Predictive habitat mapping has shown great promise to improve the understanding of the spatial distribution of benthic habitats. However, although they surely represent an important step forward in process-based ecosystem management, their predictive efficiency is not always tested by independent groundtruthing data. This is particularly true for the deep-sea environment, where sample data are always limited compared to the large extent of the areas to be mapped. The aim of this study is to apply and test different spatial models to statistically predict the distribution of three Cold-Water Coral (CWC) species (Madrepora oculata, Lophelia pertusa and Dendrophyllia cornigera) in the Cap de Creus Canyon (NW Mediterranean), based on high-resolution swath-bathymetry data and video observations from the submersible JAGO (IFM-GEOMAR). Submarine canyons act as specific hosting areas for CWCs, owing to their favourable environmental conditions, which provide habitat and shelter for a wide range of species, including commercially viable fish. Maximum Entropy (MaxEnt), General Additive Model (GAM) and decision tree model (Random Forest) were independently applied to represent non-linear species-environment relationships using terrain variables derived from multibeam bathymetry (slope, geomorphologic category, rugosity, aspect, backscatter). Relevant differences between the three models were observed. Nonetheless, the predicted areas where CWCs should be found with higher probabilities coincided for the three methods when a lower spatial scale was considered. According to the models, CWCs are most likely to be found on the medium to steeply sloping, rough walls of the southern flank of the canyon, aligning with the known CWC ecology acquired from previous studies in the area. As a final step, a probabilistic predictive ensemble has been produced merging the outcomes of the three models considered, providing a more robust prediction for the three species. The main insight is that important discrepancies can arise in using different species distribution models, especially when high spatial resolutions are considered. This could in part be the result of the different statistical assumptions behind each of the models. We suggest that a more reliable prediction could be obtained by merging models into spatial ensembles, able to reduce differences and associated uncertainties, showing hence a strong potential as an objective approach in the planning and management of natural resources.
Conference or Workshop Item