This repository describes the entire data processing and analysis workflows performed to estimate the distribution of crabeater seals in East Antarctica using a weighted species distribution model ensemble. This work is presented in the publication entitled: "Are high-resolution ocean sea ice model outputs a suitable source of environmental data to predict the distribution of species?: A case study of crabeater seal (Lobodon carcinophagus) in East Antarctica" published in TBA: DOI. The co-authors of this publication are Denisse Fierro-Arcos, Stuart Corney, Amelie Meyer, Hakase Hayashida, Andrew E. Kiss, Louise Emmerson, Colin Southwell, and Petra Heil.
The code included in this repository was developed by Denisse Fierro-Arcos.
Species distribution models (SDM) quantify the relationship between species presence of and environmental factors. They are often used to guide conservation management plans, but limited availability of environmental and biological data in undersampled regions, such as the Southern Ocean, represent an important challenge preventing us from accurately estimating species distribution.
The main aim of this study is to determine whether ocean models that have been validated from an ecological perspective are suitable to estimate the distribution of species in a data-poor region. Here we chose to use to work with ACCESS-OM2-01, a high-resolution coupled sea ice-ocean model, because of its ability to reproduce past conditions of ecologically relevant environmental factors, particularly sea ice extent (Fierro-Arcos et al 2023). We compared the predictive performance of a weighted SDM ensemble fitted with environmental data from two sources: remotely-sensed data products and ACCESS-OM2-01 model outputs.
We tested our methodology by estimating the distribution of crabeater seals (Lobodon carcinophagus) during November (late breeding period) and December (post-breeding period) in East Antarctica. Crabeater seals are the most abundant top predator in the Southern Ocean (Bengtson et al 2018) and have a highly specialised diet made up almost exclusively of Antarctic krill (Euphasia superba) (Bengtson et al 2018), (Southwell 2005), making them the one of the largest consumers of Antarctic krill (Moosa2024, Southwell 2005).
Our weighted SDM ensemble included two regression-based SDMs and two tree-based SDMs, respectively: Generalised Additive Model (GAM), (MaxEnt), Random Forests (RF), and Boosted Regression Trees (BRT). We weighted the contribution of each SDM in our ensemble based on their predictive performance.
Given that remotely-sensing data within the area and period of study were only available for a few of environmental variables that influence the distribution of crabeatear seals, we created two sets of environmental data from the ACCESS-OM2-01 model. One set matched the variables available in remotely-sensed data products (reduced ACCESS-OM2-01), and another including the full set of relevant environmental variables (full ACCESS-OM2-01). This allowed us to compare the performance of models trained with environmental data from remotely-sensed products and models, and to assess the impact of including a larger, and arguably more complete set of environmental predictors.
- Based on model performance metrics, SDMs trained with the full ACCESS-OM2-01 dataset or remotely-sensed products perform best, suggesting that SDMs fitted with these environmental datasets likely offer the most realistic predictions of crabeater seal distribution in the recent past.
- The weighted SDM ensemble made up of individually tuned SDMs outperformed individual SDM algorithms across all metrics (
$AUC_{PRG}$ ,$AUC_{ROC}$ and Pearson correlation). Machine-learning SDMs (RF and BRT) performed better than regression-based methods (GAM and Maxent). Maxent was the worst performing method and thus was excluded from weighted SDM. - Sea-ice related variables, including sea ice concentration and distance to the sea ice edge, were identified as key drivers of crabeater seal distribution. Matching results from previous attempt to uncover drivers of distributin of these animals.
- The distribution of Antarctic krill, their main prey, was only included in the full ACCESS-OM2-01 dataset. We found that when included, it was consistently identifed as one of the most influential variables of distribution. This suggests that informatin about prey is likely to lead to more realistic estimates of habitat distribution.
- We used the location of crabeater seal sighted while hauled out on sea ice, therefore our habitat estimates may not capture the full range of distribution of these animals.
- The environmental data products used here (particularly remotely-sense products) did not have a very high resolution (in the orders of meters). The resolution of environmental data matters as it influences the spatial patterns in environmental conditions that are captured. The resolution of these products should be sufficient to capture the spatio-temporal variability of environmental variables that influence the distribution of the animals of interest.
- Environmental data products used to trained SDMs should provide a true representation of environmental conditions within the spatial limits of the area of interest. This is true whether we are using remotely-sensed products or ocean model outputs.
- Prior to estimating habitat distribution for any species, an evaluation of the ability of an ocean model to realistically reproduce observed past environmental conditions within the area of interest is necessary.
- The performance of each SDM algorithm considered must be evaluated prior to developing an ensemble model, particularly if unweighted.
- We emphasise the importance of including a comprehensive suite of ecologically relevant environmental predictors in species distribution models, as a reduced set may not capture key drivers of distribution.
- Behaviour, competition, and predator-prey interactions are important for species distribution. Future work should consider the inclusion of interactions with major krill predators such as Antarctic baleen whales, to improve the accuracy of distribution predictions and identify the realised niche for crabeater seals.