Predicting fisheries from albatross movements requires accounting for individual variability in interaction

Abstract

Fisheries have major ecological impacts including bycatch of foraging seabirds, but it is often difficult to obtain comprehensive information on the presence of fishing vessels. Automatic Identification System (AIS) data can be used to monitor fisheries and their interactions with GPS-tracked seabirds, but not all vessels have their AIS operational. Bird-tied radar detectors can overcome this limit and complement monitoring, but the technology is recent and costly. We used both methods combined as a training dataset for classification algorithms, to extend the identification of interactions to GPS tracks without radar detectors nor AIS. We studied over 3 years wandering albatrosses from the French Southern Territories, interacting with toothfish and tuna longliners. We used 196 GPS tracks combined with radar detectors, to calculate different movement variables over various scales (time spent in an area, sinuosity, speed) and used a Random Forest to distinguish behaviour in presence or absence of fishing vessels. Our model reached high classification accuracy (ca. 85%) for individual birds included in the training dataset. However, we lost predictive power (around 72% of accuracy, with a drop of specificity from 76 to 59%) when predicting on individuals not included in the training dataset. Our results emphasize the importance of documenting and accounting for individual variations to use animals as sentinels. We discuss the pros and cons of different research avenues (data sampling, classification model, bird species, etc.) to eventually get to predict fisheries from bird movements only.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-14882-4.

Introduction

Fisheries often have important ecological impacts on various compartments of marine ecosystems. In the absence of mitigating measures, impacts go from overfishing of fish stocks^1,2 to alterations of benthic ecosystems (e.g. physical damages^3,4 and/or artificial feeding of scavengers⁵); to incidental captures (bycatch) of fish, sharks, seabirds, cetaceans and turtles^6,7. In seabirds for example, and particularly for albatrosses, bycatch is recognised as one of the primary extinction threats to populations worldwide⁸. These issues are exacerbated with Illegal, Unreported and Unregulated fisheries (IUU) that are likely not to follow recommendations and guidelines to try mitigating these ecological impacts. IUU is thus among the most harmful human activities for marine ecosystems^9,10.

One of the main challenges in addressing these issues has long been the difficulty of monitoring the marine environment, even for the basic task of documenting the extent and location of these activities. Autonomous vehicles for ocean sampling and surveillance offer promising perspectives but still face many challenges^11,12. Recently, two new sources of data on fine-scale vessel location have become gradually available allowing to more precisely monitor fishing activities. First, there is the availability of VMS and AIS data^13–15, now covering global scale and easily accessible. These systems provide GPS trajectories of most vessels > 12 m operating within EEZ and/or in international waters. However, these data are incomplete, missing some vessels/fleets (e.g. artisanal fleets composed of vessels < 12 m, but also a part of large industrial vessels), and of course not covering IUU^15,16.

Second, radar detectors coupled with GPS-GNSS tags, can be placed on albatrosses^15,17,18 and other seabirds^19,20 that are attracted in huge numbers around fishing vessels²¹. Since most vessels always have their radar activated, this system allows researchers to record whether seabirds approach a vessel (< 5 km¹⁸). Seabirds, particularly albatrosses, are known to cover vast distances, and are often easy to approach and equip with tracking systems at their breeding colonies, making them ideal candidates for long-term monitoring²². By comparing radar detections to AIS data it is also possible to estimate the extent of vessels encountered by birds that are not declared to the AIS system, some of which likely to be IUU (e.g. 40% in international waters of the South-West Indian Ocean, and 10–100% in subantarctic EEZ therein¹⁵). However, the GPS-radar detector is a relatively new technology, and remains expensive for wide and continuous monitoring of populations. As a result, both past and current tracking studies on albatrosses and other seabirds are still largely based on GPS data alone.

In theory, seabirds adopt specific behaviours when around fishing vessels. Being able to recognize interactions with fishing vessels only from the movement and behaviour of seabirds would thus be invaluable, for at least two reasons. It could considerably extend the monitoring of fishing activities from the large numbers of seabirds tracked worldwide with GPS every year²³. It could also permit to retrospectively analyse how much birds interacted with fisheries in the past, and better understand how they may adapt to changes in the environment and/or in fishing practices^24–27. With the ability to comprehensively assess vessel interactions along animal movements from independent data sources (AIS and/or GPS-radar detectors), there is now the opportunity to train models (e.g. machine learning algorithms) to recognize within GPS movement data some patterns linked to interactions with fisheries^28–30.

Early exploratory studies in albatrosses seemed to identify different movement patterns when associated (or not) with fishing vessels^29,30. However, the statistical approaches used were more descriptive than predictive. The first attempt to apply a machine learning approach to predict interactions between wandering albatrosses and fishing vessels lacked predictive power and concluded that there were insufficient differences between interactions with fisheries and more “natural” foraging behaviours²⁸. Considering the promising results from early exploratory analyses, two other factors could alternatively explain the lack of predictive power in the study by Carneiro et al. First the sample size may have been too small to capture the full variability of interaction behaviours. Interaction may vary in particular with the type of fisheries (e.g. trawling, longlining, etc.) and/or depending on individual bird characteristics. It is indeed known that some characteristics of foraging in seabirds can be individual-dependent (i.e. individual specialization^31–33); beyond population, age and/or sex effects^34,35 and/or can be plastic within-individuals (e.g. learning and memory³⁶), although whether and how it could affect movement patterns around fisheries is largely unknown. Second, the technical method used in¹⁷ did not directly contrast local and larger-scale movement patterns around any location, unlike initial exploratory analyses did (i.e. more continuous descriptions of Area-Restricted-Search patterns at various scales^17,29). Yet it is often assumed that it is the contrast between relatively straight and slow local path when following vessels setting or hauling a line, and much more variable path at a larger scale mirroring vessel’s own fishing movements, which may help recognize interactions from the GPS path and differentiate them from natural foraging.

The aim of this study is to develop a model able to predict seabird interactions with fishing vessels based on their GPS tracks, or at least better identify why predictions remain unsuccessful (i.e. lack of contrast with natural foraging or large variability in interaction behaviour). Compared with the earlier study by Carneiro et al., here we trained a machine learning algorithm on a different population of the same albatross species (Wandering albatross Diomedea exulans). We used a slightly larger sample size and in a region with less variability in fleets visited by albatrosses, i.e. almost exclusively longliners in our case¹⁵ compared with the diversity of both types of fisheries and prey that exists around South Georgia¹⁷. While we used a similar Random Forest modelling approach, we included unlabelled continuous descriptors of albatross movements at various scales for each GPS location (i.e. 2 to 150 km around each location), instead of segmenting the track into foraging or transit section and then predicting whether a foraging section was with a vessel or not. This allowed the model to potentially learn directly from “raw” behavioural patterns, and their variability, performed across contrasting scales. Finally, to assess how individual variations in albatross-vessel interactions might affect predictive power, rather than a lack of contrast between “natural” and “vessel-associated” foraging, we carefully designed sensitivity analyses focusing specifically on how the dataset was partitioned between training/validating and testing.

Materials and methods

Dataset

We used 193 foraging tracks during incubation shifts of 162 wandering albatrosses (Diomedea exulans) from Crozet and Kerguelen archipelagos, respectively on Possession Island and Courbet Peninsula, combined with radar detectors. Those detectors measure radio signals from the X band frequencies of marine radar devices on vessels used for navigation safety¹⁸.

Adult birds were temporarily equipped with loggers on their back feathers using Tesa tape. All birds showed no sign of damage after release neither after recapture. 83 males and 79 females were equipped (Table 1 for the distribution by breeding season), with an age range of 8 to 43 years old. Each individual was tracked for 1 to 4 trips (136 being tracked for 1 trip; 22 for 2 trips; 3 for 3 trips and 1 for 4 trips). Three logger types were used in this study: Centurion (< 80 g), XGPS (35 g) and Spoutnik (< 80 g). They all include a GPS and a radar detector (for more details about the functioning, see¹⁸), differing mainly on their real-time transmission options, with no impact on the present analyses. The loggers corresponded to 0.45–1.02% of the female body mass, and 0.37 to 0.85% of male body mass, which is well within ethical recommendations³⁷. Comité d’Ethique en Expérimentation Animale Numéro 084, Terres Australes et Antarctiques Françaises, Comité d’Environnement Polaire and Conseil National de la Protection de la Nature all approved this experiment to be carried out and all experiments were performed in accordance with these guidelines and regulations as well as with standard guidelines for bird handling and tagging.

Table 1.

Number of tracks recorded each breeding season according to the sex of the individuals.

Breeding season	2016/2017	2017/2018	2018/2019	2019/2020
Male	15	13	36	35
Female	22	15	24	33

Due to differences in frequency of GPS acquisition (from 1 point/min to 1 point/3 mins; depending of the device), tracks were interpolated using a linear method to generate a track with a point every 15 min, in order to match the sampling rate of most GPS data available without radar detectors in this population (to which the predictions of the model could have later been applied). In case of gaps in the GPS recordings of more than 4 h, the track was split in parts and the gap was not filled with interpolations.

Each bird location was assigned a binary “vessel interaction state” (yes or no; example in Fig. 1) based on both radar detections and declared GPS vessel locations (AIS data commercially obtained via CLS company, from data furnished by Orbcomm company). After bird location interpolation, we first attributed to interpolated bird locations the same “vessel radar state” as the closest radar recording. Then we spatio-temporally matched the bird’s locations with AIS vessel location interpolated at the exact same time, and associated a positive AIS interaction only when the bird location was at less than 5 km from a vessel (to match with the detection range of the radar detector¹⁸). We combined these to obtain our “correctly annotated” training data set: a positive vessel interaction state when either or both methods detected a vessel nearby, and a negative one when none of the methods detected a vessel within 5 km.

Fig. 1.

Foraging track of an adult wandering albatross from La Possession Island during incubation, overlaid with vessel-interaction locations derived from bird-borne radar measurements in the inset.

Metrics used as inputs for the predictive model

We used different metrics to characterise each bird location, either behavioural, biological or abiotic, and to train our model to associate any bird location to either an interaction state (1) or a no-interaction state (0).

We first used four different behavioural metrics to describe the bird’s movement at different scales around each GPS location: Sinuosity, i.e. the degree of curvature within an area; overall time spent in an area (First Passage Time, hereafter FPT, stricto sensu, i.e. without a segmentation procedure to infer behavioural state^38,39); estimates of average speed and estimates of variability in speed. We expected that interactions with vessels might result in specific scale-dependent patterns of speed, sinuosity and FPT²⁹: at a local scale birds attending a vessel might reduce speed and follow relatively straight path as the longline is set but at a larger scale if birds follow vessels for longer periods they may alternate between phases of variable speed, sinuosity and FPT.

Sinuosity was calculated at different spatial scales (radius values varying from 2 to 150 km; referred as Sinuos.r2 to Sinuos.r150 in Fig. 3) as the observed distance travelled by the bird to cross a circle of radius r centred on the focal location, divided by the minimal theoretical distance to cross this circle, i.e. 2r. FPT was defined as the time taken to cross the same circles as those used for sinuosity (thus reflecting not only sinuosity but also speed within the circle; referred as FPT.r2 to FPT.r150 in Fig. 3).

Fig. 3.

Relative variable importance for the base random partitioning model with every variable included (for the one with only trajectories-based metrics, see ESM 1).

Those values of circle radius are quite arbitrary, but they are chosen to fit in 3 categories: small-scale behaviours, i.e. behaviour that can take place in the direct vicinity of a location, and that matches our interaction definition (2 and 5 km); mid-scale behaviours that can take place in the visual detection range of a vessel by the bird (10 and 20 km) and large-scale behaviours expected to reflect more global activity state of the bird (50, 100 and 150 km).

For speed, we used two types of metrics, the mean speed and the standard deviation of speed. Although our definition of interaction is based solely on proximity to a vessel (i.e., distance), we assumed that albatrosses typically need to land on the water to engage in such interactions. In addition, we expected that, when interacting, birds would often take off and land repeatedly to keep pace with moving vessels. As a result, we hypothesized that these interactions would be associated with low average speeds but distinct patterns of speed variability across scales. To avoid complete redundancy with sinuosity and FPT, instead of calculating mean and standard error of speed within circles of radius r, we calculated them for various temporal scales (i.e. number of locations around the focal point, representing mean and standard deviation of speed from 30 min to 4 h before/after the focal location; respectively referred as Spd.mean.0.5 to Spd.mean.4 and Spd.sd.0.5 to Spd.sd.4 in Fig. 3). The standard deviation was calculated as the square root of the variance observed in each subsample of speeds.

Secondly, we also used two biotic metrics about individual birds, age and sex, as this species is known to have different strategies across their life and depending on their sex^22,35, which might affect patterns of movements when they interact with vessels. In our study population most individuals are of known age and sex so that we were confident the model could then be applied to any GPS tracks in our study population. However, this choice might be a limit if the model was to be applied to other study populations where such information might not be available.

Thirdly, we used as inputs some simple abiotic environmental variables associated to each location: bathymetry (as wandering albatrosses are known to change their behaviour along shelf edges, and fishing vessels in the area often operate over specific bathymetric features³⁶), day/night (night being defined as a local solar elevation below − 6° under the horizon) since albatrosses are mostly diurnal, and latitude as a proxy for the type of fisheries the bird can encounter (in the north and in oceanic waters they can encounter mainly tuna longliners, while around Crozet and Kerguelen, in neritic waters, more south, they encounter mainly toothfish longliners¹⁸). Because the use of these metrics as inputs reflected the specificity of our study system, we investigated how model performance varied with or without these parameters as inputs.

Base model design

We first pooled all our locations, and split it in two parts: a training and testing dataset, containing 80% of the data, and a validation dataset containing the remaining 20%. For a given random split of data, we balanced both training and validation dataset to have the same number of “0” and “1” vessel-interaction states in each (i.e. a 1:1 class balance). For this we kept all our “1” (80% in the training pool and 20% in the validation pool) and randomly sampled the same corresponding numbers of “0”. Down-sampling helps reduce bias toward the majority class in imbalanced datasets, particularly for rare events such as interaction events in our case, but comes at the cost of discarding real data, potentially reducing variability coverage and increasing the risk of model instability or overfitting.

The overall logic of this procedure is depicted in the upper part of Fig. 2. We report results from only one random split but repeated runs of different randomized splits gave very similar results, i.e. our results for this base model do not appear sensitive to the exact random partition chosen.

Fig. 2.

Scheme summarising the individual-partition model design. For the simple model design, 100% of the data went to the upper part of the scheme.

We used a random forest model with a cross validation of 10 folds with 3 repeats, which is for us a good compromise between robustness and computational cost, and to account for randomness in the model, and then validated it on the validation data set. To assess the fit we calculated 4 metrics: Out-Of-Bag error (hereafter OOB), accuracy, sensitivity and specificity, explained below.

OOB is an estimate of the model’s prediction error calculated using the subset of data not included in the bootstrap samples for training each tree.

Accuracy is the proportion of correct predictions (both true positives (TP) and true negatives (TN)) out of all predictions (out of TP, TN, false positives (FP) and false negatives (FN)). For a 2-state classification, an accuracy of 0.5 depicts a model that attributes states randomly.

Sensitivity is the proportion of true positives correctly identified by the model.

Specificity is the proportion of true negatives correctly identified by the model.

We tuned the model on three parameters: the number of trees (hereafter ntrees), the number of variables randomly sampled as candidates each split (hereafter mtry) and the minimum number of observations that must be in a terminal (hereafter node size).

Sensitivity analyses: filtering out or not part of the data for training and testing

In a second step, we tested whether model predictive power improved when the model was run on a pre-selection of locations more likely to be associated with interactions. For this we tried two additional models. The first included only locations by day (filtering out night locations), as albatrosses are expected to forage (and interact with vessels) mainly during the day⁴⁰. The second included only locations by day and with a mean local speed of less than 10 km h⁻¹ between the previous and the following locations, because albatrosses need to sit on water when attempting to feed⁴¹, resulting in much lower apparent speed when interacting with vessels⁴². For those two models, the filtering out of locations happened after the metric calculation and before the model training. The input metrics for the models, the training/testing splitting procedures and the performance statistics were otherwise the same as for the base model.

Sensitivity analyses: assigning individuals to either training or testing pools of data

To better understand how reliably our predictions could be extrapolated to actual new data from new individuals and conditions, we built a last model (referred as Individual-based partitioning model in further section) using a different, more controlled splitting procedure between training and testing data (Fig. 2). For this we split the total pool of individual birds (n = 156) in two parts: a training and testing pool of individuals, containing 80% of the different birds, and a validation dataset containing the remaining 20%. We then re-sampled our 80% training pool of individuals in two parts, a training data set containing 80% of the locations (from random individual origin), and a testing one containing 20%. We then applied the same method as before, and assessed the fit with two different accuracy/sensitivity/specificity scores: one for “known birds” (testing model predictions on the testing dataset from the training pool of individuals) and one for “unknown birds” (testing dataset from the testing pool of individuals). Only one OOB score is calculated for this model (referring to the training dataset from the training pool of individuals).

The whole analysis is run in R version 4.0.5⁴³. For random forest model building, we used randomForest⁴⁴ and caret⁴⁵ packages.

Results

The 193 tracks from 162 individuals represented 183 361 points at sea, of which 13 158 were classified as 1, i.e. overall 7.2% of albatrosses time at sea was spent close to an identified vessel. These interacting locations formed 1493 interaction events (i.e. series of interacting locations without more than 2 h interruption of interaction), from 162 individuals (mean duration 2 h 12 min ± 2 h 34 min).

Base random partitioning model

After model tuning, the best combination of nodesize, mtry and trees was respectively 1, 10 and 2000. With the parameters, the Out-of-Bag (OOB) error estimated was 0.16, indicating a relatively strong predictive power of our model. Overall accuracy of the model was quite high, with an accuracy of 0.83, sensitivity of 0.88 and specificity of 0.78, i.e. overall better predictions of interacting locations than non-interacting locations. In decreasing order of importance, the variables mostly used by the model were: large-scale sinuosity, age, 1-h mean speed, large-scale FPT, speed standard error, small- and mid-scale sinuosity and mean speed. Mid and small-scale FPT, bathymetry, sex and dayshift were less used for predictions (See Fig. 3 for more details about relative variable importance). This is consistent with the observed differences in behaviour observed between interacting vs. non-interacting locations (Fig. 4 presents the contrast in behaviour between interaction states for the most important variable of each category of movement variable, while ESM 2 includes all other trajectory-based variables). Corroborating this, when we removed every non-trajectory-based variable (i.e. bathymetry, age, sex and dayshift), in case this model was to be used on other datasets where such information was missing and/or irrelevant, the results were similar (OOB = 0.18; accuracy = 0.86; sensitivity = 0.86; specificity = 0.76). In addition, using latitude as a proxy for potential differences in interactions with different fleets didn’t improve accuracy. The two models filtering out either night locations or both high-speed locations and night locations showed the exact same characteristics as the main simple model, both in performance and in variable importance.

Fig. 4.

Comparison between the boat-related interaction state for the most important mode of each category of trajectory-based variable (Sinuosity, Speed and FPT; for every trajectory based-variable, see ESM 2).

Individual-based partitioning model

The model accounting for potential individual variability, when tested on “known birds” (testing dataset from the training pool of individuals), showed similar characteristics as the main base model, both in terms of variable importance and in performance statistics (OOB = 0.14, Accuracy = 0.86, sensitivity = 0.89, specificity = 0.83). When applied to “unknown birds” however, the model predictions showed a drastic drop in accuracy (0.72) through specificity (0.59), while the sensitivity stayed approximately the same (0.85). This sudden drop in specificity depicted an increase in the number of false positive, although the model still performed slightly better than chance.

In terms of true interaction events (i.e. genuine sequences of consecutive locations where the bird was associated with a vessel, grouped together into a single behavioural event), 60.5% were matched by at least one predicted interacting location. This proportion rose to 78.3% when considering only true events lasting more than 1 h (i.e., > 4 locations). The false positive predictions showed no clear patterns in relation to interaction events.

Discussion

Predicting seabird interactions with fishing vessels based solely on their GPS tracks would open several important applications. These include monitoring vessels at low costs from simple bird-borne locational sensors, potentially quantifying and mapping IUU activities when complemented by other vessel data sources¹⁵; and analysing the evolution of interactions in past and future seabirds’ movement data, to e.g. assess changes in attraction to vessels and potential natural rescue mechanisms to bycatch^24,26. Our baserandom partition model identified interactions with fisheries with good accuracy and specificity (i.e. >0.8) using only trajectory-based metrics, i.e. sinuosity, speed and FPT measured and contrasted at different scales around each location. This is extending previous exploratory results on albatrosses^29,30 suggesting actual predictive power. However, when applying the model to data from new individuals (i.e. not included in the training data set), our predictive power dropped from 0.86 to 0.72 for the global accuracy, and from 0.89 to 0.59 for the specificity. This is for now insufficient to extrapolate predictions of interactions with fishing vessels to cases where only GPS data is available, but promisingly it still provides slightly better predictions than chance. Below we discuss in more details how our results inform which methods to try in the future, in order to improve our capacity to predict fisheries from seabird tracking data.

The hypothesis defended previously was that predictions were limited by natural foraging mirroring vessel-associated foraging²⁸. This hypothesis could suggest to primarily seek for new or more accurate ways to contrast natural foraging and vessel-associated foraging. Additional behavioural details may be gained by increasing the frequency of acquisition of GPS locations and/or deploying other types of sensors: immersion sensors to better detail the alternance of sitting and flying behaviours (although this was not found to have a strong effect²⁸), microphones or proximity sensors to gain insights into the social environments likely to be more intense in aggregations around vessels^21,46,47, more accurate 3D data through altimeters, magnetometers and/or accelerometers^48,49, to test for differences in flying altitudes and/or take-off frequencies⁵⁰. All these ideas hold promise for using seabirds as allies in monitoring fisheries in the future. However, they would require new deployments (along with radar detectors and AIS for training) and therefore would not support retrospective analyses of past data. Additionally, they are likely to be more complex and costlier than simply using standard GPS data to predict fisheries activity based on movement metrics. Here, our base model suggests that it seems already possible to efficiently contrast natural from fisheries-associated foraging even with coarse GPS data alone (one location every 15 min). In addition, our results suggest that rather than the contrast, it is the variability of the interaction behaviour itself that is currently limiting progress: the same limitation might also apply when integrating other sensors. Our study thus suggest it is still worthwhile to address this question with only coarse GPS data.

A second avenue often suggested to reach higher predictive power is to improve the processing rather than the tracking data itself, by exploring other movement metrics. Yet we found here that FPT was not a determining factor when speed and sinuosity were already considered over various scales, suggesting to us that other combinations of these two-basic metrics (speed and sinuosity) are unlikely to provide major improvements in predictive power. We acknowledge however that even slight improvements in the accuracy, sensitivity or specificity scores could still be very useful, especially since our procedure report scores for sub-samples of 0 from the original dataset. Another way to improve the processing could be the use of other models. Convolutional Neural Networks (CNN), for instance, could eliminate the need to define metrics ‘subjectively’, by applying models directly to raw tracks⁵¹. CNN models generally require even larger datasets than RF to be able to provide robust results. The 196 tracks of this study for example are still probably too few, despite being quite large for seabird biologging studies. Moreover, the good performance of our base model compared with the drop-in performance of the individual-partitioning model suggest that our movement metrics already captured some of the important behavioural variation to identify interactions, but rather lacked generality across individuals or interactions. Poor performance of the individual-based partitioning model may stem from a small sample size and limited behavioural information across age and stages for example, as expected for phenomenon with high variability, which requires larger and better-targeted samples to be effectively captured—something our base model results suggest may still be achievable. It seems unlikely that we could overcome this lack of extrapolation power solely by applying CNNs to our data. Instead, we believe that progress would primarily come from increasing the total sample size of annotated data, with the aim of capturing the full range of variation in interaction behaviour⁵².

Our results also emphasize that the data sampling design on individuals is likely to be crucial to better account for individual variability. Variation in the time spent interacting and the proportion of encountered boats that were attended has already been documented in this species across populations of the Indian Ocean³⁴, but also within the Crozet population depending on individual characteristics like age or sex, and on contextual factors such as fishing activity or presence of cetaceans around vessels^35,40,53. However, these previous results did not explore resulting effects on movement patterns around vessels at different spatial or temporal scales. Trying to better describe and characterize this variability (e.g. continuous or categorical variations in movement patterns during interactions, intra- and/or inter-individual variations in interaction patterns) could help better define efficient machine learning model designs in the future. Many more explaining factors could also be advanced to predict part of the variability in interaction behaviour. These can be: (i) fisheries practices (e.g. type of fishing gears, targeted species, bait used, use of mitigation measure; as well as the individual behaviours and preferences of vessel captains^54,55 (ii) abiotic (local meteorological or oceanographic conditions^50,56, and/or (iii) biotic factors: other species/individuals attending the vessel changing the competition/dominance⁴⁷, availability of alternative prey options and individual body condition^19,57, constraints to relay the partner at the nest⁵⁸. Continuously monitoring all these variables is impossible in practice, and the fewer predictive variables required, the greater the potential for broadly applying the model to both new and historical data. Here, the few additional variables that we tested, chosen based on strong a priori expectations (e.g. bathymetry, latitude as a proxy for type of fishing fleet, age, sex), did not substantially improve predictive performance, likely because their effects were conditional on other unmeasured variables. Furthermore, variability could also stem from “cognitive” factors (memory, learning, personality^32,35,36), i.e. how individuals react to these environmental conditions. Rather than (or in addition to) exploring an endless range of possible contextual input variables, another strategy would be to better partition the variability, into within-individual (i.e. the same individual behaves differently across fisheries encounters and foraging trips), and between-individuals (i.e. individual specializations). For example, if individuals are specialized, extrapolating to new individuals will remain challenging regardless of how many contextual variables are included. In such case, one viable monitoring strategy could be to consistently track the same individuals over time. In our study, the lack of repeated observations per individual (due to limited battery size of early versions of the radar-detector devices) limited our ability to explore this partitioning, and we recommend giving particular attention to this aspect of sampling design in future efforts to use seabirds as monitoring allies.

Future attempts to identify fisheries from seabird movement patterns may also consider relying on different seabird species. In white-capped albatrosses (Thalassarche cauta) for instance, difference of sinuosity between natural and vessel-associated foraging appeared of greater contrast than those previously measured in wandering albatrosses^29,30. Moreover, beyond the question of contrast with natural foraging, the variability in interaction behaviour itself might be extreme in wandering albatrosses. In this species, some foraging trips extend thousands of kilometres to Antarctica or to subtropical latitudes while other trips can remain within 50 km of the colony^22,40; leading to high variability in the shape of resulting trajectories. It is this wide range that makes wandering albatrosses an excellent ally to monitor fisheries and IUU through radar detectors¹⁵, however, it may become a disadvantage when aiming to automate the recognition of archetypal movement patterns. It is not clear to what extent other wide-ranging seabirds strongly attracted to fisheries, such as white-chinned petrels⁵⁹ or individuals at other life stages (e.g. wintering birds⁶⁰); would exhibit less variability in their movements, while still providing broad monitoring coverage. Another option could be to begin with a less ambitious monitoring scope by focusing on smaller-range seabirds, such as mollymawk albatrosses, and/or on chick-rearing stages⁶¹. These birds may exhibit less variability in range and movement patterns, while also more readily allowing for repeated high-frequency GPS tracking of individuals⁶².

In conclusion, our study highlights how crucial it is not to neglect individual variability when trying to use animals as allies for environmental monitoring. There are many other contexts than monitoring fisheries with seabirds where individual variability matters^63,64. Machine learning tools are holding enormous promises but also open a labyrinthic amount of technical possibilities to navigate through (type of behavioural data collected, type of statistical model, explanatory variables included) that may easily mask the importance of the initial sampling design. Theoretical ecology has now largely recognized the widespread occurrence of individual specialization (i.e. persistent between-individual differences⁶⁵); and its importance for fundamental ecological and evolutionary processes⁶⁶. Progress in cognitive sciences also increasingly emphasize the scope for complex within-individual variability in behaviour⁶⁷. We thus recommend to carefully choose how individuals are sampled (repeat data on individuals, choice of individuals, etc.), because this cannot later be changed, in contrast with the type of models or environmental variables included therein. We have discussed here how this individual variability can also help inform all other methodological choices for incremental improvements of predictions, from the relevance of using more complex models to the consideration of different species as platforms for monitoring. We hope this will efficiently inspire future machine learning-assisted attempts to use seabirds as allies in monitoring fisheries but also more generally animals as environment sentinels.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

E.L.S., H.W., J.C. and M.C.: Study design; A.C., H.W., J.C. and A.P.: Data acquisition; E.L.S., A.C., A.P., J.C. and M.C.: Data processing; E.L.S.: Data analyses; E.L.S, J.C. and M.C.: writing and interpretation of the results; J.C.: administration and funding of the project. All authors edited and revised the manuscript. All authors approved the final version of the manuscript and agreed to be held accountable for the content therein.

Data availability

Data are available in figshare at: https://figshare.com/articles/dataset/Data_OCEAN_SENTINEL_Weimerskirch_et_al_2018_2019/10289096?file=20777529 (doi: 10.6084/m9.figshare.10289096⁶⁸.

Declarations

Competing interests

The authors declare no competing interests.