Environmental parameters associated with colony distribution on Skomer island
Wilcoxon rank sum tests for all colony classifications revealed statistically different slope angles between colonies and non-colonies on Skomer Island (taking any positive count of birds as colony presence, W = 611724994, p-value < 2.2e-16). Unoccupied sections were generally less steep than occupied sections (median slope angles 45o and 51o respectively) and the densest and largest colonies were associated with the steepest cliffs (Fig. 1). Slope angle varied with cliff orientation, with mean slope angle being lowest for cliffs with S and SE orientations. Interestingly, cliffs facing SW have relatively high mean slope angles (Fig. 1a). Despite this, most occupied sections were orientated away from the SW (Fig. 2a). Both the largest, as well as the densest colonies, were mainly orientated towards the NW and E.
Models of wind and slope angle significantly predicted colony presence and absence on Skomer, with the model correctly identifying 80% of the colonies and 93% of avoided sites for a SW wind, where presence was taken as the 10 largest colonies (Table 1). In fact, models of colony location performed well across all modelled wind directions except for SE winds, indicating that colonies are characterised by particular sets of flow characteristics in most scenarios. Indeed, in each wind direction, one colony classification could be predicted by airflow parameters alone, without slope angle. The lowest model performance was seen with SE winds where only one colony classification yielded a reasonable model fit (McFadden41 R2≥ 0.2, table 1). In this model, colony presence was predicted by slope angle alone (Table 1).
Across the different colony classifications, the pseudo R2 was lowest overall in models of any occupancy (1st classification SW 0.28; NW 0.24, SE 0.12, NE 0.38) compared to those predicting the 10 largest colonies (2nd classification SW 0.59; NW 0.58,SE 0.32, NE 0.74) or the 11 densest (3rd classification SW 0.17; NW 0.40, SE 0.15, NE 0.67) (Table 1). The overall accuracy and true skill statistics followed the same general trend, being highest for the 10 largest colonies (Table 1). The sensitivity tended to be somewhat lower than specificity in the 1st classification (sensitivity was: SW 0.63; NW 1.00, SE 0.21, NE 0.60 and specificity was: SW 0.69; NW 0.21, SE 0.96, NE 0.84), but increased in the 2nd and 3rd classifications indicating a better ability to predict true presence compared to true absence for the largest and densest colonies.
While mean slope angle was included in two of three models for SW, NW and NE wind directions, airflow parameters always had a higher effect size (SI Table 2). A narrow set of airflow parameters was identified as significant in predicting colony presence across wind directions (Table 1). Furthermore, there was broad agreement between the airflow characteristics identified within each wind direction, irrespective of the way colonies were classified. Pressure statistics were included in all SW models, while TKE and horizontal wind speed parameters were included in all NW and NE models, respectively.
Unexpectedly, wind speed (in terms of median or skewness) was not included in the simplest top models of SW winds, instead colonies were associated with a lower median pressure and pressure gradient (Table 1), both of which are linked to lower exposure (Fig. 2, Fig. 3).
Colonies were associated with higher turbulence compared to unoccupied sites, particularly in the case of NW and NE winds (Table 1). In NE winds, colonies experienced both higher wind speeds (Fig. 4) (positive horizontal median, 1st classification and negative horizontal skewness, 2nd and 3rdclassifications) and higher turbulence (here TKE estimates in the 1st and 3rd classifications).