Innovative use of depth data to estimate energy intake and expenditure in Adélie penguins Free

Energy is the fundamental currency shaping the life-history strategies of animals (Burger et al., 2019; Kressler et al., 2023; Pontzer and McGrosky, 2022). Individuals within populations acquire and spend energy to fuel processes such as movement, body maintenance and growth, thermoregulation and reproduction (Gower et al., 2008; Noakes et al., 2013; Steinhart et al., 2005). As energy acquisition from the environment is limited, individuals' performance and trade-offs in energy allocation directly impact life-history traits such as survival and reproduction, and, therefore, fitness and population dynamics (Brown et al., 2004; Mogensen and Post, 2012; Morano et al., 2013). In addition to fuelling all physiological processes, energy acquisition (i.e. foraging) itself often requires physical activity and thus energy expenditure (Pontzer and McGrosky, 2022). Environmental variations in biotic (e.g. prey availability) and abiotic factors, and animal internal states can affect how energy is spent and acquired, and lead to adjustments in foraging strategies (Byrne et al., 2022; Chevallay et al., 2022; Egert-Berg et al., 2021).

In order to understand individual trade-offs, it is important to consider both energy acquisition and expenditure within the same framework. Yet, simultaneously measuring energy intake and expenditure in wild animals is challenging. Numerous species forage in areas almost impossible to sample (i.e. deep oceans, remote land, sky), making observations difficult. Approaches in eco-physiology, such as doubly-labelled water (DLW), respirometry, heart-rate monitors or oesophagus temperature sensors, have proven useful in providing data to validate energy expenditure and intake (Froget et al., 2004; Hicks et al., 2020; Nagy et al., 1999; Ropert-Coudert et al., 2001). However, these approaches are costly, logistically difficult, invasive and, therefore, challenging to consistently use in long-term species monitoring programs. Given the difficulties in simultaneously measuring energy expenditure and energy acquisition in wild animals, information comes from different data sources (English et al., 2024), making long-term analysis or study comparisons difficult.

Recent advances in biologging technologies have allowed researchers to validate the use of accelerometers to accurately estimate animal behaviours and energetics across a wide range of taxa and habitats (English et al., 2024; Wilson et al., 2020). When validated with the DLW technique, for example, these devices allow estimation of activity-specific energy expenditure (Sutton et al., 2021; Hicks et al., 2020). By recording tri-axial body acceleration at high resolution (25–100 Hz) on species such as white sharks (Watanabe et al., 2019), southern elephant seals (Gallon et al., 2013) and penguins (Kokubun et al., 2011), accelerometers have also been used to detect putative prey capture attempts. Yet, because of their novelty, historical time series of accelerometer data are often not available.

In contrast, time-depth recorders (TDRs), developed in the 1980s, have been extensively used since the 1990s (Ropert-Coudert et al., 2009) to reconstruct dive profiles, investigate feeding behaviour and estimate energy expenditure of diving marine predators (Chappell et al., 1993a,b; Viviant et al., 2014). Despite lower accuracy, data collection via TDRs presents several advantages compared with data collected with accelerometer tags. TDRs record data at lower resolution (usually 1 Hz or coarser), reducing memory and battery consumption. They are not affected by their location on the body of the animal, unlike accelerometers, which need to be placed at or near the center of the mass of the animal to measure body acceleration. Because of their simplicity, TDRs can also be smaller and lighter than accelerometers for the same recording duration, which is ethically important to minimize the impact of data loggers on animals. Owing to their lower memory and power consumption, TDRs are also able to record data for longer time periods than accelerometers, which allows, for example, researchers to record multiple consecutive foraging trips on smaller animals. Practically, TDRs can be used to study the year-round diving and feeding behaviour of marine vertebrates, such as in Adélie penguins (Lescroël et al., 2023), emperor penguins (Houstin et al., 2022) and beluga whales (Storrie et al., 2022), which is still difficult with accelerometers as memory and battery consumption are very demanding. Moreover, the coarser data resolution, compared with accelerometer data, generates smaller datasets which require less analytical and computational power. These advantages are very important when working on long-term species monitoring programmes and when aiming to link individual behaviour to fitness (reproduction and survival) and, ultimately, population dynamics.

Despite its advantages and longer history, validations of energy expenditure and acquisition with TDR-derived parameters in the wild are rare (Chappell et al., 1993b; Elliott, 2016) compared with accelerometers. When no direct calibration with DLW is available, most studies quantified time–activity budgets from biologging data or direct observations and calculated energy budgets using activity-specific metabolic rates estimated via biomechanical and bioenergetic modelling (Croll et al., 1991; Stanley, 2002; Dunn et al., 2023). It has been shown that these methods tend to underestimate energy expenditure for flying and diving birds (Elliott, 2016), which advocate for the need to build these calibrations from experiments in the wild when possible. Historical TDR-derived proxies such as bottom phase duration and number of undulations within dives have been used as proxies of feeding activity (Bost et al., 2007; Lescroël et al., 2021; Viviant et al., 2014). Yet, validation of such metrics is rare (but see Watanabe et al., 2014), and recent papers tend to show that these depth-derived metrics alone do not effectively reflect the feeding activity of marine predators compared with accelerometers (Allegue et al., 2023; Brisson-Curadeau et al., 2021).

To address this knowledge gap, we focused on Adélie penguins, one of the most abundant Antarctic seabird species and ecosystem sentinel of the Southern Ocean (Barbraud et al., 2020; Forcada and Trathan, 2009), as our model species. This species mostly forages on two species of krill (Euphausia superba and E. crystallorophias) (Ratcliffe and Trathan, 2012), and is therefore highly dependent on sea-ice conditions (Kokubun et al., 2021; Michelot et al., 2020). Bottom phase duration and undulations have been extensively used to describe its feeding activity. Yet, the thresholds used to define these two parameters are often different across studies. Bottom phase duration is sometimes considered spanning from the first to the last time vertical velocity was <0.25 m s⁻¹ (Ropert-Coudert et al., 2007), sometimes it is considered to be below the 40% deepest part of a dive (Lescroël et al., 2021), and it sometimes spans from the first to the last undulation (Bost et al., 2007). Similarly, calculations of the number of undulations within a dive are derived either from changes in vertical velocity alone or with the latter in addition to different intensity thresholds of vertical velocity (Bost et al., 2007; Lescroël et al., 2021; Ropert-Coudert et al., 2001). Creating a simple, validated and objective framework to study feeding activity based on depth data could therefore ease study comparisons.

The Adélie penguin colony located on Île des Pétrels, Antarctica, has been extensively monitored since the 1960s. Since the mid-1990s, TDRs have also been regularly deployed, followed by accelerometers since 2016. Hence, only 8 years of accelerometry data are currently available, compared with 25 years for TDRs. During the 2018–2019 breeding season, breeding Adélie penguins were fitted with loggers recording both accelerometry and TDR data (Hicks et al., 2020). Importantly, DLW measurements were collected from these individuals, allowing the calculation of accurate data on energy expenditure (Speakman, 1997). We took advantage of this diverse ecological data collection to develop a framework allowing estimation of energy balance of Adélie penguins from depth data. We used DLW measurements and TDR data to predict energy expenditure from depth data only. We combined behavioural classification based on accelerometry (Chimienti et al., 2022) and the power of machine learning (Pichler and Hartig, 2023) to estimate feeding activity on solely depth data. We compared our results with other historical TDR metrics and accelerometers to answer the following research questions: (1) can depth data be used to predict energy expenditure of marine predators; and (2) can machine learning help to objectively estimate feeding activity of marine predators from depth data? Furthermore, as foraging (i.e. actively searching and hunting preys) is known to be costly for marine predators (Jeanniard-du-Dot et al., 2017; Elliott et al., 2013), we investigated the cost of fine-scale feeding (i.e. prey catching within foraging activity, hence the process that provides energy intake).

The study colony is located on Île des Pétrels, Antarctica, next to the Dumont D'Urville research station, in Adélie Land (66°40′S, 140°01′E). From 21 December 2018 to 11 January 2019, 58 breeding Adélie penguins [Pygoscelis adeliae (Hombron & Jacquinot 1841)] (24 females and 34 males) were tracked and monitored. All individuals were in their chick-guarding stage, where parents alternate mostly 1-day trips at sea to forage and feed their chicks (Ainley, 2002). Individuals were captured at their nest when both parents were present. To limit disturbance, we only captured one of the partners from each nest. We performed molecular sexing at Centre d'Etudes Biologiques de Chizé (CEBC) as previously described (Marciau et al., 2023) to confirm the sex of each individual a posteriori.

This study was approved and authorized by the ethics regional committee number 084. Permits to work in Antarctica were delivered by the Terres Australes et Antarctiques Françaises (TAAF) under the advice of Comité d'Environnement Polaire and Conseil National de la Protection de la Nature. All experiments were performed in accordance with the guidelines of these committees.

All statistical analyses and data manipulations were performed in R version 4.3.1 (https://www.r-project.org/).

Daily energy expenditure (DEE, kJ day⁻¹) was measured using the doubly labelled water (DLW) technique, as described in Hicks et al. (2020). Individuals were blood sampled from the tarsus vein upon capture to have a background sample. Then, they were weighed and injected with 0.3 ml DLW kg⁻¹ body mass into the pectoral muscle. Birds were individually marked for future identification and kept in an enclosure for approximately 2 h to allow DLW to equilibrate. Following that period, another blood sample was taken from the tarsus vein and data loggers (Axy-Trek, Technosmart, Italy, 40×20×8 mm, 14 g, less than 0.5% of individual body mass) recording tri-axial acceleration at 100 Hz and pressure at 1 Hz were deployed on the central back region of breeding Adélie penguins and secured using waterproof adhesive Tesa tape and two Colson plastic cable ties. Each bird was then released back to its nest and left for a foraging trip at sea.

Birds were recaptured at their nest after a foraging trip, or 3 days after DLW injection. Deployment duration ranged from 46.53 to 78.43 h with an average of 54.82±1.43 h. Trip duration (from first to last dive) ranged from 12.24 to 31.94 h with an average of 22.60±0.98 h. A new blood sample was taken from the tarsus vein and birds were weighed (see Supplementary Materials and Methods for detailed protocol).

Upon recovery, data were downloaded and processed using the R programming language. After calculating depth (±0.1 m) from pressure (Leroy and Parthiot, 1998), a custom function (see available code at https://github.com/bendps/tdr2nrj) was used to perform the zero-offset correction. Static acceleration was calculated by smoothing each axis over a 1-s period. Then, dynamic body acceleration (DBA) was calculated by subtracting static acceleration from the raw acceleration value. Vectorial DBA (VeDBA) was calculated as the square root of the sum of the squares DBA of the three axes (Qasem et al., 2012; Wilson et al., 2006).

To allow easier inter-study comparisons, we also report mass-specific DEE (kJ g⁻¹ day⁻¹), which was obtained by correcting DEE by the mass of the individual upon equipment.

Using a random forest (RF) algorithm, we aimed to identify periods of feeding in the diving behaviour of Adélie penguins. As a reference, we used the accelerometry-derived behavioural classification from Chimienti et al. (2022). This algorithm uses unsupervised machine learning to classify Adélie penguin movements into several behaviour, including putative feeding activity. As this classification was done at 25 Hz, we summarised it at 1 Hz. To ensure a conservative approach, we classified each 1 Hz data point as feeding only when at least half of the corresponding 25 Hz were labelled as feeding behaviour (Machado-Gaye et al., 2024).

To quantify larger-scale movement, we calculated the rolling means and standard deviations of vertical velocity and vertical acceleration over 5-s windows.

Before running the RF model, we performed variable selection to reduce its complexity. We filtered variables based on their correlation and variable importance measure (VIM, from the ‘Boruta’ R package; Kursa and Rudnicki, 2010). Whenever two variables were highly correlated (correlation >0.8), the one with the lowest VIM was excluded from the RF model.

The remaining variables were used in a RF built using ‘tidymodels’ (https://www.tidymodels.org/) and ‘ranger’ (Wright and Ziegler, 2017) R packages. We randomly selected and assigned half of the deployments to train the RF, while the other half was kept to test model performance. The train and test datasets had the same sex ratio. We trained the RF over six variables and parametrised it on different numbers of trees (i.e. 50, 100, 500, 1000). We also tuned the mtry parameter, indicating the number of variables randomly sampled as candidates at each split, between 1 and 6 for each number of trees. Finally, the training dataset was further split into training (75%) and testing (25%), and a 5-fold cross-validation procedure was performed on the final training data. The best model was selected based on two widely used metrics, namely, accuracy and area under the receiver operating characteristic (ROC) curve (AUC) (Cutler et al., 2007; Poisot, 2023).

Unless stated otherwise, means±s.e.m. are provided. All model assumptions were checked using the plot_model() function from the ‘sjplot’ (https://CRAN.R-project.org/package=sjPlot) R package.

To estimate energy expenditure from depth data, we modelled the DLW-derived DEE using multiple linear regression models with sex, mass upon equipment, time spent in each behaviour, and sum of vertical movement over each behaviour at the daily level. Following Hicks et al. (2020), we used the average VeDBA across foraging trip as our null model. We compared all model combinations using Akaike's information criterion corrected for small sample size (AICc). To estimate the prediction power of our model, we implemented a bootstrap procedure to limit the effect of our small sample size. Over 1000 iterations, the dataset was randomly separated in a train (50%) and a test (50%) dataset. After fitting the best model to the train dataset, we compared its depth-derived DEE with the DLW-derived DEE on the test dataset using root mean squared error (RMSE). For each coefficient, we used bootstrapping to calculate a 95% confidence interval to estimate model stability.

To estimate our accelerometry-derived proxy of feeding activity from depth data, we modelled at the daily level the accelerometry-derived proxy of time spent feeding at-sea (T_feeding) using several predictors: the RF-derived proxy T_feeding, and two historical proxies to quantify feeding intensity, namely, bottom duration (i.e. time spent in the 20% deepest part of a dive; Bestley et al., 2015; Carter et al., 2017) and number of undulations [i.e. any change in depth over a 1-s period, from a negative vertical velocity to positive or vice versa (i.e. descending to ascending or vice versa), during the bottom phase of a dive, as described by Lescroël et al., 2021]. Again, all model combinations were compared using AICc.

All models within a ΔAICc of 2 contained both time- and movement-based parameters, as well as body mass (Table 1). Sex was never retained in any of the best models. In comparison, the most parsimonious model performed significantly better than the best model using only time-based parameters (ΔAICc=2.70; Table 1) and the VeDBA model (ΔAICc=7.10).

The bootstrap procedure confirmed the robustness of our model to predict DEE from depth data (ρ=0.81 [0.66; 0.89], R²=0.66 [0.44; 0.81], RMSE=344.67 kJ day⁻¹ [264.48; 439.73], error rate=12.01 [9.22; 15.32]; Fig. 2) and was coherent with model coefficient calculated on the full dataset.

Considering mass-specific DEE as response variable gave similar results, and the best selected model contained the same behavioural variables (time spent diving and amount of vertical movement in the sub-surface phase) and also the sex variable (adjusted R²=0.70; Fig. S2, Table S2).

We trained a RF to detect accelerometry-derived feeding activity from depth-derived parameters only. During model tuning, accuracy and AUC plateaued when mtry reached 3 and ntrees 1000. With these parameters, the out-of-bag (OOB) error estimated the RF algorithm was 0.03, indicating the good predictive power of our model. In decreasing order of importance, retained variables were the 5-s rolling SD of vertical acceleration, 5-s rolling mean of vertical velocity, depth, dive duration, 5-s rolling mean of vertical acceleration, and vertical acceleration (Fig. S3). Overall accuracy of the model was high, with a balanced accuracy of 0.83, specificity of 0.95 and sensitivity of 0.72. When evaluating the fine-scale detection of feeding events proxies, we observed that the model tends to group short consecutive prey catching attempts (i.e. continuous period of feeding). Nonetheless, the overall feeding duration was very similar to the accelerometry-derived reference proxy and our RF algorithm detected assumed feeding events during the ascend phase of the dives (Fig. 3).

We investigated whether our RF algorithm could predict accelerometry-derived time spent feeding at-sea and compared it with historical proxies (bottom phase duration and number of undulations). Our RF efficiently predicted accelerometry-derived time spent feeding over a foraging trip from TDR data (R²=0.81, AICc=18.22). This model was significantly better at predicting the reference accelerometry-derived time spent feeding than bottom phase duration (R²=0.51, AICc=40.76) and number of undulations (R²=0.32, AICc=48.78; Table 2). Using multiple feeding proxies (e.g. predicted time spent feeding and bottom duration) did not significantly increase the predictive power of our model (Fig. 4; Table S3). All proxies were predicting time spent feeding better than the null model with trip duration only (R²=0.04, AICc=57.15).

On average, penguins spent 1.84±0.11 h feeding per trip (0.81±0.05 h day⁻¹). Over that same period, individuals expend on average 2958.89±67.92 kJ day⁻¹. Penguins spending more time feeding per day displayed comparatively higher DEE (estimate=413.67±172.44, P=0.02, adjusted R²=0.09; Fig. 5; Table S4). Mass upon departure from the colony was not retained in the final model (Table S4).

We present a novel method to quantify energetics (i.e. energy acquisition and expenditure) based on depth data recorded from diving marine predators. We demonstrate the reliability of this method using foraging trips from 48 Adélie penguins during the chick-rearing stage in the 2018–2019 breeding season.

Recordings from depth data provide information on animal movement in just one spatial dimension (y-axis), unlike accelerometry, which covers three spatial dimensions (x-, y- and z-axes). Furthermore, the temporal resolution of depth data is lower (1 Hz for TDR) compared with that of accelerometers (25 Hz or more). Yet, despite the lower spatial and temporal resolution, our results show that solely based on depth data, the machine learning algorithm RF can be used to identify proxies for fine-scale (1 Hz) feeding behaviour with good accuracy when trained from accelerometer data (accuracy=0.83).

Our RF model was able to predict the feeding pattern detected with the accelerometers, which has been shown to be a reliable proxy for prey capture attempts (Brisson-Curadeau et al., 2021; Kokubun et al., 2011; Schoombie et al., 2024). However, the model grouped short putative prey catching attempts in single feeding bouts, which was expected given the lower resolution of depth data. Therefore, further investigation (e.g. using video cameras for example) is needed to assess what is being caught during these prey catching attempts (Del Caño et al., 2021; Sutton et al., 2020). Historical proxies such as bottom duration or number of undulations were solely used to estimate proxies of feeding intensity without knowledge of when feeding was performed (Bost et al., 2007; Lescroël et al., 2021). Moreover, these historical proxies rely on the hypothesis that feeding is occurring mainly, if not totally, during the bottom phase of the dive (Deagle et al., 2008; Falk et al., 2000). Yet, accelerometry (Chimienti et al., 2022), oesophagus sensors (Ropert-Coudert et al., 2000) and video data (Del Caño et al., 2021; Sutton et al., 2020) have shown that penguins also feed during the ascent phase of their dive. Owing to their high resolution, accelerometry-derived feeding proxies have been shown to predict actual feeding events more accurately than historical TDR-derived proxies (Brisson-Curadeau et al., 2021). By training our RF on accelerometer-derived feeding proxies, our method was able to identify more precisely feeding activity within depth data (e.g. during the ascent phase of dives), which was not possible when using historical proxies. Moreover, the method presented in this paper presents the advantage of not needing manually set arbitrary thresholds to define behavioural modes such as bottom phase or undulations, allowing transferability to other diving marine predators. Yet, because RF is trained on a given set of data, applications of our current model to unseen data would need data inspection (and possibly re-training) to ensure that the variables used in the model fall within the range of the training dataset. Moreover, the 2 m threshold used in this study, which defines dives, may bias our feeding prediction as other penguin species are known to capture prey during dives less than 2 m deep (Gómez-Laich et al., 2018; Schoombie et al., 2024). Future high-resolution detailed analysis of accelerometer combined with video data could inform us on the importance of such feeding behaviour.

Overall, our method could likely be applied to other marine predators with TDR and accelerometer data availability to train species-specific RF algorithms. Indeed, the metrics used in this paper to estimate feeding are a simple and objective way of describing movement of diving predators. Using this method allows researchers to closely reproduce the accelerometry-derived feeding pattern from depth data only.

Contrary to historical proxies such as bottom dive phase duration and undulations, using this method would have the advantage to provide a simplified and unbiased comparison between studies. Yet, further investigations are needed to evaluate how our algorithm efficiently detects feeding for different types of prey, as foraging tactics and movement can change depending on the prey (Bowen et al., 2002). Krill is the main prey item of Adélie penguins, but they are also known to feed on fish or squid (Ratcliffe and Trathan, 2012). Therefore, our dataset is likely to contain feeding of prey other than krill. Hence, our algorithm might also be set to investigate feeding of other penguin species if the metrics derived from TDR data are similar.

Moreover, we found that DEE increased with the proportion of time spent feeding per day. Although significant, our time spent feeding only explained 9% of the observed variance. This important unexplained variability underlines that active foraging (i.e. searching and hunting prey), rather than fine-scale feeding, is energetically costly (Jeanniard-du-Dot et al., 2017; Yeates et al., 2007). Studying energetics as both intake and expenditure rather than focusing on expenditure will also allow to better understand the importance of individual foraging strategies and trade-offs. Therefore, even if our proxies for time spent feeding need further validation from verified feeding (e.g. camera loggers, oesophagus sensors), our innovative methodology should improve our ability to study individual energetic trade-offs at large spatial and temporal scales. Here, we identified a positive linear relationship between time spent feeding and DEE. Further studies could relate feeding duration to energy intake using mass gain measures and diet energy content estimation to better describe the extent of that relationship. Alternatively, validation of feeding detected through accelerometers, and prey identification using camera loggers (Sutton et al., 2020; Watanabe et al., 2019) combined with mass gain data over a foraging trip could link our depth-derived foraging proxy to energy intake.

With a R² of 0.68, our low-resolution TDR-based model provides comparable results to other DLW calibrations based on high-resolution accelerometers on marine and terrestrial species such as Adélie penguins (R²=0.75) (Hicks et al., 2020), little penguins (R²=0.78) (Sutton et al., 2021) and polar bears (R²=0.70) (Pagano and Williams, 2019). This equation could be applied to the numerous Adélie penguin long-term TDR data collected across Antarctica (Cimino et al., 2023; Lescroël et al., 2023; Riaz et al., 2020), thus offering a great opportunity to reliably investigate regional variations in energy budgets. In line with what was previously known (Hicks et al., 2020), we showed that assigning different calibration coefficients to different behaviours enhanced the predictive power of our model. Our model shows penguins mostly expend their energy while diving and transiting (i.e. T_dive and VM_subsurface; Eqn 4). Also, sex was not present in any of the best performing models, making our framework applicable in scenarios where individual sex is not known.

It is important to note that during the studied season, there was open water accessible next to the colony, and therefore, penguins did not have to walk long distances to access open water, reducing their energy expenditure (Watanabe et al., 2020). Therefore, our calibration might not reflect energy expenditure in areas and years when long walking periods are needed to access open water. Yet, the proposed model could still be relevant when focusing on the at-sea part of the foraging trips. Further investigations would be needed to assess whether our model can also be used to estimate at-sea energy expenditure during the other phase of the breeding period, incubation, during which foraging trips are typically longer (10–15 days) and diving behaviour is different with shallower and less frequent dives (Chappell et al., 1993a; Lescroël et al., 2023). Yet, in another study, Chappell et al. (1993b) found that the field metabolic rate of Adélie penguins was not significantly different in incubation and chick rearing.

Historically, TDR was the first type of logger deployed on wild animals (Kooyman, 1965). The described method can easily be applicable to other marine predators, with often longer time-series available for TDR data compared with accelerometers, allowing the study of long-term trends. With environmental changes affecting energy availability (Duncan et al., 2015), these long-term estimations of energetics are crucial to deciphering how changing environmental conditions will affect individual life history strategies. When undertaking such analysis, it is crucial to ensure that the RF model is trained on data that accurately represents the long-term range for all the parameters (e.g. dive depth, dive duration, etc.). Even if the metabolic rate of Adélie penguins is thought to be stable across the range of sea temperature they encounter (Ainley and Wilson, 2023), applying our DEE equation to different environment (e.g. higher temperature) would need further investigations especially considering changes in thermoregulatory costs (Grunst et al., 2022). Alternatively, development of bioenergetic and biomechanic models would be a valuable extension as they would allow further exploration of theoretical limits of energy expenditure and allocation across environmental conditions (Chimienti et al., 2020; Gallagher et al., 2021; Goedegebuure et al., 2018; Sibly et al., 2013).

In conclusion, we show that lower-resolution TDR data can be used to estimate energetics similarly to accelerometers. Our results demonstrate how the application of machine learning approaches allows researchers to re-analyse datasets and more accurately predict energetics of diving marine predators compared with historical methods. Feeding activity is expected to be related to mass gain (Lescroël et al., 2021). Therefore, with prior knowledge of a species’ diet, calibrating depth-derived feeding proxy with mass gain could allow us to refine this simple framework by estimating energy intake directly. Because it is based on the widely used TDR data, this framework could be applied to several long-term marine species monitoring programs. This would allow researchers to study how extrinsic (e.g. environmental variations) and intrinsic (e.g. body condition) variables impact energetic trade-offs in individuals.

We thank the Institut Polaire Francais Paul-Emile Victor (IPEV program P1091) for its financial and logistical support. We thank the ‘Service d'Analyses Biologiques du CEBC’ for its expertise, and especially Cécile Ribout for her technical help in conducting laboratory analyses.