Study site and roost sites
The study was conducted in Hong Kong, which is located on the south coast of China (22° 09′‒22° 33′ N, 113° 50′‒114° 26′ E). It is a compact and highly populated city, with the majority of 7.48 million population residing on only 24.4% of 1110 km2 total land (Planning Department 2020). Hong Kong has a humid subtropical monsoon climate, tending towards temperate for nearly half the year. Due to monsoons and typhoons, 80% of the precipitation is concentrated between May and September (Hong Kong Observatory 2020). The mean daily minimum temperature varies monthly, ranging from 14.6 to 26.9 °C. In contrast, the native habitats of Yellow-crested Cockatoo in Indonesia are almost entirely tropical. The mean daily minimum temperature remains fairly constant throughout the year, averaging 23.2 °C. The minimum temperature variations between Hong Kong and Indonesia are relatively greater between December and March than other months, with Hong Kong being 5.6 °C to 8.6 °C colder during this season (Additional file 1: Table S1). This indicates a possible night-time cold stress of Yellow-crested Cockatoo in the winter of Hong Kong.
The studied roosting flock, which can be regularly observed and easily tracked, assembles on the northern coast of Hong Kong Island. This district is a prime financial and commercial centre of the city, characterised by dense and tall buildings. It also contains two large urban parks, i.e. Hong Kong Park (8.2 ha) and Hong Kong Zoological and Botanical Gardens (5.6 ha), which are the major feeding grounds for Yellow-crested Cockatoo, providing abundant and various plant food sources (Fig. 1).
The roosting population was counted once a week in clear evenings for two years (Year 2014/15 and 2015/16) from March 2014 to March 2016. As the evening progressed, birds in pairs or small flocks successively flew into the pre-roost gathering sites in Hong Kong Park from all directions. The observation lasted from late afternoon to early evening, beginning when the individuals flew to the pre-roost aggregation sites until no conspicuous flying and calling came from the flock at the final roost site (Davis 1955). We located the birds visually using binoculars (Nikon Monarch 5, 10 × 42), standing at a 30-m high observation tower in Hong Kong Park (Fig. 1). The observation tower afforded an unobstructed view to track the birds as they entered the pre-roost gathering sites and then settled into the roost sites, with a few exceptional locations where the birds could not be seen but their presence could later be identified by faecal droppings (Gorenzel and Salmon 1995). The number of cockatoos attending the pre-roost aggregation was counted every 5 min until the cockatoos flew into their roost site(s). The roosting flock size was counted at the roost site; otherwise, it was estimated from the final number obtained at the pre-roost site when not all individuals could be distinguished at the roost site. The pre-roosting aggregation would sometimes split up to occupy two to three roost sites, that the one greater in number was regarded as the major roost while the others were minor roosts.
In total, seven roost sites were identified to be used by Yellow-crested Cockatoo during the 105 observational days. Five sites were located on trees (BT, CT, GT, PT and RT), one on a building roof (CR), and the other on a cluster of lamp posts (LP). Given the occupied months and night minimum air temperature of occupied days of each roost (Additional file 2: Fig. S1), we classified the roost sites into three distinct groups, i.e. spring roosts (GT and PT), summer roosts (BT, CR and LP), and winter roosts (CT and RT). As roost CR was not accessible, we excluded it from all microhabitat-related analyses.
Data collection
To investigate whether land use around roosts affects roost site selection, we measured the proportion of land use types within a 50-m radius plot around each roost delineated by Google Earth Pro (Google Inc.). We modified the classification of land uses defined by Yap et al. (2002), whereby “tree-dominated area” refers to lands covered by tree canopies, “built-up area” represents developed areas with buildings, and “open space” consists of areas covered by roads, low-growing vegetation or other vacant lands.
To explore the level of human disturbance of each roost site, we quantified human disturbance using multiple anthropogenic components (Gorenzel and Salmon 1995; Peh and Sodhi 2002; Yap et al. 2002; Jaggard et al. 2015). Specifically, we collected data on (1) pedestrians: night-time occurrence of pedestrians passing by the roosts in 10-min durations; (2) traffic: night-time car volume on the nearest roads to the roosts in 10-min durations; (3) night-time noise level at the roost sites; (4) night-time light intensity at the roost sites; (5) mean roosting height above ground; and (6) distance to the nearest main road, building, streetlight and tree. We measured the mean number of pedestrians and number of vehicles of each roost by conducting a 10-min survey soon after the roosting flock settled down, once a month from July 2015 to February 2016. Similarly, we measured the noise levels and light intensity once a month, using an extendable sound meter and light meter, respectively. We held the instruments at the roosting height at four directions around each roost, soon after the roosting flock settled down. We recorded the light intensity three times per sampling evening and the average decibel for a total of 5-min per sampling evening. We measured the mean roosting height using a rangefinder (Nikon Forestry Pro). We estimated the distance to the nearest main road, building, streetlight, and tree on the Hong Kong GEOINFO MAP (www.map.gov.hk/gm/) since the measurements could not be accurately made on the ground. The mean values of the anthropogenic disturbance variables are given in Additional file 1: Table S1.
As described above, the night-time temperature in the winter of Hong Kong is much lower than that in the native habitats of Yellow-crested Cockatoo. We measured the microclimate of each roost site when the winter roosts were occupied, and recorded the microhabitat temperature at each roost site by positioning a thermometer on an extendable pole at the roosting height. The temperature was recorded for 5 min soon after the roosting flock settled down on each roost on 12 randomly selected days from late November 2015 to January 2016, from which the mean values (Tmicro) were calculated.
Data analysis
We compared the roosting flock size of each roost site using Kruskal–Wallis test followed by Wilcoxon rank sum tests. We conducted Pearson’s chi-squared test to determine whether roost sites and land use types were independent of each other. We used Pearson residuals to measure the discrepancy between observed and expected values (Friendly 1994; Friendly and Meyer 2015). The formula is:
$$\text{Pearson residual} = (\text{actual} - \text{expected})/\sqrt{\text{expected}}$$
Cut-off points of Pearson residuals at ± 2 and ± 4 implied that the residuals were significant at α = 0.05 and α = 0.0001 levels respectively (Meyer et al. 2006). We performed a hierarchical clustering analysis using Ward’s method to investigate how the roosts were grouped by human disturbance variables. Thereafter, we performed a partial least squares determinant analysis (PLS-DA) to maximise the variation between seasonal roost groups and identify the type of human disturbance important in roost site selection. PLS regression is designed explicitly for analytical situations where predictor variables are highly correlated and/or the sample size is smaller than the number of observations (Carrascal et al. 2009). The special case of PLS-DA is a supervised method where the response is a categorical variable (Pérez-Enciso and Tenenhaus 2003), like the seasonal roost pattern in the present study. This method extracts components (or latent variables) from predictor variables and indicates its importance in explaining the response variable. The results of the PLS-DA model include R2X and R2Y scores which indicate the explained variance and Q2Y score which indicates the predictive variance of model fit (Wu and Guo 2018). The performance of PLS-DA model was assessed via overall misclassification error rate and significance of R2Y and Q2Y by leave-one-out cross-validation and permutation testing (100 cycles) respectively (Thévenot et al. 2015; Rohart et al. 2017). The predictor variables were standardised before PLS-DA regression. We performed a one-way repeated measures ANOVA followed by paired t-tests to determine whether there was a significant difference of microclimate temperature between winter roosts and other roosts. We conducted the Mauchly’s test of sphericity to test if the variances of each paired roost site were equal and assumption of sphericity was met. Prior to the analysis, Tmicro values were standardised to eliminate the variation among sampling times.
We used R version 3.5.1 (R Core Team 2018) to conduct the statistical analyses. Prior to statistical analyses, we checked the data normality by Shapiro–Wilk normality tests and homogeneity of variance by F-tests. We used the car package (Fox and Weisberg 2019) for one-way repeated measures ANOVA, vcd package (Meyer et al. 2017) for Pearson chi-squared test and mosaic plot (Friendly 2013), ggtern package (Hamilton and Ferry 2018) for ternary diagram, cluster and factoextra packages (Kassambara et al. 2017; Maechler et al. 2019) for clustering analysis, and ropls and mixOmics packages (Thévenot et al. 2015; Rohart et al. 2017) for PLS-DA analysis.