Field methods

We captured 43 Turtle-doves between 2012 and 2016 during their breeding season (May and July) using either a baited walk-in drop trap or a baited whoosh net (details of the processes and equipment used are provided in the Supporting Information). Seventeen birds were caught on farmland in the southeast of the UK, and 26 birds on predominantly arable farmland in the north (eight birds), west (four birds) and south (three birds) of France, and in the predominantly deciduous Chizé Forest in the west (11 birds) of France (Fig. S1). We fitted solar PTT-100 s Platform Terminal Transmitters (PTT; Microwave Telemetry Inc., Columbia, MD, USA) to the backs of 42 birds; one of the captured birds in the UK was released after being ringed as it was too small to be tagged without compromising its welfare. We used either a flat Teflon ribbon (Lormée et al. 2016), or 2-mm braided nylon in the 2014 and 2016 seasons (Willemoes et al. 2014). Transmitters were programmed with duty cycles of 10 h switched on and 48 h off (from 2012 to 2014), 8 h on/15 h off from 2015 (France only) and 4 h on/19 h off in 2016 (UK only). Satellite data were received through the ARGOS satellite-based positioning system. Of the 42 tagged birds, 14 (four from the UK and ten from France) produced data for at least one entire winter season (from September to May), with four of these providing data for two or more consecutive winters (Table S1). We do not have information about the remaining 28 birds in the wintering area either because the bird died, or because their tag failed to transmit locations for the full period required. In most cases, it was not possible to discern between these two outcomes.

Tracking data

Tracking data preparation was performed in R (version 4.1.2; R Core Team 2021). We applied an Azimuthal Equidistant projection (EPSG: 54032) and filtered the fixes to remove unrealistic locations using a speed threshold of 90 km/h (Lormée et al. 2016). We identified a position as being in the wintering area if there was a cessation in the directed southward trajectory, limited daily movements (≤ 30 km), lower average speeds (≤ 15 km/h) and the bird remained settled in an area for more than 15 days. The start of spring migration was identified as a consistent northward trajectory. Turtle-dove home-ranges might be geographically distinct in early and late winter (Eraud et al. 2013, Lormée et al. 2016). Inspection of the tracks revealed sudden southwards movements of some of the birds, moving to distinct new wintering sites by the beginning of December. We therefore split the data into two wintering periods: from their arrival in the wintering area in September to the end of November (early winter) and from the beginning of December to the end of the wintering season in May (late winter).

Mapping the winter distribution of satellite-tagged Turtle-doves

We defined the landscape area as the extent of the presence of the wintering tagged birds enclosed by the 100% Minimum Convex Polygon (MCP_[100]) around all the birds' fixes with accuracy classes labelled as LC3, LC2 and LC1, which have an estimated precision of 250, 500 and 1500 m, respectively (Argos 2016) (Fig. 1).

To select the core areas within the birds' home-ranges (the area within which an animal typically lives, moves and conducts its regular activities) (Burt 1943, Warwick-Evans et al. 2018) we applied a ‘time spent in area’ approach as a proxy for spatial utilization (Warwick-Evans et al. 2018). This approach minimizes errors due to the disparity of duty cycles and the number of fixes per bird, and the clustered use of the areas (Aarts et al. 2008, Noonan et al. 2019). We applied a Lambert Azimuthal Equal Area projection centred on the winter positions, created a grid of 2 × 2-km cell size covering the landscape area (133 186 cells), and estimated the time spent per cell for each bird and wintering season using the ‘tripGrid’ function in the ‘trip’ R package (Sumner et al. 2023). Eight hundred cells (0.6%) had at least one fix. For each bird, we ranked the time spent in each cell, and the cells that accounted for the top 96% of the total time spent by a given bird in a wintering season were selected to define the core areas (Casper et al. 2010, Soanes et al. 2013). We calculated two measures of individual home-range size for each of the birds: the MCP_[95] and the MCP_[50], using the fixes in the core areas with the ‘adehabitatHR2’ package (Calenge & Fortmann-Roe 2023) to facilitate comparison with other studies.

To define the environment available to the bird at two different scales, we selected the fixes with the highest spatial precision (LC3) (1) within the landscape area for landscape-scale analyses and (2) within the core areas for core areas-scale analyses. We thinned both samples to ensure that points were separated by at least 2 h and were more than 1 km apart (Willemoes et al. 2014). Then we generated two sets of pseudo-absence locations, one each for landscape and core areas, matching the number of occurrences by bird, month, year and ecoregion in each instance. For both landscape and core areas, a 250-m radius buffer (reflecting the location uncertainty for an LC3 fix) was placed around both the occurrence and pseudo-absence locations to extract the available environmental conditions, including land cover classes as a proportion of the total area in the buffer. Any pseudo-absence locations that fell within 500 m of a presence location, thus overlapping its habitat buffer, were excluded, and alternative pseudo-absence points were generated until no overlaps occurred.

Environmental variables

We considered potential correlates of distribution based on previous research in the non-breeding and breeding seasons and the availability of suitable remote sensing sources for the region. Variables selected were: land cover (Dunn & Morris 2012, Moreno-Zarate et al. 2020, Bermúdez-Cavero et al. 2021), climate (Marx & Quillfeldt 2018, Keller et al. 2020), vegetation phenology (Eraud et al. 2009), topography and availability of freshwater (Jarry & Baillon 1991, Dunn & Morris 2012).

Land cover composition was extracted from the Copernicus Global Land Service, epoch 2015 (CGLS, V2.0) (Buchhorn et al. 2019). Level 1 land cover classes were used to identify shrubland (shrub), grassland (grass), bareland (bare), cropland (crop) and urban/built-up (urban), and level 2 classes were used to split the level 1 forest class into closed forest (for_closed) and open forest (for_open) (Table S2).

Climate data for the wintering seasons from 2014 to 2018 were obtained from the CRU TS 4.03 dataset (Harris et al. 2020). The variables used were monthly precipitation (pre), monthly average daily mean temperature (tmp), monthly average daily maximum temperature (tmx) and monthly average daily minimum temperature (tmn). Climate variables for each point were extracted for the individual month in which the presence or pseudo-absence was recorded. We also derived a variable for the total precipitation accumulated during the preceding summer (PREsum) by summing the rainfall from June to October, the wet season in the wintering area (Fig. S2).

Decadal values of the normalized difference vegetation index (NDVI) were acquired from the CGLS NDVI (Smets et al. 2016), and mean monthly NDVI values calculated to align with the temporal resolution of the climate data. As with the climate variables, NDVI for each point was extracted for the individual month in which the presence or pseudo-absence was recorded.

Altitude (alt) was obtained from the SRTM DEM (Jarvis et al. 2008), from where we also derived the slope (slp). Distance to watercourses, rivers, tributaries and other minor streams (river_dist) was extracted for each position from the AQUAmaps based on the shortest distance in kilometres (FAO 2014), as a proxy indicator of fresh water. Table S2 provides a description of the variables.

Habitat selection modelling

For each of the four sets of analyses (early and late winter at landscape and core areas scales), we examined the presence of factors with near-zero variance (the proportion of unique values and the ratio of the frequency of the most prevalent value to the second) that could lead to different statistical problems, especially due to a large proportion of zeros (Bolker et al. 2009, Zuur & Ieno 2016). In both the landscape and core areas, and in early and late winter, the classes of ‘closed forest’ and ‘waterbodies’ were predominantly (99%) zero values (i.e. absent). Similarly, ‘urban extent’ for the landscape scale and ‘bare land’ for the core areas scale were almost all zeros (Table S3) and we excluded these from further analyses at the relevant scale. We checked for correlations between all covariates and retained those with a Spearman's rank correlation coefficient ρ < 0.7, and checked their collinearity, selecting those with a variation inflation factor (VIF) ≤ 3 (Zuur et al. 2010) (Table S3). At the landscape scale, all three covariates related to temperature (monthly maximum, minimum and mean) were highly correlated, so we selected the monthly averaged maximum temperature as a potential source of physiological stress for the species in the wintering areas, especially during late winter (dry season) (McKechnie et al. 2016). We inspected whether there was a significant difference in locations between early and late winter in terms of longitude and latitude, distance to water and land cover, by applying Wilcoxon–Mann–Whitney tests.

We built univariable models (models with a single independent variable) for the environmental variables selected in the previous step for each of the four sets of analyses. By comparing the Akaike Information Criterion (AIC) of each one against the intercept-only model we identified which variables were informative. Informative variables within each set were then included in a series of multivariable models. At the landscape scale, we fitted six different models for each winter period (early and late). These were land cover variables only; climate, topology and NDVI; land cover, climate, topology and NDVI, and repeating each of these with the addition of the distance to a river. At the core areas scale, we fitted models with land cover factors, including and excluding the distance to the river. To facilitate the interpretation of the results, we plotted the partial effects for both early and late winter in the same plot for the landscape and core scales.

To identify correlates of the distribution of Turtle-doves in the wintering season, and to allow the fitting of non-linear relationships, we applied generalized additive mixed models (GAMMs) using the package ‘GAMM4’ in R (Wood & Scheipl 2020). We applied a logit link function with covariates as smooth terms and bird identity as a random effect. Smoothing parameters were estimated via maximum likelihood (ML), and we applied penalized cubic regression splines. To facilitate the ecological interpretation of the individual effects and the comparability of the results across the models, we fixed the number of knots to four. Model selection was based on an information-theoretic approach (Burnham & Anderson 2002), where models with a difference in AIC ≤2 were considered equally plausible (Burnham & Anderson 2004).

Proportions of shrubland and grassland cover were collinear for both early and late winter at the landscape scale (Tables S3 and S4), so we fitted separate models. Models containing grassland cover always had a lower AIC and Bayesian Information Criterion (BIC), so here we present only the models containing grassland cover. In the core areas, proportions of shrubland and grassland cover were also collinear for early winter, whereas proportions of cropland and grassland cover were collinear for late winter. Therefore, we present the results for the models containing grassland cover.

All the statistical analyses were performed in R version 4.1.2 (R Core Team 2021). Additionally, we used the packages: ‘usdm’ to calculate the VIF (Naimi 2017), ‘caret’ to detect near-zero variance in the covariates (Kuhn et al. 2022) and ‘performance’ to extract the AIC values (Lüdecke et al. 2020). We plotted the partial effects of the predictors and extracted the predicted values with the packages ‘effects’ (Fox et al. 2022) and ‘ggplot2’ (Wickham et al. 2023), respectively.

Wintering area description

The area enclosed by the MCP_[100] (landscape area) covered 528 634 km² across Mauritania, Senegal, The Gambia and Mali and encompassed areas of tropical and subtropical grasslands, savannahs and scrublands, and the Inner Niger Delta (Fig. 1 and Fig. S1). The region is characterized mainly by a hot semi-arid climate, with a dry season from November to April, and a cooler, wetter season from May to October which becomes increasingly humid further south (Fig. S2).

All but two of the satellite-tagged birds spent early winter in southern Mauritania and northern Senegal, often along the Senegal River and tributaries between the two countries, before moving further south, with a change in the median latitude from early to late winter: $$$ {\tilde{x}}_1 $$$ = 15.02°N, interquartile range (IQR) = 0.75 versus $$$ {\tilde{x}}_2 $$$ = 14.20°N, IQR = 1.31, respectively (log_eW Wilcoxon = 12.64, P ≤ 0.001; Fig. S3). Although birds did move eastwards, there was no overall change in median longitude between early and late winter: $$$ {\tilde{x}}_1 $$$ = 10.86°W, IQR = 4.66 versus $$$ {\tilde{x}}_2 $$$ = 10.66°W, IQR = 6.19, respectively (log_eW = 12.27, P = 0.532; Fig. S3). The median distance to the nearest watercourse was similar for both early and late winter: $$$ {\tilde{x}}_1 $$$ = 0.95 km, IQR = 1.99 versus $$$ {\tilde{x}}_2 $$$ = 1.05 km, IQR = 2.19, respectively (log_eW = 12.22, P = 0. 348; Fig. S4).

Across the wintering area, the land cover of shrubland, grassland, cropland and open forest accounted for more than 95% of the area (Fig. S5). Within the 250-m buffers around presence locations, the extent of open forest and shrubland was significantly lower in the areas occupied in early winter than in the areas occupied in late winter: open forest: $$$ {\tilde{x}}_1 $$$ = 4.72 versus $$$ {\tilde{x}}_2 $$$ = 5.70, respectively (log_eW = 12.145, P = 0.002; shrubland: $$$ {\tilde{x}}_1 $$$ = 16.32 versus $$$ {\tilde{x}}_2 $$$ = 21.39), respectively (log_eW = 12.182, P = 0.041). The occupied area of grassland was significantly lower in the late winter areas: $$$ {\tilde{x}}_1 $$$ = 10.75 versus $$$ {\tilde{x}}_2 $$$ = 0.00, respectively (log_eW = 12.490, P < 0.001). There was no significant difference in the extent of croplands within the buffers between early and late winter (Fig. 2).

We identified a total of 54 core areas (non-contiguous clusters of intensively used cells), with a median of four core areas per individual and wintering season (range 1–8). The size of the home-ranges as the MCP_[95] was on average 46.06 km² (± 75.02 km² standard deviation (sd)) ranging from 0.40 to 543.10 km² and the area defined by the MCP_[50] averaged 6.70 km² (± 14.37 km² sd) and ranged between 0.03 and 82.71 km².

Correlates of distribution at the landscape scale

In both early and late winter, the multivariable model for the presence of Turtle-doves with the lowest AIC included all the land cover covariates with the climatic and the topographic covariates (Tables 1 and 2). During both wintering periods, open forest and cropland cover were significantly positively associated with Turtle-dove presence. The likelihood of presence was highest when open forest cover exceeded 20% and cropland cover surpassed 25% (Fig. 3a; Table 1). Shrubland and grassland cover were significant only during early winter, with the positive effect maximized at 25–35% shrubland cover and around 50% grassland cover. Beyond these thresholds, higher cover percentages decreased the probability of presence. Proximity to watercourses always increased the likelihood of presence, especially within 2 km. Altitude had a negative effect, with birds unlikely to be present above 400 m (Fig. 3b). Monthly precipitation was uninformative for late winter (Table S4). In both early and late winter, the random effect (bird ID) was found to be significant (Table 2). Distance to water was the best performing variable in the univariable models, for both early and late winter (Table S5). In both early and late winter, the random effect (bird ID) was significant (Table 2).

The shapes of the relationships between distribution and environmental correlates were similar for both early and late winter, except for summer rainfall and NDVI. In early winter, the maximum likelihood of presence was in areas with between 400 and 500 mm of summer rainfall and NDVI values of 0.5. In late winter, the maximum likelihood of presence was in areas with more than 600 mm of summer rainfall and with NDVI values between 0.3 and 0.4. In early winter, the maximum likelihood of presence was in areas with monthly precipitation of around 50 mm. There was no significant relationship with monthly precipitation in late winter. A maximum monthly temperature of 38°C increased the likelihood of presence in early winter, whereas in late winter, temperatures over 40°C increased this likelihood (Fig. 3b).

Correlates of distribution in core areas

In early and late winter, the models with the lowest AIC (best models, Burnham & Anderson 2004) included land cover and distance to water (Tables 3 and 4). The random factor of bird ID was non-significant for both winter periods (Table 4). In both periods, there was a positive relationship between open forest cover and likelihood of presence above 20% open forest cover, and extent of open forest was the best performing univariate model for both early and late winter (Tables S7 and S8). However, an increase in the distance to water and, to a lesser extent, urban cover, had a negative effect on the presence of birds. In early winter, likelihood of presence was maximized with grassland cover of 25–50%, while in late winter, increases in grassland cover over 25% had a negative effect (Fig. 3a).