Remote sensing of emperor penguin abundance and breeding success

Emperor penguins, *Aptenodytes forsteri*, are highly vulnerable to climate change due to their reliance on stable land-fast sea ice for breeding. Predicted rapid declines in Antarctic fast ice extent threaten the species, with projections suggesting a 90% colony loss by the end of the century. The remoteness and harsh conditions of their habitat have resulted in insufficient study of most colonies. Only two colonies, Pointe Géologie and Atka Bay, have extensive ground-truth population counts and detailed reproductive data. These colonies serve as crucial benchmarks for understanding the species' life cycle, including their arrival at breeding colonies between late March and early May, courtship, egg-laying (followed by female foraging trips), male incubation (64 days on average), chick hatching (July-August), and chick-rearing (until fledging between November and January). Satellite-based surveys, essential due to the inaccessibility of most colonies, provide the bulk of available data for population size estimation. However, even high-resolution imagery (0.3 m/pixel) offers only area-based estimates, not individual counts. The viability of area-to-abundance conversion is compromised by imaging uncertainties, area-to-number conversions, and large phenological occupancy fluctuations. These fluctuations arise from huddling behavior (influenced by temperature, wind, solar radiation, and humidity) and annual phenological patterns (linked to sea ice extent and prey availability). The polar night prevents satellite imagery during the incubation stage (June-July), leading to uncertainties in breeding pair estimates. Consequently, current satellite-based surveys are sufficient for gauging population size and long-term trends but not for assessing short-term factors impacting breeding success, unless a colony disappears entirely. This study aims to develop a method to compensate for these uncertainties and provide reliable estimates of annual breeding pairs and breeding success based on colony area measurements during austral spring and summer.

Existing literature highlights the vulnerability of emperor penguins to climate change, primarily focusing on sea ice decline and its impact on breeding success (Barbraud & Weimerskirch, 2001; Forcada & Trathan, 2009; Jenouvrier et al., 2014, 2019; Trathan et al., 2020). Studies have used satellite imagery to estimate population size (Fretwell & Trathan, 2020; Fretwell et al., 2012; Labrousse et al., 2021, 2023), revealing both the discovery of new colonies and challenges in accurately assessing population trends due to variability in colony occupancy and image resolution limitations (Labrousse et al., 2021). The impact of sea ice extent on prey availability and foraging behavior has also been investigated (Kirkwood & Robertson, 1997; Zimmer et al., 2008; Watanabe et al., 2012; Houstin et al., 2022), emphasizing the complexity of the relationship between environmental factors and penguin breeding success. The challenges of using satellite imagery for accurate population estimates are well-documented, with uncertainties arising from various factors, including imaging processes, conversion of colony areas to individual counts, and fluctuations in colony occupancy. The limited availability of ground-truth data for most emperor penguin colonies underscores the need for robust methodologies for analyzing remotely sensed data (Fretwell & Trathan, 2019). Previous studies established the use of a windchill model for predicting colony density based on environmental factors (Richter et al., 2018). This work builds upon previous research by developing a more comprehensive approach that integrates a phenological model with a refined windchill model to improve the accuracy of population estimates and breeding success assessment.

The study uses a three-step method to estimate emperor penguin abundance and breeding success from satellite imagery. First, a windchill model is used to convert colony area to individual counts. This model estimates colony density (animals per square meter) as a function of air temperature, wind speed, solar radiation, and relative humidity. The model parameters were estimated using data from Pointe Géologie and Atka Bay colonies between September and December, encompassing 538 measurements of colony area, individual counts, and meteorological variables. The windchill model's predictive power was assessed using R² and geometric error. Second, a phenological model describes how the number of individuals present in the colony varies throughout the breeding season (March 1 to February 28). This mechanistic model incorporates parameters representing arrival time, courtship duration, female absence duration, foraging trip duration, and chick-feeding periods. The model was benchmarked against ground-based individual counts from Atka Bay (3 breeding seasons) and Pointe Géologie (10 breeding seasons). Model parameters were estimated using a Markov chain Monte Carlo (MCMC) approach. Geometric errors and R² values were used to evaluate the model's accuracy in predicting adult and chick counts, as well as breeding success (fledging chicks relative to breeding pairs). Third, the phenological model is inverted to infer breeding pairs and breeding success from sparse counts obtained near the end of the breeding season. This method was benchmarked using data from ground-based and satellite-based images. Specifically, the methodology involved manual selection of colony-covered areas from ground-based images (using Clickpoints software) and calculation of occupied area. The windchill model parameters were estimated using correlations between colony density fluctuations and meteorological variables. The phenological model was developed based on known breeding activity patterns, fitted to observed data using an MCMC approach, and validated by comparing model predictions to observed numbers of individuals. The model includes parameters for arrival date, courtship duration, female absence duration, foraging trip duration, and chick-feeding periods. The model's ability to predict the annual breeding success (fledging chicks/breeding pairs) was assessed by comparing model predictions to observed chick counts. The application of the windchill and phenological models to satellite imagery involved using previously published colony area data, obtaining meteorological data, predicting colony density using the windchill model, multiplying this density by the colony area to obtain animal counts, fitting the phenological model to these counts, and finally deriving the number of breeding pairs and fledging chicks. The results were then compared to available ground-truth data. The study also compared model predictions of phenological events (arrival, female departure, etc.) to manual observations.

The windchill model, predicting colony density from meteorological variables, showed an R² value of 0.32 and an average geometric error of 39% for density predictions. However, when combined with measured area to estimate the number of individuals, the model achieved an R² of 0.93 and an average geometric error of 11%, with 60% of data points within the 1-sigma interval. The phenological model accurately captured the temporal fluctuations of weekly individual counts at both Atka Bay and Pointe Géologie, showing an average geometric error of 16% and an R² of 0.91 for all data points. While predicting the number of fledging chicks effectively (R² = 0.45 overall), it showed higher accuracy at Pointe Géologie (R² = 0.64) compared to Atka Bay (R² = 0.12). The model also showed good accuracy in predicting phenological event timing, though limitations in the frequency of ground-truth observations hindered a full evaluation of its ability to predict inter-annual variations. Significant correlations were found between model parameters and breeding success, particularly foraging trip durations during the crèching phase (R² values ranging from 0.39 to 0.55). Applying the combined models to satellite imagery data from Coulman Island, Atka Bay, and Stancomb-Wills Glacier in 2011, the study found good agreement between estimated and available ground-truth breeding pairs (average geometric error of 28%, R² = 0.88). Benchmarking with ground-based images from Atka Bay and Pointe Géologie also yielded good agreement (average geometric error of 21%, R² = 0.92, 74% of data within one standard deviation of the prediction). Analysis revealed a striking increase in breeding pairs at Atka Bay between 2008-2011 and 2018-2020, possibly due to immigration from a nearby colony that experienced breeding failure. Comparisons with previous studies showed discrepancies in estimated population sizes, attributed to improvements in methodologies accounting for phenological variations. Despite the good correlation between model estimates and ground truth, the model showed some difficulties predicting the timing of phenological events in years with unusual events, highlighting a need for more extensive data.

The developed method effectively addresses the challenges associated with estimating emperor penguin abundance and breeding success from remotely sensed data. The combined windchill and phenological models offer a significant improvement over existing methods, providing more accurate and robust estimates that account for the complex interplay between environmental factors and penguin behavior. The model's ability to extrapolate from sparse data is particularly valuable for monitoring remote colonies where ground-based surveys are limited. The findings highlight the importance of considering phenological variations when interpreting satellite imagery data, as neglecting these variations can lead to significant errors in population estimates. The strong correlations between foraging patterns and breeding success underscore the influence of environmental factors, such as sea ice extent, on reproductive success. The study’s success in predicting breeding success, using only information on adult penguin abundance, opens up possibilities for large-scale monitoring programs using satellite data. While the model performs well, further research is needed to refine the model parameters by incorporating a wider range of environmental data and incorporating variation in individual penguin condition. This research enhances our ability to monitor the health and resilience of emperor penguin populations and anticipate the effects of future climate change. The model can provide early warning signals of breeding anomalies that inform conservation efforts. The method's potential for large-scale application and integration with future high-resolution satellite data makes it a valuable tool for monitoring the health of the Southern Ocean ecosystem.

This study presents a novel methodology for estimating emperor penguin abundance and breeding success using a combined windchill and phenological model applied to satellite imagery data. The model accurately predicts breeding pairs and fledging chicks, accounting for phenological variations and improving upon existing methods. The results demonstrate the effectiveness of this approach, providing insights into population dynamics and breeding success. Future research should focus on validating the model across a wider range of colonies and environmental conditions, incorporating data on individual penguin condition, and exploring the integration of the model with advanced remote sensing techniques.

The model's accuracy is partially dependent on the accuracy of the input meteorological data. In some cases, meteorological data from weather stations may not fully reflect the local conditions at the colony site. Furthermore, the model's predictive power is influenced by the availability and quality of ground-truth data used for model fitting and validation. The relatively limited number of ground-truth observations, especially for some phenological events and the Atka Bay colony, may hinder precise inter-annual variation predictions. Finally, the model's assumptions, such as the fixed number of foraging trips, might not fully capture the nuanced variability of individual penguin behavior.