Assessing the reliability of species distribution models under changing environments: A case study on cetaceans in the North-East Atlantic

Abstract

1. Species Distribution Models (SDMs) assume stable relationships between species and their environment from which predictions are made. These relationships are likely to vary with changing environments, and predictions might depend more on modelling choices than on empirical data. Reliability assessments of predictions are necessary to support policy-making. 2. We identified environmental extrapolations among potential predictions of cetaceans’ distribution from 2005 to 2020 in the North-East Atlantic and calculated the percentages of calibration data with similar environments (nearby data), supporting these predictions. Thus, the assessment of reliability is generic, as evaluated before model fitting. 3. Predictions on continental shelves were extensively supported by the calibration data and were more reliable throughout the year than predictions on continental slopes and abyssal plains, which were more supported in summer. Predictions off Portugal were particularly uncertain due to the lack of surveys in this region of deep, warmer waters with seamounts. 4. The high effort between May and July led to a southern winter shift of nearby data, following the decrease in temperatures. A large part of the predictions between December and April was extrapolated due to the low coverage of the winter primary productivity drops, spring peaks and cold waters. They were based on data collected during other seasons and regions, and given the large spatial extent of the area, and the seasonality and regionality of the cetacean distributions, reliable winter predictions might be restricted to geographic areas where winter surveys took place. These predictions are more uncertain and warrant caution. 5. Synthesis and applications: extrapolations and nearby data highlighted environmental gaps to predict cetacean distributions in the North-East Atlantic, which could be covered by future surveys. This informs model users of regions and periods when predictions reliability becomes uncertain. SDMs are invaluable tools for supporting conservation applications and, despite the warnings that have been issued, the degree of information available for predicting distribution is still rarely reported. We recommend adding this assessment as routine information on the reliability of predictions.

Keywords

• Climate change
• Decision-making
• Environmental similarity
• Extrapolation
• Nearby data
• Robustness