Abstract: Species distribution models (SDMs1) are crucial for guiding management in a changing world. However, limited understanding of algorithm selection, ensemble weights and temporal transferability assessment undermines confidence in their predictions. Transferable predictive models, based on objective and proven selection criteria could therefore provide effective tools for defining species-environment relationships.
This study developed a framework for generating SDMs in the marine environment that improves models? temporal transferability. The methodological approach steps were: 1) Collection of predictors related to species ecology and their records and species grouping according to their ecological requirements. Twenty-one seaweeds were used as a case study. Environmental and distribution data were divided into two independent periods to evaluate temporal transferability. 2) A model for each species was built in each period with nine algorithms (Generalized Linear Model, Generalized Additive Model, Multivariate Adaptive Regression Spline, Mixture Discriminant Analysis, Classification and Regression Trees, Support Vector Machine, Flexible Discriminant Analysis, Random Forest, MAXENT) and projected into the other period. Predictor contributions to the final models were obtained. 3) Assessment of predictive performance for each model was made using the area under the receiver operating characteristic curve and true skill statistics metric for both models? accuracy and temporal transferability capabilities. All values were over 0.8 for all groups. In turn, the geographical pattern of all models were shown to be ecologically coherent.
The algorithms and their weights that fit best were used to generate transferable models over time in the marine environment and retained for each species. In general, machine learning algorithms produce models with higher sensitivity than regression-based approaches. This methodology sets the scene for further inquiries in the marine environment when developing consistent practices for model development and transferability.
Results are satisfactory for broad application in marine research, allowing a comparative framework between species predictions and facilitating the use of transferable models, especially in climate change studies across large areas. In addition, the proposed methodological approach is a cost-effective tool for dealing with a high number of species in marine environments. All data are freely available, so the methodology can be reproduced for marine researchers with different objectives.
Otras publicaciones de la misma revista o congreso con autores/as de la Universidad de Cantabria