ocean color is the South Korean Geostationary Ocean Color Imager (GOCI) that
provides 500 m resolution ocean color data for the Korean Sea and the northern East China Sea.
5 CONCLUSION
To conclude, deterring the impact of illegal fishing will improve the health and sustainability of
our oceans. The impact of illegal fishing reaches into regional and local economies and even causes
geopolitical tensions which further reduces the growth of developing countries’ economies. Countries
have begun improving maritime monitoring and law enforcement; however, recent advances rely on a
retrospective analysis of individual vessel spoofing which is unable to provide real-time information.
Going forward, it is vital that law enforcement be able to prevent illegal activities. To make steps in
this direction, we have developed a machine learning algorithm that can accurately predict whether
12

a Chinese fishing vessel is operating illegally with 69-96% accuracy in the Patagonian Shelf region
of Argentina. In addition to prediction, our use of machine learning has informed us of why illegal
fishing happens in the region based on changing ocean conditions. These results offer a promising
step towards real-time monitoring and likelihood estimates of "hot spots" where illegal fishing
is likely to occur based on changes in oceanographic conditions. As new information and data
become available, these valuable tools will allow maritime law enforcement to improve their ability
to monitor and enforce illegal activity in the oceans, thus reducing illegal fishing and improving the
sustainability of fish stocks.
6 ACKNOWLEDGEMENTS
Correspondence: A. John Woodill, Oregon State University, johnwoodill@gmail.com. We would like
to thank Global Fish Watch for providing the vessel location data for our analysis. This work was
supported by NASA under the award "Securing Sustainable Seas: Near real-time monitoring and
predicting of global fishing fleet behavior" (Award No. 80NSSC19K0203). Additional support for
developing seascapes was provided from NASA under the award " Dynamic Seascapes to Support
a Biogeographic Framework for a Global Marine Biodiversity Observing Network " (Award No.
80NSSC18K0412)
7 DATA AVAILABILITY
The AIS data to support the findings in this study are available at Global Fish Watch (https:
//globalfishingwatch.org/datasets-and-code/). Oceanographic data are available at NASA
Ocean Color (https://oceancolor.gsfc.nasa.gov/) and ERDDAP (https://cwcgom.aoml.noaa.
gov/erddap/griddap/noaa_aoml_seascapes_8day.graph). Code to support the findings in this
study are available online.
13

References
Agnew, D. J., J. Pearce, G. Pramod, T. Peatman, R. Watson, J. R. Beddington, and T. J. Pitcher
(2009). Estimating the worldwide extent of illegal fishing. PloS one 4 (2), e4570.
Arkhipkin, A., P. Brickle, and V. Laptikhovsky (2013). Links between marine fauna and oceanic
fronts on the Patagonian Shelf and Slope. Arquipélago. Life and Marine Science 30, 19–37.
Belhabib, D., W. W. Cheung, D. Kroodsma, V. W. Lam, P. J. Underwood, and J. Virdin (2019).
Catching industrial fishing incursions into inshore waters of Africa from space. Fish and Fisheries .
Chawla, N. V. (2009). Data mining for imbalanced datasets: An overview. In Data mining and
knowledge discovery handbook, pp. 875–886. Springer.
Crespo, G. O., D. C. Dunn, G. Reygondeau, K. Boerder, B. Worm, W. Cheung, D. P. Tittensor,
and P. N. Halpin (2018). The environmental niche of the global high seas pelagic longline fleet.
Science advances 4 (8), eaat3681.
Csirke, J. (1987). The Patagonian fishery resources and the offshore fisheries in the South-West
Atlantic. FAO Fisheries Technical Paper (FAO).
de Souza, E. N., K. Boerder, S. Matwin, and B. Worm (2016). Improving fishing pattern detection
from satellite AIS using data mining and machine learning. PloS one 11 (7), e0158248.
Elith, J. and J. R. Leathwick (2009). Species distribution models: ecological explanation and
prediction across space and time. Annual review of ecology, evolution, and systematics 40, 677–
697.
ERDDAP (2016). 8-day global seascapes.
FAO (2018). The state of world fisheries and aquaculture 2018-meeting the sustainable development
goals.
Ford, J. H., D. Peel, D. Kroodsma, B. D. Hardesty, U. Rosebrock, and C. Wilcox (2018). Detecting
suspicious activities at sea based on anomalies in automatic identification systems transmissions.
PloS one 13 (8), e0201640.
14

Harkell, L. (2018). Chinese squid jiggers amass near Argentine waters as official season start nears.
Undercurrent News .
Kavanaugh, M. T., M. J. Church, C. O. Davis, D. M. Karl, R. M. Letelier, and S. C. Doney (2018).
Aloha from the edge: Reconciling three decades of in situ eulerian observations and geographic
variability in the north pacific subtropical gyre. Frontiers in Marine Science 5, 130.
Kavanaugh, M. T., B. Hales, M. Saraceno, Y. H. Spitz, A. E. White, and R. M. Letelier (2014).
Hierarchical and dynamic seascapes: A quantitative framework for scaling pelagic biogeochemistry
and ecology. Progress in Oceanography 120, 291–304.
Kavanaugh, M. T., M. J. Oliver, F. P. Chavez, R. M. Letelier, F. E. Muller-Karger, and S. C.
Doney (2016). Seascapes as a new vernacular for pelagic ocean monitoring, management and
conservation. ICES Journal of Marine Science 73 (7), 1839–1850.
Kleisner, K. M., J. F. Walter III, S. L. Diamond, and D. J. Die (2010). Modeling the spatial
autocorrelation of pelagic fish abundance. Marine Ecology Progress Series 411, 203–213.
Kroodsma, D. A., J. Mayorga, T. Hochberg, N. A. Miller, K. Boerder, F. Ferretti, A. Wilson,
B. Bergman, T. D. White, B. A. Block, et al. (2018). Tracking the global footprint of fisheries.
Science 359 (6378), 904–908.
Laptikhovsky, V., A. Arkhipkin, and P. Brickle (2013). From small bycatch to main commercial
species: explosion of stocks of rock cod patagonotothen ramsayi (regan) in the Southwest Atlantic.
Fisheries research 147, 399–403.
Lubchenco, J. and K. Grorud-Colvert (2015). Making waves: The science and politics of ocean
protection. Science 350 (6259), 382–383.
Mallory, T. G. (2013). China’s distant water fishing industry: Evolving policies and implications.
Marine Policy 38, 99–108.
Mendenhall, E., C. Hendrix, E. Nyman, P. M. Roberts, J. R. Hoopes, J. R. Watson, V. W. Lam, and
U. R. Sumaila (2020). Climate change increases the risk of fisheries conflict. Marine Policy 117,
103954.
15

Miller, N. A., A. Roan, T. Hochberg, J. Amos, and D. A. Kroodsma (2018). Identifying global
patterns of transshipment behavior. Frontiers in Marine Science 5, 240.
NASA Ocean Biology Processing Group (2014). Moderate-resolution imaging spectroradiometer
(modis) aqua 11Âţm day/night sea surface temperature data.
NASA Ocean Biology Processing Group (2017). Modis-aqua level 3 mapped chlorophyll data version
r2018.0.
Österblom, H. and Ö. Bodin (2012). Global cooperation among diverse organizations to reduce
illegal fishing in the Southern Ocean. Conservation biology 26 (4), 638–648.
O’Farrell, S., J. N. Sanchirico, I. Chollett, M. Cockrell, S. A. Murawski, J. T. Watson, A. Haynie,
A. Strelcheck, and L. Perruso (2017). Improving detection of short-duration fishing behaviour
in vessel tracks by feature engineering of training data. ICES Journal of Marine Science 74 (5),
1428–1436.
Petrossian, G. A. (2018). A micro-spatial analysis of opportunities for IUU fishing in 23 Western
African countries. Biological Conservation 225, 31–41.
Pramod, G., K. Nakamura, T. J. Pitcher, and L. Delagran (2014). Estimates of illegal and unre-
ported fish in seafood imports to the USA. Marine Policy 48, 102–113.
Profeta, D. (2018). Argentina, the next battleground for illegal fishing. China Dialogue Ocean.
Roberts, D. R., V. Bahn, S. Ciuti, M. S. Boyce, J. Elith, G. Guillera-Arroita, S. Hauenstein, J. J.
Lahoz-Monfort, B. Schröder, W. Thuiller, et al. (2017). Cross-validation strategies for data with
temporal, spatial, hierarchical, or phylogenetic structure. Ecography 40 (8), 913–929.
Roberts, J. J., B. D. Best, L. Mannocci, E. Fujioka, P. N. Halpin, D. L. Palka, L. P. Garrison, K. D.
Mullin, T. V. Cole, C. B. Khan, et al. (2016). Habitat-based cetacean density models for the us
atlantic and gulf of mexico. Scientific reports 6 (1), 1–12.
Russo, T., A. Parisi, M. Prorgi, F. Boccoli, I. Cignini, M. Tordoni, and S. Cataudella (2011). When
behaviour reveals activity: Assigning fishing effort to métiers based on vms data using artificial
neural networks. Fisheries Research 111 (1-2), 53–64.
16

Sala, E., J. Mayorga, C. Costello, D. Kroodsma, M. L. Palomares, D. Pauly, U. R. Sumaila, and
D. Zeller (2018). The economics of fishing the high seas. Science Advances 4 (6), eaat2504.
Scales, K. L., E. L. Hazen, M. G. Jacox, C. A. Edwards, A. M. Boustany, M. J. Oliver, and S. J.
Bograd (2017). Scale of inference: on the sensitivity of habitat models for wide-ranging marine
predators to the resolution of environmental data. Ecography 40 (1), 210–220.
Shaver, A. and S. Yozell (2018). Casting a wider net: The security implications of illegal, unreported,
and unregulated fishing.
Soykan, C. U., T. Eguchi, S. Kohin, and H. Dewar (2014). Prediction of fishing effort distributions
using boosted regression trees. Ecological Applications 24 (1), 71–83.
Sumaila, U. R., J. Alder, and H. Keith (2006). Global scope and economics of illegal fishing. Marine
Policy 30 (6), 696–703.
UN DESA (2016). The Sustainable Development Goals Report 2016.https://doi.org/10.18356/
3405d09f-en.
Watson, J. R., E. C. Fuller, F. S. Castruccio, and J. F. Samhouri (2018). Fishermen follow fine-scale
physical ocean features for finance. Frontiers in Marine Science 5, 46. 10.3389/fmars.2018.
00046.
Watson, J. R. and A. J. Woodill (2019). Anticipating illegal maritime activities from anomalous
multiscale fleet behaviors. arXiv preprint arXiv:1910.05424.
Watson, J. T. and A. C. Haynie (2016). Using vessel monitoring system data to identify and
characterize trips made by fishing vessels in the United States North Pacific. PloS one 11 (10),
e0165173.
Worm, B. and T. A. Branch (2012). The future of fish. Trends in ecology & evolution 27 (11),
594–599.
17

Class Name Description
1 North Atlantic Spring,
ACC Transition
Cold (4-7 C), low dynamic topography,
moderate nutrient stress
2 Subpolar-Subtropical
Transition
Variable SST (12-21C), Oligotrophic to
mesotrophic, moderate to high nutrient stress
7 Temperate Transition Temperatures ranging from 11-15 C,
moderate to high salinities (>34 psu), low
mesotrophic conditions ([chl-a] 0.2 to 0.5 mg
m-3 with positive sea level anomalies indicative
of convergent zones
12 Subpolar Cool (6-9 C), mesotrophic, low CDOM
14 Cool Temperate
Blooms Upwelling
Subpolar and temperate waters – indicates
upwelling which is almost always associated
with productive fisheries
15 Tropical Seas Warm, salty ( 35psu), low mesotrophic
conditions
17 Subtropical Transition
(low nutrient stress)
Warm, mesotrophic, low nutrient stress
19 Arctic/Freshwater
influenced Subpolar
Shelves
Lower salinity <32, high CDOM
21 Warm, Blooms, High
Nuts
18-22 degrees C, high chl-a (>2 mg m-3), high
CDOM, salinity variable but relatively fresh
(32 psu)
Table 1: Summary of Patagonian Shelf Seascapes
Table provides a list of the most prevalent seascapes in the Patagonian Shelf region.
18

Variable Abbr. Description Source
Sea Surface
Temperature
SST Skin sea surface temperature in
units of Celcius.
NASA Ocean Color Level-3
MODIS-Aqua 8-day SST at 4km
resolution.
Chlorophyll CHL Near-surface concentration of
chlorophyll-a
NASA Ocean Color Level-3
MODIS-Aqua 8-day CHL at 4km
resolution.
Seascapes SEA Classification (n=30) of ocean
features as seascapes derived from
CHL, SST, and photosynthetically
available radiation (PAR).
NOAA ERDDAP 8-day SEA at
5km resolution.
Latitude LON Geographical coordinate of fishing
vessel along the east-west axis.
AIS Global Fishwatch Data
Longitude LAT Geographical coordinate of fishing
vessel along the north-south axis.
AIS Global Fishwatch Data
Inside EEZ I-EEZ Binary variable describing whether
fishing vessel is located inside the
EEZ
AIS Global Fishwatch Data
Distance to
EEZ
D-EEZ The distance a fishing vessel is from
the EEZ (in kilometers)
AIS Global Fishwatch Data
Table 2: Summary of Random Forest Variables
Table provides a list of variables, abbreviations, a brief description and the source of the data used in
the random forest classifier.
19

Random
Forest
Random
Forest
Random
Forest
Random
Forest
Random
Forest
Random
Forest
all
variables
top five
variables
Seascape,
SST, CHL,
Month
all variables
2km EEZ
all variables
5km EEZ
all variables
10km EEZ
(1) (2) (3) (4) (5) (6)
F-1 Score
2012 69% 84% 89% 92% 80% NA
2013 70% 74% 77% 58% 47% 0
2014 55% 69% 75% 79% 45% 0
2015 89% 93% 71% 81% 78% 92%
2016 80% 96% 73% 92% 40% 67%
Table 3: Summary of Main Results
Table provides results from our analysis across different model specifications shown in column: (1)
using all variables; (2) top variables from random-forest; (3) Seascape and Distance measures; (4-6)
sensitivity analysis of vessels distance to EEZ being inside 2km, 5km, and 10km. Table provides
an F-1 Score to evaluate model performance for each predicted year 2012-2016. See Table S1 for
additiona accuracy metrics.
20

Random
Forest
Random
Forest
Random
Forest
Random
Forest
Random
Forest
Random
Forest
all
variables
top five
variables
Seascape,
SST, CHL,
Month
all variables
2km EEZ
all variables
5km EEZ
all variables
10km EEZ
(1) (2) (3) (4) (5) (6)
Average-Precision
2012 0.90 0.80 0.85 0.96 0.82 NA
2013 0.79 0.75 0.67 0.41 0.22 0
2014 0.83 0.84 0.77 0.86 0.49 0.20
2015 0.96 0.94 0.77 0.95 0.89 1
2016 0.92 0.99 0.79 0.99 0.87 0.71
Average-Recall
2012 0.99 0.97 0.97 0.99 0.99 NA
2013 0.79 0.86 0.92 0.82 0.72 0.50
2014 0.69 0.76 0.82 0.87 0.67 0.50
2015 0.91 0.96 0.97 0.84 0.82 0.92
2016 0.83 0.97 0.80 0.95 0.63 0.77
Area-Under-Curve
2012 0.90 0.78 0.88 0.96 0.8 NA
2013 0.79 0.75 0.71 0.46 0.18 0
2014 0.83 0.85 0.76 0.87 0.48 0.21
2015 0.96 0.94 0.78 0.95 0.89 1
2016 0.92 0.99 0.79 0.99 0.87 0.71
Table S1: Summary of Additional Results
The table provides results from our analysis across different model specifications shown in column: (1) using
all variables; (2) top variables from random-forest; (3) Seascape and Distance measures; (4-6) sensitivity
analysis of vessels distance to EEZ being inside 2km, 5km, and 10km. The table provides Average-Precision,
Average-Recall, and Area-Under-Curve accuracy metrics to evaluate model performance for each predicted
year 2012-2016.
21

Figure 1: Map of Patagonian Shelf with Seascapes and Illegal/Legal Fishing Vessels
Map of the Patagonian shelf region with the EEZ (orange dashed line), five most prevalent seascapes
(colored in blue), and the location of vessels on March 13, 2016. Chinese vessels fishing illegally, i.e.
within the Argentine EEZ, are highlighted by red markers.
22

Figure 2: Number of Illegal Events per Month and Year
Figure provides the number of illegal events by month and year in the Patagonia Shelf region. The
majority of illegal Chinese fishing occurs in February and March when the squid fishery is most
active.
23

Figure 3: Precision-recall Plot using Random Forest Classifier with all Variables
The figure provides a precision-recall plot for the random forest classifier with all variables. The
precision-recall is used to evaluate the accuracy of the model as a function of different probability
thresholds in the classifier for each prediction year. A model that is highly predictive bows towards
(1, 1) in the plot. Additional statistics are provided to evaluate the performance of the classifier.
An F-1 score evaluates the number of true and false positives to provide an accuracy score of 100%
if perfect. Area-under-curve (AUC) provides the integral of the precision-recall plot where a 1.0
represents a perfect prediction. Average-precision (AP) and Average-recall (AR) weights the average
precision/recall across thresholds where 1.0 provides a perfect prediction.
24

Figure 4: Best Predictors from Random Forest Classifier
The top five predictors used for predicting illegal fishing on the Patagonia shelf from the random
forest classifier for the primary model (Panel A) and model with only oceanographic variables (Panel
B). In Panel A, location variables (longitude and latitude) were the most important predictors; i.e.
it is an obvious condition that in order for a Chinese vessel to fish illegally in the Argentine EEZ, it
must first be in the Argentine EEZ. In Panel B, sea surface temperature and chlorophyll were the
best predictors.
25

Figure 5: Seascape Transition from Legal to Illegal Region
The number of fishing vessels that change status from fishing legally to illegal as they move to
a given seascape (on the left). The behavior of fishing vessels shows they will choose to operate
illegally if a seascape associated with productive fishing (e.g. the temperature bloom upwelling
seascape) is within the Argentine EEZ.
26

Figure 6: Seascape with Large and Small Illegal Vessel Activity
Map of the Patagonian Shelf with high levels of illegal fishing (Panel A n=76) and low levels of
illegal fishing (Panel B n=3) for two days in 2016. The EEZ is shown as an orange dashed line. Red
dots represent illegal fishing vessels and black dots represent legal fishing. An important seascape
associated with fishing – the "temperate bloom upwelling" seascape is shown in dark blue. In
general, as the area covered by this seascape diminishes we see a rise in the number of illegal fishing
events. This suggests that a spatial concentration of fishing effort, driven by changes in the spatial
distribution of certain seascapes, can promote illegal fishing in the region.
27