Breeding phenology plasticity to climate revealed by automated monitoring of two Eudyptes penguin species across the sub-Antarctic Authors: Samanta Dodino, Ignacio Juárez Martínez, Alice J. Edney, Klemens Pütz, Melina Barrionuevo, Esteban Frere, Ana Millones, Annick Morgenthaler, Petra Quillfeldt, Katharina Reusch, Andrea Raya Rey, Tom Hart Published: 2026 ABSTRACT The timing of breeding is a key component of a species’ realised niche and has been shown to be flexible to environmental change. Using a network of time-lapse cameras around Patagonia and the sub-Antarctic, along with citizen science assisted AI data extraction, we demonstrate that semi-automated extraction of phenological parameters is feasible without extensive human presence at colonies. Our comparative analysis of two related penguin species across the Polar Front reveals shifts in breeding phenology linked to environmental variation. The Southern Rockhopper Penguin showed a potential contraction of its breeding season, with consistent arrival but earlier departures in recent years. In contrast, Macaroni Penguins arrived and departed earlier, suggesting a temporal shift rather than a shorter season. These phenological changes correlated with winter-integrated SST rather than late-winter conditions, while reduced SST during pre-moult periods extended pre-moult duration and delayed migration. Despite marked phenological shifts in one species, breeding success remained stable, suggesting that timing changes may reflect flexible responses to environmental shifts rather than hidden survival costs. Species-specific differences in baseline phenology and temporal trends suggest that while some seabirds show phenological plasticity to ocean warming, stable breeding success may mask subtle but consequential impacts on sub-Antarctic population dynamics.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-54931-0. Introduction Climate change has altered the timing of seasonal activities for many species, but the capacity to adjust their phenology varies1–4. Many organisms rely on predictable environmental cues such as photoperiod to regulate key life-history events, which potentially could constrain their responses under rapid environmental change (i.e. temperature)5–7. Large-scale assessments across marine, freshwater, and terrestrial systems show that phenological responses differ systematically among trophic levels, with secondary consumers generally advancing more slowly or exhibiting lower sensitivity to climate change than primary producers and primary consumers8,9. This differential response increases the likelihood of temporal mismatches between interacting trophic levels. Seabirds, as long-lived marine predators, may be particularly sensitive to such mismatch. In these species, environmentally driven changes in breeding conditions have been linked to variation in survival and reproductive performance10. Evaluating shifts in breeding phenology and breeding success is therefore essential for understanding how climate-driven environmental change may influence demographic outcomes in marine predators11. The South Atlantic Ocean and the adjacent Southern Ocean are experiencing measurable warming trends12, with major implications for upper-trophic level predators such as penguins. These species often have limited capacity to adjust rapidly to environmental shifts, potentially leading to mismatches between optimal foraging areas and prey availability13–15. In the genus Eudyptes, long-term environmental variability has been linked to changes in foraging behaviour, energy acquisition, and breeding outcomes16–19. For instance, Southern Rockhopper Penguins (Eudyptes chrysocome) have shown altered winter dispersal patterns in response to oceanographic anomalies20, while Macaroni Penguins (Eudyptes chrysolophus) may face reduced reproductive success when foraging under suboptimal conditions21. Although the specific pathways remain unclear, these studies highlight the potential for changing ocean conditions to disrupt key life-history stages, emphasizing the need for long-term studies of phenology and reproductive performance22. Southern Rockhopper and Macaroni Penguins are both currently listed as Vulnerable due to global population declines23. These species breed in different regions of the South Atlantic Ocean/Southern Ocean. Southern Rockhopper Penguins breed on temperate and sub-Antarctic islands north of the Antarctic Polar Front (APF), between 46°S and 54°S24, while Macaroni Penguins breed primarily on oceanic islands south of the APF, between 55°S and 65°S25. This spatial separation is maintained in part by oceanic fronts like the APF, which can act as semi-permeable biogeographic barriers limiting dispersal and contributing to ecological and genetic divergence26. Both species exhibit strong philopatry to their breeding sites27,28 and are believed to show fidelity to foraging zones associated with frontal systems29,30. As a result, each species predominantly exploits different inter-frontal zones, characterized by distinct water masses, prey communities, and climate-related stressors31,32. Their high site-fidelity and apparent consistency in foraging strategies implies potentially limited behavioural plasticity in response to rapidly changing environmental conditions33. However, evidence from isotopic and genomic studies suggests that Southern Rockhopper Penguins can exhibit occasional dispersal and niche shifts under unfavourable conditions, which may facilitate colony founding and modulate population connectivity34. Tracking data also suggest some plasticity in spatial distribution and habitat use, showing some seasonal overlap, with individuals occasionally crossing fronts during the non-breeding period30,35. As both species face ongoing population declines, understanding how regional oceanographic regimes influence their breeding phenology and breeding success is essential for assessing their vulnerability and adaptive capacity in a warming South Atlantic Ocean/Southern Ocean. Time-lapse cameras offer a low-cost, non-invasive method to monitor the phenological and breeding behaviour of colonial seabirds36–42. By reducing the need for repeated field visits, remotely operated camera network minimizes human presence at colonies, limiting potential disturbance to breeding individuals while enabling extensive monitoring programs across species’ ranges in remote and difficult conditions with minimal researcher effort38,42,43. However, the substantial volume of datasets generated by the camera network presents significant analytical challenges. In this context, advancements in technologies such as artificial intelligence (AI) (e.g., Pengbot44 and citizen science initiatives (e.g., Penguin Watch45: www.penguinwatch.org, Seabird Watch: www.seabirdwatch.org) have substantially increased the speed and scalability of data processing42, allowing large datasets to be analysed efficiently while maintaining minimal disturbance to study populations. In this study, we investigate whether environmental changes influence penguins through modifications of their breeding behaviour at the colony, focusing on both phenology and breeding success. By including the geographic representation of the species’ breeding locations, we ensure that observed temporal trends are representative of their responses to environmental change and allow the detection of landscape-level patterns across much of their distribution. We evaluated the interannual variability in phenology and breeding success of two Eudyptes congeners, the Southern Rockhopper and Macaroni Penguins at ten sites across the South Atlantic Ocean/Southern Ocean. Using time-lapse cameras and analyses combining AI (Pengbot44) with citizen science (Penguin Watch), we investigated phenological and breeding success patterns in relation to environmental conditions (air temperature, sea surface temperature, sea surface anomalies and chlorophyl-a). We hypothesized that seasonal and colony-specific differences in phenology timing and breeding success are driven by environmental conditions (with potential links to prey availability). Specifically, we expected differences in start/end dates of the breeding season and breeding success under different environmental conditions across colonies and seasons. We had no a priori directional hypotheses regarding the environmental predictors. There is currently no unified expectation for how penguin species in general, and these focal species and colonies in particular, respond to environmental variability. Previous work has shown that both unusually cold and unusually warm SST anomalies can negatively affect adult survival19, indicating that departures from typical conditions may be detrimental. Additionally, spatial differences in SST have been linked to variation in foraging range and maximum travel distances, with higher SST around colonies associated with longer foraging trips, yet without consistent effects on breeding success46. Therefore, rather than predicting that specific environmental conditions would uniformly advance or delay phenology, we evaluated how breeding timing, the duration of breeding stages, and reproductive performance vary across environmental gradients in order to identify potential optimal conditions and quantify the sensitivity of Southern Rockhopper and Macaroni Penguins breeding parameters to oceanographic variability across latitudes. Materials and methods Study area and camera deployment Field work was conducted between 2013 and 2024 at ten sites across the South Atlantic Ocean/Southern Ocean : seven at Southern Rockhopper Penguin colonies north of the Polar Front and three at Macaroni Penguin colonies on South Georgia, south of the Polar Front (Fig. 1, Table S1). Southern Rockhopper Penguin cameras were deployed at: Isla de los Estados, Falkland Islands/Islas Malvinas (Beauchene Island, Bleaker Island, Hummock Island/Isla Montículo, Pebble Island/Isla Borbón, and New Island/Isla Goicoechea), and at Isla Pingüino (see Table S1 for more details). Macaroni Penguin cameras were deployed at Hercules Bay, Rookery Bay and Cooper Bay (see Table S1 for more details). Fig. 1Camera network from Southern Rockhopper (Eudyptes chrysocome) and Macaroni (Eudyptes chrysolopus) Penguins’ colonies at the southwestern Atlantic Ocean. Dots represent camera locations for Southern Rockhopper (violet) and Macaroni Penguin (green) colonies. Southern Rockhopper Penguin colonies (south to north): (1) Bahía Franklin, Isla de los Estados; (2) Beauchene Island, Falkland/Malvinas Islands; (3) Bleaker Island, Falkland/Malvinas Islands; (4) New Island/Isla Goicochea, Falkland/Malvinas Islands, (5) Hummock Island/Isla Monticulo, Falkland/Malvinas Islands; (6) Pebble Island/Isla Borbón, Falkland/Malvinas Islands; (7) Isla Pingüino. Macaroni Penguin colonies (south to north): (8) Cooper Bay, South Georgia; (9) Rookery Bay, South Georgia; (10) Hercules Bay, South Georgia. The dashed line represents the Antarctic Polar Front (APF). Map was made with Natural Earth and ETOPO147 using the R package ‘marmap’48. Camera network from Southern Rockhopper (Eudyptes chrysocome) and Macaroni (Eudyptes chrysolopus) Penguins’ colonies at the southwestern Atlantic Ocean. Dots represent camera locations for Southern Rockhopper (violet) and Macaroni Penguin (green) colonies. Southern Rockhopper Penguin colonies (south to north): (1) Bahía Franklin, Isla de los Estados; (2) Beauchene Island, Falkland/Malvinas Islands; (3) Bleaker Island, Falkland/Malvinas Islands; (4) New Island/Isla Goicochea, Falkland/Malvinas Islands, (5) Hummock Island/Isla Monticulo, Falkland/Malvinas Islands; (6) Pebble Island/Isla Borbón, Falkland/Malvinas Islands; (7) Isla Pingüino. Macaroni Penguin colonies (south to north): (8) Cooper Bay, South Georgia; (9) Rookery Bay, South Georgia; (10) Hercules Bay, South Georgia. The dashed line represents the Antarctic Polar Front (APF). Map was made with Natural Earth and ETOPO147 using the R package ‘marmap’48. Time-lapse cameras (Reconyx Hyperfire HC500) were installed at the Eudyptes colonies at the beginning of the first accessible breeding season (depending on logistical constraints) and following the methodology proposed by Jones et al.49. Once deployed, cameras remained in place at the colonies and operated year-round. Cameras were generally programmed to capture images at regular hourly intervals covering daylight hours, with some sites extending coverage to 24 h depending on logistical constraints. The cameras were checked at least once per year, generally between November and December (subject to site-specific logistical constraints), to retrieve the images stored on memory cards and to replace the batteries. Data classification via artificial intelligence (Pengbot) and via citizen science (Penguin Watch project) To estimate the amount of penguins per image in such a large-scale dataset we used a computer vision ConvNet called Pengbot44(https://github.com/CarlosArteta/penguin-counter). Pengbot estimates an amount of penguin per pixel so these can be summed to produce a total amount of individuals per image. An example of input images, output matrices and counts are given in Appendix 1 (Figure S1). However, this AI was originally trained on Antarctic penguin colonies, resulting in some noise and false positives at sites where Eudyptes penguins breed among vegetation (tussock) (e.g. Hummock/Montículo Island). In those cases, the estimation was carried out using Penguin Watch (www.penguinwatch.org), a citizen science project which allows volunteers from around the world to view images and mark penguin presence. Currently, this project has over 80,000 registered volunteers who help tag adults, chicks, eggs, and “others” (e.g., predators, exotic species, etc.). Each image is analysed by ten volunteers before being retired from view, to validate the classification criteria. To group the xy coordinates of multiple volunteer clicks and produce a single “consensus click” per individual we applied hierarchical agglomerative clustering49 using the ‘dbscan’ package in R50. Finally, to extract the metadata from each photograph (dates and times, camera id, temperature), we used the ‘exiftoolr’ package in R version 4.4.051. Phenology estimation We estimated four key phenological events from colony attendance patterns: (1) breeding season arrival, (2) pre-moult start, (3) pre-moult end, and (4) winter departure. We used the median number of penguins per image per day to estimate phenology milestones from changes in penguin colony attendance. For each colony and season, baseline mean and standard deviation were calculated from predefined reference periods. Reference periods were defined a priori based on species-specific breeding phenology and known periods of minimal or stable colony attendance24,25, and were kept constant across years. To estimate “arrival”, we used overwinter presence as a baseline. For Southern Rockhopper Penguins, the winter reference period corresponded to 1 July–15 September, and for Macaroni Penguins, it corresponded to 1 July–15 October. Arrival was defined as the first day on which the daily median exceeded the winter mean plus three standard deviations, sustained for three consecutive days. A threshold of three standard deviations was chosen for arrival because of the low numbers and variance of overwinter counts, reducing the risk of false positives from short-term fluctuations. For further assurance we required the arrival date to be the first of three consecutive days above that threshold. If fewer than 20 daily median values were available within the winter reference period, arrival was recorded as missing, as a minimum sample size was required to obtain a representative and statistically stable estimate of the baseline mean and variance. For pre-moult period detection, reference periods corresponded to 15 January-15 March for Southern Rockhopper Penguins and 15 February-15 March for Macaroni Penguins. “Pre-moult start” was defined as the first day in which median values dropped to or below the mean plus two standard deviations of the reference periods. “Pre-moult end” was estimated as the first day the median values exceeded the mean plus two standard deviations. Again, these estimates needed to be the first of three consecutive days meeting the requirements. A lower threshold (two standard deviations) and a smaller minimum sample size requirement (> five daily values) were applied for pre-moult because colony attendance during this period exhibits sharper and more sustained changes. These pronounced shifts allow reliable detection with a less conservative threshold and reduce sensitivity to small baseline sample sizes. For colony “winter departure” estimation, reference periods corresponded to 1 April-30 June for Southern Rockhopper Penguins and from 1 May-30 June for Macaroni Penguins. Departure was estimated as the first day in which daily median presence dropped below the mean of the reference period plus three standard deviations. As with arrival, we required at least three consecutive days meeting this condition to confirm the event. Similar to arrival, the higher threshold for departure was applied to reduce false positives due to high day-to-day variability in late-season counts. If fewer than 15 values were available, departure was recorded as missing to avoid unstable variance estimates. A minimum of 15 daily values was required to estimate the departure baseline, reflecting the higher counts and variability during this period compared to overwinter conditions, whereas arrival required 20 days due to the lower magnitude of winter counts. For phenology analyses we considered each species separately to avoid masking species-specific patterns, as breeding schedules differed between species. Southern Rockhopper Penguin phenology Due to the logistical difficulties of accessing these remote colonies for equipment maintenance and data retrieval, together with occasional weather-related camera failures, there were various years for which Southern Rockhopper Penguin colonies lacked data (see Table S1). To avoid dataset gaps, we established four “colony groups”: Isla de los Estados (IDLE), Southeastern Falkland/Malvinas Islands (Beauchenne and Bleaker Islands) (MFSE), Northwestern Falkland/Malvinas Islands (Hummock Island/Isla Montículo, Pebble Island/Isla Borbón, and New Island/Isla Goicoechea) (MFNW), and Isla Pingüino (PMIP). Phenological events were analyzed using generalized linear mixed models (GLMMs) with a Gamma distribution and a log link, as dates deviated from normality and displayed a right-skewed distribution with a mean–variance relationship. Each phenological event was modelled separately, with the day of the season (defined as days since 1 July) as the response variable, colony group and season as fixed effects, and camera identity as a random effect. Random effects were specified for camera identity to account for potential pseudo-replication. Preliminary analyses including colony as an additional random effect revealed negligible variance and no improvement in model fit, so only camera identity was retained. Pairwise comparisons between colony-groups were performed for Southern Rockhopper Penguins using estimated marginal means with Tukey-adjusted p-values (‘emmeans’ package52). Macaroni Penguin phenology Gaps in Macaroni Penguins data prevented within-colony comparisons so we treated all three colonies as a single colony group (South Georgia) (GEOR), assessing only seasonal differences. Macaroni Penguin phenological data showed no significant deviation from normality (Shapiro–Wilk test: p = 0.32; residuals approximately normally distributed based on inspection of Q–Q plots), so linear mixed models (LMMs) were used. The models included the day of the season for each phenological event as the response variable, season as a fixed effect, and camera identity as a random effect. As with Southern Rockhopper Penguins, we fitted separate models for each phenological event. Model selection For all models (Southern Rockhopper and Macaroni Penguins), we used an information-theoretic approach (MuMIn package, version 1.47.8,53) to compare model performance. The final model was selected based on the corrected Akaike’s Information Criterion (AICc), prioritizing the model with the lowest AICc and the highest Akaike weight (wi). Model selection was based on AICc and Akaike weights (wi), with models having ΔAICc ≤ 2 considered equally plausible54. While the statistical significance of individual parameters was not used to choose models, confidence intervals excluding zero were used to assess the biological relevance of explanatory variables. For all analyses, candidate models were first ranked using AICc and Akaike weights, and biological relevance was subsequently assessed only for the best-supported model. 95% confidence intervals were examined to determine whether explanatory variables had meaningful effects. To quantify the relative contribution of random effects, we calculated the proportion of total variance attributable to camera identity from the variance components of each mixed model. This variance partitioning allowed us to assess the relative importance of within- versus among-colony variability in phenological timing. Despite differences in model families between species (GLMMs with Gamma distribution for Southern Rockhopper and LMMs with Gaussian errors for Macaroni Penguins), the analytical framework—including fixed and random effects structure and separate modelling of each phenological event—was applied consistently across species. Finally, we visualized the phenological events of each site using circular time plots (‘ggplot2’ package55; functions geom_bar() and coord_polar()), which provide an intuitive representation of the penguin annual cycle. Bars indicate the relative frequency (proportion) of observations occurring at each time of year (scaled 0–1 within sites), rather than raw counts. Normalization accounts for the unbalanced sampling effort among sites, allowing direct comparison of phenological patterns even where the number of observations differed. Breeding success estimation Breeding success was estimated manually due to limitations in chick detection by both AI and Penguin Watch volunteers. The small size and low-contrast colours made chick identification challenging. Breeding success is a camera-based metric estimated as the number of surviving chicks divided by the number of active nests in view in a given season. Studies in chick mortality indicate that chick loss happens predominantly during incubation and early guard stage56,57. The following developmental stage, crèche, is therefore used in on-site studies to count surviving chicks and estimate productivity or breeding success. This is impossible to recreate from cameras reliably as our view of the colony is limited by the camera frame and tussock grass. Chick survival might be underestimated here if or when crèches are occurred outside the portion of the colony in frame. Conversely, chick survival can be overestimated if crèches in frame but are populated with chicks coming from nests outside the camera view. For this reason, we devised ‘walkout’ as a metric that can indicate chick independence as well as reliably linking the surviving chicks to the nests in view. Walkout was defined as the first observed event of chick independence from their parents, either because both adults left the chicks unattended for the first time or because the chicks are seen just outside of their nest area. After the first walkout event chicks become increasingly more independent and perform walkouts more frequently until all chicks are about and the previous colony structure disappears. We assume the first event of chick separation to be a marker of their ability to function independently (i.e. thermoregulate and move about). No reduction in the number of chicks was observed at any nest in the period after walkout where they still go back to their nest to sit and be fed so we believe this metric is equivalent to the survival until crèche. We evaluated differences between colony groups using LMM with breeding success as the response variable, colony group and years as the explanatory variables and camera identity as the random effect (to account for possible pseudo-replication). A single model was fitted for both species; species was not included as a fixed effect because each camera monitored only a single species, preventing separation of species and camera effects and potentially causing singularity issues. Models fitted separately for each species showed the same pattern, but splitting the data reduced statistical power. Following the principle of parsimony, we therefore present the results from the combined model. Combining species thus represents a trade-off between statistical power and species-specific inference; results allow robust estimation of overall patterns, but subtle species differences should be interpreted with caution. Model selection followed the same information-theoretic framework described in the phenology section, with AICc used to rank candidate models and confidence intervals used to assess biological relevance. Variance partitioning was calculated from the random-effect variance components to estimate the contribution of camera identity. Environmental drivers of breeding success and phenological variation between colonies To identify environmental drivers of breeding success and phenological variation across colonies, both species were analyzed jointly, irrespective of species-specific phenological schedules. This design enabled the identification of common environmental influences on congeners with contrasting breeding chronologies. We used the following oceanic variables: sea surface temperature (SST), sea surface temperature anomalies (SSTA), and chlorophyll-a (Chl-a) as a proxy for primary productivity. Additionally, we used camera-recorded temperature at each colony to describe at-colony temperature conditions. Satellite images for SST and SSTA were obtained from NOAA (https://pae-paha.pacioos.hawaii.edu/erddap/griddap/dhw_5km.html), and Chl-a data were downloaded from NASA-OCEAN DATA COLOR (https://oceandata.sci.gsfc.nasa.gov/api/file_search). To model the relationship between phenological events and environmental variables we used a generalized additive mixed model (GAMM, package ‘mgcv’ in R58) with a Gamma distribution, as phenological event dates deviated from normality. To model breeding success, we used Gaussian distribution. GAMMs are well-suited for ecological data as they allow for flexible, non-parametric modelling of complex relationships, particularly when predictor effects are expected to be non-linear59. Prior to modelling, environmental covariates were checked for collinearity, with a cut-off criterion of r = 0.6 for inclusion in the same model. All the models included camera identity as a random effect. Models were fitted using restricted maximum likelihood (REML), which estimates smoothing parameters via penalized likelihood and helps prevent overfitting. The effective degrees of freedom (edf) of each smooth term, reflecting the complexity of the fitted relationships, are reported to allow assessment of model flexibility. We selected the best model according to AIC. For breeding success, we generated raster layers representing the average values of SST, SSTA, and Chl-a during the chick-rearing period. For Southern Rockhopper Penguins, we considered December and January, while for Macaroni Penguins, we focused on January and February (representing most of the chick stage period). Environmental values were extracted within a 50 km radius from the colony, reflecting the maximum foraging distance reported for these species during this stage24,60–63. Additionally, temperature data recorded by time-lapse cameras at the colonies were used to estimate average temperatures during the chick-rearing period, allowing for an integrated assessment of both at-sea and on-land environmental conditions. We found that SST and SSTA were correlated with Chl-a (r = 0.6) and SST with camera-recorded temperature (r = 0.8). Considering that Chla- provides a better proxy of marine productivity relevant to penguin foraging success64, we retained Chl-a and on-land temperature in the model and excluded SST and SSTA to prevent multicollinearity. The model included breeding success as the response variable, with on-land temperature and Chl-a as smooth terms. To understand variation in arrival dates, we considered the environmental conditions during the at-sea distribution of penguins throughout the entire winter period and during the last month of winter as potential predictors of arrival timing. By analysing both periods, we aimed to assess whether arrival dates are influenced by cumulative environmental conditions over the entire winter (long-term pattern) and/or by immediate, short-term conditions experienced in the final month before arrival. For Southern Rockhopper Penguins, we generated mean rasters for the entire migration period using monthly data from April to September, and for the last month using only September. For Macaroni Penguins, we used data from April to October for the full winter period and only October for the last month. We extracted the average value of each environmental variable, considering published data on each colony’s winter distribution28,60,65–67. For Isla de los Estados, Isla Pingüino and Hummock Island/Isla Montículo, we had access to the full raw tracking dataset, which allowed us to extract values using a 4 km buffer around the raw points, considering (i) all positions recorded during the winter period and (ii) only those from the last migration month. For colonies without raw tracking data, we rasterized published distribution maps28,65 into shapefiles using the kmeans() and rast() functions of the ‘terra’ package and we extracted environmental values according to the shapefiles. From these shapefiles, we extracted average values for the full winter period and, to approximate the last-month distribution, we refined the shapefiles based on the maximum distance from colonies observed in the raw datasets during the final migration month (~ 1000 km). We ran one model using winter migration values and another using last-month values. We found a correlation between SST and SSTA for the whole winter period and the last month (r > 0.90 in both comparisons). Consequently, for each period, we ran one model including arrival dates as the response variable, with SST and Chl-a as smooth terms, retaining SST as the temperature metric due to its more direct interpretability. For phenological variation in departure, we considered the environmental conditions during the at-sea distribution throughout the pre-moult period and the at-colony conditions during the moult. By including these periods, we aimed to assess whether departure dates are influenced by cumulative environmental conditions over the pre-moult period (medium-term pattern) and/or by immediate, short-term conditions experienced in the last month during the moult before departure. For the pre-moult period, we considered the environmental conditions within an 800 km radius around each colony, based on an average foraging range for both species estimates from previous studies (Southern Rockhopper Penguins:35,68; Macaroni Penguins:27,69,70). Temperature data recorded by time-lapse cameras at the colonies were used to estimate average temperatures during the moult period. We found that SST and SSTA were correlated (r = 0.99). Consequently, we ran a model including departure dates as the response variable, with SST, Chl-a and at-colony temperature as smooth terms. For phenological variation in the pre-moult period, we used the generated raster of the pre-moult trip described above. We evaluated the pre-moult duration (differences between arrival and departure dates) in relation to the environmental variables. We found that SST and SSTA were correlated (r = 0.99). We ran one model with SST and Chl-a as explanatory variables. Results Southern Rockhopper Penguin phenology For all Southern Rockhopper Penguins phenological events, the best-supported models consistently retained colony group as a fixed effect (Table S2), indicating that most phenological variation was structured at this level (Figs. 2 and 3). Year was retained for winter and pre-moult departures, but not for breeding arrival. Temporal trends varied by event: the negative estimate for year in the winter departure model indicates that departures occurred earlier in more recent years, whereas the positive estimate in the pre-moult departure model suggests that pre-moult departures occurred later in recent years (Table S2). Variance partitioning indicated that camera identity contributed negligibly to variation in breeding arrival and winter departure (less than 0.001% of the total variance in both cases), indicating that these events were structured almost entirely at the colony-group level. In contrast, camera-level effects were more pronounced during the pre-moult trip stages, accounting for 8.8% of the total variance in pre-moult trip departure and 1.9% in pre-moult arrival. These results suggest that, while differences among colony groups account for most of the variation across all stages, additional heterogeneity associated with camera identity may occur, particularly during the pre-moult period. Regarding colony-groups contrasts, there was a significant phenological split between the pair Isla de los Estados-Southeastern Falkland/Malvinas Islands (IDLE-MFSE) and Isla Pingüino-Northwestern Falkland/Malvinas Islands (PMIP-MFNW) (p < 0.01), with no significant within-pair differences (IDLE-MFSE: p = 0.82; PMIP-MFNW: p = 0.99). Arrival occurred approximately 8 days earlier and departure approximately 36 days earlier for IDLE and MFSE colonies ([Mean day of the season ± SD]: Arrival= October 2nd ± 1.1 days; Departure= March 20th ± 12.2 days, Table S3) compared to PMIP and MFNW colonies (Arrival= October 10th ± 1.4 days; Departure= April 25th ± 7.3 days, respectively). For pre-moult “start” and “end”, pairwise comparisons indicated significant differences between IDLE and all other colonies (p < 0.001 for all comparisons, Table S3). IDLE individuals left for the pre-moult trip around 10 days earlier ([Mean day of the season ± SD]: February 1st ± 0.5 days) and came back to moult around 20 days earlier (February 20th ± 0.9 days) compared to MFSE, MFNW and PMIP (Pre-moult start= February 21st ± 1.5 days; Pre-moult end= March 13th ± 8 days) (Fig. 3). Macaroni Penguin phenology For Macaroni Penguins, the best-supported models retained year as a fixed effect for colony arrival and winter departure (Tables S4 and S5), indicating that timing of these events has shifted across years. Negative estimates for year in both models indicate that, in more recent years, arrivals and departures occurred earlier. In contrast, year was not retained in models for pre-moult start and end dates, suggesting no significant temporal trends for these events (Table S4 and S5). Variance partitioning showed that camera identity explained a substantial proportion of the total variance for several events. For winter departure, 90.0% of the total variance was explained by camera effects, indicating substantial heterogeneity across locations. Similarly, 72.7% of the variance in pre-moult trip departure and 57.0% in pre-moult arrival were attributed to differences between cameras. These results suggest considerable within-area variability in the timing of these stages, particularly for winter departure and pre-moult events. Macaroni arrival occurred on average November 2nd ± 2.6 days (Fig. 2) and ended on average April 20th ± 8.5 days (Figs. 2 and 4). Pre-moult trips started on average March 3rd ± 3.5 days and ended on average March 20th ± 5.9 days (Figs. 2 and 4). Fig. 2Phenological events of the Southern Rockhopper Penguins (Eudyptes chrysome) breeding at Isla de los Estados, southeastern and northwestern Falkland/Malvinas Islands and Isla Pingüino, and Macaroni Penguins (Eudyptes chrysolopus) breeding at South Georgia. Circular bars show the relative distribution of observations over time (scaled 0–1 within colony groups). This normalization accounts for years with missing data among colonies and enables clearer comparison of phenological patterns across groups. Phenological events of the Southern Rockhopper Penguins (Eudyptes chrysome) breeding at Isla de los Estados, southeastern and northwestern Falkland/Malvinas Islands and Isla Pingüino, and Macaroni Penguins (Eudyptes chrysolopus) breeding at South Georgia. Circular bars show the relative distribution of observations over time (scaled 0–1 within colony groups). This normalization accounts for years with missing data among colonies and enables clearer comparison of phenological patterns across groups. Fig. 3Median adult presence of Southern Rockhopper Penguins (Eudyptes chrysome) at several study colonies across all seasons available. Data is derived from the citizen science consensus (Penguin Watch) and artificial intelligence (Pengbot) classifications. Here we only show only one northern (Hummock/Montículo Island) and one southern (Bleaker Island) colony for the Falkland/Malvinas as example. Median adult presence of Southern Rockhopper Penguins (Eudyptes chrysome) at several study colonies across all seasons available. Data is derived from the citizen science consensus (Penguin Watch) and artificial intelligence (Pengbot) classifications. Here we only show only one northern (Hummock/Montículo Island) and one southern (Bleaker Island) colony for the Falkland/Malvinas as example. Fig. 4Median adult presence of Macaroni Penguins (Eudyptes chrysolopus) at two colonies across all available seasons. Data is derived from citizen science consensus clicks (Hercules Bay) and artificial intelligence (Rookery Bay) classifications. Median adult presence of Macaroni Penguins (Eudyptes chrysolopus) at two colonies across all available seasons. Data is derived from citizen science consensus clicks (Hercules Bay) and artificial intelligence (Rookery Bay) classifications. Breeding success Model selection identified the null model as the best-supported model (Table S6), indicating that neither colony identity nor year explained a substantial portion of the variance in breeding success (Fig. 5). Variance partitioning showed that 33% of the total variability was explained by differences among cameras, while 67% remained unexplained. Fig. 5Breeding success of Southern Rockhopper Penguins (Eudyptes chrysome) breeding at Isla de los Estados, southeastern and northwestern Falkland/Malvinas Islands and Isla Pingüino, and Macaroni Penguins (Eudyptes chrysolopus) breeding at South Georgia across the seasons. Breeding success of Southern Rockhopper Penguins (Eudyptes chrysome) breeding at Isla de los Estados, southeastern and northwestern Falkland/Malvinas Islands and Isla Pingüino, and Macaroni Penguins (Eudyptes chrysolopus) breeding at South Georgia across the seasons. Environmental drivers of breeding success and phenological variation For breeding success, model selection indicated that the best-supported model included only summer Chl-a (Table S7). The analysis revealed a significant non-linear effect of Chl-a (edf = 2.27, p < 0.001), while summer SST had no significant effect (edf = 1.05, p = 0.30). Breeding success increased with summer Chl-a up to ~ 5–6 mg/m³, after which the effect plateaued (Fig. 6). Colonies in low-productivity areas (e.g., Isla de los Estados, South Georgia) exhibited lower success, while those in more productive waters (e.g., North-Falkland/Malvinas, Isla Pingüino) performed better. The plateau should be interpreted cautiously due to limited data at the upper end of the chlorophyll-a range. Variance partitioning showed that Chl-a accounted for ~ 33% of the total variance in breeding success, with the remaining variance unexplained by the model. Fig. 6Smoothed effect of Chl-a from the chick-rearing period on breeding success of Southern Rockhopper and Macaroni Penguins as estimated by a GAMM model. The solid line represents the estimated smooth function, and the shaded area indicates the 95% confidence interval. Smoothed effect of Chl-a from the chick-rearing period on breeding success of Southern Rockhopper and Macaroni Penguins as estimated by a GAMM model. The solid line represents the estimated smooth function, and the shaded area indicates the 95% confidence interval. For arrival phenology, the best-supported model included only sea surface temperature (SST) during the entire winter migration period as a predictor (Table S7). This model revealed a significant non-linear effect of SST (edf = 6.34; p < 0.001), with a U-shaped relationship indicating that earlier arrivals occurred at intermediate SST values (2–5 °C), while both colder (< 2 °C) and warmer (> 6 °C) temperatures delayed arrival. This suggests the existence of a thermal optimum associated with earlier return timing (Fig. 7, left panel). Variance partitioning showed that SST accounted for 93.6% of the total variance in arrival dates. For the model based on SST during the final month of the winter migration, the best-supported model also included SST as the sole predictor (Table S7). A significant non-linear effect was again observed (edf = 2.68; p < 0.001), although the U-shaped relationship was less pronounced. Earlier arrivals were associated with SST values around 3–5 °C, while lower (< 1 °C) and higher (> 5 °C) temperatures were linked to later arrival dates. Variance partitioning showed that SST accounted for 50.2% of the total variance in arrival timing (Fig. 6, right panel). Fig. 7Smoothed effect of SST from whole winter (left) and the last month of winter migration (right) on arrival dates of Southern Rockhopper and Macaroni Penguins as estimated by a GAMM model. The solid line represents the estimated smooth function, and the shaded area indicates the 95% confidence interval. Smoothed effect of SST from whole winter (left) and the last month of winter migration (right) on arrival dates of Southern Rockhopper and Macaroni Penguins as estimated by a GAMM model. The solid line represents the estimated smooth function, and the shaded area indicates the 95% confidence interval. For colony departure phenology, the best-supported model included SST and Chl-a during the pre-moult trip (Table S7). The model identified significant complex non-linear effect of both SST (edf = 5.39, p < 0.001) and Chl-a (edf = 6.85, p < 0.001) on departure timing (Figure S2). In contrast, the median on-land temperature during moult did not have a significant effect (edf = 1.00, p = 0.79) and was not retained in the final model. Despite these significant effects, the model explained only 11.7% of the variation in departure dates, indicating that other factors likely contribute to variability in colony departure timing. For pre-moult duration, the best-supported model included both SST and Chl-a (Table S7). Both predictors showed significant non-linear effects on pre-moult duration (SST: edf = 2.28, p < 0.001; Chl-a: edf = 8.54, p < 0.001). The effect of SST was moderately non-linear and generally decreasing across the observed range suggesting that higher SST levels were associated with shorter pre-moult durations, whereas Chl-a exhibited a more complex response pattern (Figure S3). Despite these significant effects, the model explained only 1.6% of the variation in departure dates, indicating that other factors likely contribute to variability in pre-moult duration. Discussion Using a combination of time-lapse cameras, AI-based image analysis and citizen science contributions, we were able to estimate the phenology and breeding success of Southern Rockhopper and Macaroni Penguins across multiple colonies in the South Atlantic Ocean/Southern Ocean. A standardized monitoring framework across years and sites allowed us to compare patterns among colonies and to assess spatial and species-level variation in responses to environmental variability. Temporal comparisons across years revealed phenological trends, such as shifts in timing of arrival and departure, which can serve as early indicators of climate-driven behavioural responses. Meanwhile, other breeding milestones such as pre-moult trips remained stable, which could be a show of resilience or increase the possibility of a trophic mismatch. Our results indicate that breeding success increases with rising chlorophyll-a concentration, up to a point where it plateaus at higher levels. Similarly, SST exhibited non-linear effects on key phenological events, with extreme SST values linked to later arrivals and departures. These environmental indicators provide insight into species’ ecological thresholds and preferences, and may inform predictions of population responses under future climate scenarios. Examining species-specific phenology, Southern Rockhopper Penguins showed significant inter-colony variability but high consistency in colony arrival and pre-moult end timings across years. However, the latest years in study showed a significant shift to earlier pre-moult start and winter departures, suggesting a possible contraction of their breeding season. Consistency in their arrival dates across years could suggest their migratory timing could be influenced by endogenous circannual rhythms. This aligns with findings in other long-distance migrants (e.g., skuas), where migration schedules remain fixed despite changing environmental conditions at breeding sites71, possibly due to strong internal regulation. In contrast, Macaroni Penguins displayed a different seasonal pattern: later years were linked to earlier colony arrivals and departures, though no seasonal effect was observed during the pre-moult period. These findings underscore the contrasting phenological responses between the two species. Macaroni Penguins exhibit inter-annual variability in arrival dates, suggesting a more flexible migratory strategy. While krill is their main prey, their diet al.so includes amphipods and fish—especially when krill is scarce72–74. This dietary plasticity is matched by behavioural plasticity: in poor foraging years, they increase trip duration and range63. Such adjustments may also influence migration schedules, allowing individuals to modulate arrival based on foraging conditions. Similar patterns have been documented in other migratory seabirds (e.g., Cape Petrels, Royal Penguins, Common Terns, Black-browed Albatrosses3,4,75,76. Inflexible behaviour, i.e. an inability to adjust arrival timing in synchrony with environmental change, could lead to a mismatch between peak resource availability and key reproductive stages, such as egg-laying or chick rearing. Phenological adaptation in seasonal environments is a fundamental life-history trait77, and failure to track shifts in food availability, especially under rapid climate change, may result in reduced breeding success or population declines78. Although we did not find a change in arrival dates in Rockhoppers, other phenological shifts have been documented. For instance, plasticity in the post-arrival phase (e.g. shortening the pre-laying period) might partially compensate for mismatches, but only under low pre-breeding mortality risk79. Dehnhard et al.80 observed for Southern Rockhopper Penguins in New Island that clutch initiation date advanced under high Southern Annular Mode, reflecting colder environmental conditions and higher food availability. However, that study also mentioned that the range of phenotypic plasticity appeared low compared to other seabirds’ species, including penguins. In contrast, Pygoscelid penguins have shown far greater plasticity in arrival dates, advancing their settlement phenology rapidly in response to climate change22. Where such compensation is not possible, persistent mismatch may threaten population resilience. Oceanographic variability can influence penguin wintering grounds by altering food availability. For example, in Magellanic Penguins, changes in the strength of the Río de la Plata plume affect prey distribution and timing of reproduction81. In Southern Rockhopper Penguins, however, consistency in arrival dates at colonies appears to be influenced more by local prey availability and environmental conditions during the reproductive period than by conditions in the wintering areas. Although for many colonies, only a single year of published tracking data was available, or the available data correspond to years different from those in which phenological observations were collected, it is likely that winter conditions varied across the years. For instance, penguins at Isla de los Estados concentrate their activities in highly productive waters with specific oceanographic conditions (e.g., SST, Chl-a;20). However, interannual variation in oceanographic conditions indicated that colder and more variable winters were linked to broader foraging and higher subsequent breeding success20. At Isla Pingüino, Southern Rockhopper Penguins maintain a consistent oceanographic niche (temperature and chlorophyll-a) but shift their spatial distribution in years with positive sea surface temperature anomalies, resulting in greater dispersal across the Argentine Shelf67. Taken together, these findings suggest that spatial fidelity during the winter may vary regionally and depends on oceanographic conditions but this do not influence arrival dates at colonies. Environmental factors play a key role in shaping migratory phenology75,82,83. In our study, phenological events and breeding success were influenced by environmental factors, which aligns with our expectation of variations across different environmental conditions. Sea surface temperature (SST) emerged as the primary predictor of inter-colony variation in arrival timing for both penguin species (Fig. 6). Colonies experiencing intermediate winter SSTs—such as Isla de los Estados and the southern Falkland/Malvinas Islands—tended to arrive earlier, whereas those in low winter SST (e.g., South Georgia) or high winter SST waters (e.g., Isla Pingüino, northern Falklands/Malvinas) showed later arrival dates. This spatial pattern may be consistent with a thermally mediated foraging efficiency mechanism, where moderate SSTs may enhance prey availability or capture success, allowing individuals to meet energetic thresholds for breeding sooner. Supporting this, previous studies (e.g., Carroll et al.84) have shown that foraging success in penguins peaks at intermediate SSTs and declines at thermal extremes. It is worth noting that the effect of SST on arrival is stronger if considering the whole winter, rather than just the last month prior to arrival (R² = 0.936 vs. 0.502). This indicates that Eudyptes penguins may likely respond to environmental conditions integrated over the full non-breeding season, rather than to proximate conditions prior to colony arrival. Also, the last-month model exhibited a weaker U-shaped response and failed to accurately capture the early arrival of southern colonies such as Isla de los Estados, which experienced moderate SSTs (~ 4 °C) but arrived among the earliest. Moreover, the curve appears disproportionately shaped by colder colonies like South Georgia, where low SSTs are consistently linked with delayed arrivals. This may have biased the model, leading to convergence in predicted arrival dates for southern colonies that differ markedly in observed phenology. Penguins may rely on broader-scale environmental integration to time migration, rather than reacting solely to immediate pre-arrival thermal cues. For Southern Rockhopper Penguins from Isla de los Estados and the Falkland/Malvinas Islands, survival probability has been shown to follow a U-shaped relationship with sea surface temperature (SST) anomalies, with reduced survival associated to pre-moulting period observed during both positive and negative deviations from average SST19. For colony departure phenology, both SST and Chl-a during the pre-moult period showed significant non-linear effects on departure timing. However, the model explained only 11.7% of the variance, indicating that these environmental predictors, while statistically significant, are not sufficient to account for variation in departure dates. We also found no effect of air temperature during the moult departure timing, suggesting that terrestrial thermal conditions do not constrain this event. Nevertheless, other unmeasured variables—such as wind, precipitation, or colony microhabitat— may influence departure phenology85,86 and should be considered in future work. The limited explanatory power of our departure model highlights the need to explore additional ecological and behavioural mechanisms. One possibility is that carry-over effects—such as earlier or later arrival influencing moult onset, or variability in the duration of the pre-moult foraging trip—may affect departure timing. These sequential dependencies across the annual cycle are not captured by SST or Chl-a alone. These hypotheses remain largely unexplored for Rockhopper and Macaroni Penguins and merit further investigation. Regarding pre-moult duration, we found that higher SST during the pre-moult trip were significantly associated with shorter time spent at sea, suggesting a potential link between favourable oceanographic conditions and foraging efficiency prior to moult. In contrast, the relationship with Chl-a was more complex and did not show a clear pattern, indicating that the influence of productivity on pre-moult duration may not be straightforward. Whitehead et al.70, reported that Macaroni and Eastern Rockhopper Penguins from Marion Island undertook longer pre-moult trips, when reduced productivity may limited prey availability, highlighting that environmental conditions can influence trip duration. They also showed that, across the species’ range, pre-moult trip duration varies considerably, with shorter trips at colonies closer to cooler waters (e.g., South Georgia) and longer trips at more distant colonies (e.g., Crozet and Kerguelen), suggesting that the distance to favourable foraging areas and local oceanographic features are important determinants of trip length. Furthermore, penguins concentrate foraging within productive Antarctic waters where mesoscale activity and eddies enhance prey availability, indicating that SST and productivity are only part of a complex suite of factors shaping pre-moult behaviour. Despite the significant effects of SST and Chl-a, these variables explained only a small fraction of the variation in pre-moult duration, whereas colony identity accounted for much more, indicating that local or colony-specific factors may be stronger drivers of pre-moult behaviour. This underscores the complex interplay between oceanographic conditions and pre-moult timing. Additional environmental or biological factors, not captured in this analysis, may therefore also play a role in shaping pre-moult timing. Future research should aim to identify and include these potential drivers (e.g., mixed layer depth as indicator of prey distribution) and the mechanisms by which this translates into changes in trip duration/phenology events (e.g., penguin body condition) to gain a more comprehensive understanding of pre-moult duration in penguins. The findings of this study have important conservation implications, as species-specific differences in phenological responses could suggest varying sensitivities to environmental change. The relative stability of breeding success indicates some resilience, but the potential contraction of the Southern Rockhopper Penguins breeding season and shifts in Macaroni Penguin timing could impact population dynamics in the long term. Meta-analyses have shown that single-brooded seabirds tend to shorten their breeding seasons as climate warms, which can lower lifetime reproductive potential and resilience to environmental variation87. Moreover, studies on Magellanic Penguins (Spheniscus magellanicus) demonstrate that delayed and compressed phenology leads to shorter chick-rearing periods, poorer fledgling condition, and ultimately reduced recruitment88, effects that may similarly impact Southern Rockhopper or Macaroni Penguins if seasonal mismatches increase. Consequently, even if mean breeding success remains unchanged year-to-year, these phenological shifts could undermine population stability over time. Our study demonstrates that large-scale monitoring using a combination of AI and citizen science is not only feasible but effective for detecting fine-scale phenological variation across remote seabird colonies. This non-invasive approach enables long-term, high-resolution monitoring while minimizing human disturbance, providing a powerful framework for tracking climate-driven changes and informing conservation strategies. However, some limitations remain, including temporal gaps that could be addressed through continuous monitoring of the same colonies. Although cameras are deployed year-round, in some cases data gaps occur because extreme environmental conditions can occasionally cause cameras to fall or malfunction. Exploring strategies to minimize these interruptions would help secure more continuous datasets. Spatial gaps also persist, particularly for Macaroni Penguins, where only one site has been studied. Differences between colonies may also exist across latitudinal gradients. In addition, variance partitioning revealed contrasting levels of within-colony variability across species: while camera effects were negligible for most phenological dates in Southern Rockhopper Penguins, they explained a large proportion of variance in Macaroni Penguins, particularly for winter departure and pre-moult events. This indicates that some phenological parameters may be more sensitive to fine-scale heterogeneity, highlighting the value of replication within colonies and broad geographic coverage. Future efforts should focus on overcoming these limitations to provide more comprehensive insights into the long-term dynamics of these vulnerable species. Conclusion This study highlights the effectiveness of AI and citizen science for large-scale monitoring of marine sentinel species, revealing distinct phenological shifts in Southern Rockhopper and Macaroni Penguins. Southern Rockhopper Penguins showed a potential contraction of their breeding season, showing inflexibility in arrival but departing earlier in later seasons, while Macaroni Penguins arrived and departed earlier, suggesting a temporal shift rather than a shorter season. Despite these changes, breeding success remained stable, indicating some resilience, and suggesting that population trends may be more strongly influenced by overwinter survival. Species-specific responses could indicate varying sensitivities to environmental change, underscoring the need for continued monitoring to inform conservation strategies for these vulnerable species. By providing novel phenological and breeding data for Southern Rockhopper and Macaroni Penguins, our study adds fine-scale information that complements global assessments of penguin populations89. Supplementary Information Supplementary Material 1 Supplementary Material 1