Fig. 3

Habitat suitability for Nightingales in West Africa based on weighted averages of four random forest models. (a) Mean predicted suitability between 2010 and 2020. (b) Change in suitability between 2001 and 2020. Geolocator-derived non-breeding locations occupied at mid-winter for birds breeding in the UK (circles), France (squares) and Italy (triangles) are shown. Geolocator locations whose 100 km buffers do not encompass any suitable habitat (as determined by the mean threshold that maximised TSS) and therefore potentially reflect calibration errors resulting in inaccurate latitudes, are shown in red in (a). (c) Distribution of suitability values within non-breeding location buffers for 2001 and 2020, the width of the shape corresponding with the frequency of data points at each value. Points indicate the mean with 84% confidence intervals (nonoverlapping confidence intervals significantly different at P = 0.05)⁷. The dashed line indicates the threshold used to determine suitable habitat. (d) Mean Decrease in Accuracy⁶⁹ for the top ten variables determining suitability based on 40 random forests (see Methods). Climate variables are the averages between 2001 and 2020, apart from annual accumulated precipitation and the Palmer Drought Severity Index (PDSI), which were extracted for the year associated with the record. Values for the precipitation variables refer to the quadratic term.