Geese were captured at Poyang Lake in the Yangtze River Floodplain, Jiangxi Province, China (29°N, 116°E) during the 2014-2018 wintering seasons, and equipped with GPS - GSM (Global Positioning System - Global System for Mobile Communications), solar- powered loggers. Further background on capture and deployment methods can be found in Si et al. . Each logger was set to record GPS (Global Position System) positions every 2 h. The collected data used in this study include bird ID, latitude and longitude (degree), time of record, and speed (km/hour). We used tracking data from 28 individuals of tundra bean goose and 55 individuals of greater white-fronted goose for further analyses.
The start and end dates of wintering / spring staging in the Yangtze River Floodplain and stopover in the Northeast China Plain for each individual were defined as the first day it arrived/left the specific site and showing continuous presence/absence. For individuals with multi-year data, the year with the highest amount of GPS records was used. Records with a GPS location error over 30 m and/or speed over 1 km/hour (presumably while in flight instead of foraging) were removed from downstream analyses. The sunrise and sunset time were calculated based on the geographic location of each GPS record using algorithms provided by the National Oceanic & Atmospheric Administration (NOAA) (https://www.esrl.noaa.gov/gmd/grad/solcalc/). Day and night locations were defined as the locations recorded in the periods between 1 h before sunrise to 1 h after sunset, and 1 h after sunset to 1 h before sunrise. For each individual bird, day and night counts of locations, start and end dates at the wintering and stopover region, and logger information are summarized in Additional file 1: Table S1.
Habitat use at both regions
We produced a 2015 China land cover map at a 30 m resolution according to Liu et al.  with an overall accuracy over 91% and used this to compare the habitat use of tracked geese at their wintering and stopover regions. Land cover types were regrouped to farmland, woodland, grassland, water, wetland (i.e., swamp, mud flat and bottomland), built-up land, and bare land (i.e., sand, desert, saline-alkali land and bare land). To compare the use of different food resources, land cover types were regrouped and the percentage of tracked locations during the day and at night on farmland, water/wetland/grass (grassland are included to cover wet meadows), and others (the rest land cover types) was calculated in both the wintering and the stopover region.
Compare human pressure on habitats in both regions
For each species, we defined the study area as the range of a 50 km buffer around the GPS locations of a specific species recorded during the day at their wintering or stopover region (Fig. 1). We chose 50 km because the maximum foraging flight distance for American and European geese are generally smaller than this [8, 16, 17], and hence both used and unused area should be included. We then compared the human pressure on habitats for each species at each region. The 2009 Global Terrestrial Human Footprint Map with a 1 km resolution  was used to extract the human pressure on farmland and wetland/grass located within the study area of each bird species in both the wintering and the stopover region. This standardized human footprint map (with a range of 0–50) was generated by integrating remotely sensed and bottom-up survey information that measure direct and indirect human pressures on the environment, including data on the extent of built environments, crop land, pasture land, human population density, night-time lights, roads, railways, and navigable waterways. To test whether human pressure differs between the wintering and stopover region, a Mann-Whitney U test was conducted, as the data were not normally distributed.
Resource selection function modelling
We tested if geese actively select farmland and wetland/grass close to their roosting area and with a relatively low human pressure, using mixed-effect resource selection function modelling. For each region, within the study area of each species, we used daytime GPS records located on farmland and wetland/grass as presence, and randomly generated an equal number of absences on these two land cover types in the part of the study area where geese were not present (Fig. 1). The minimum distance between an absence point and its closest presence point was set to 1 km. To take the resource availability into account, the number of absences generated on each food resource type was calculated based on the farmland and wetland/grass composition in the study area of each species in each region (tundra bean geese in the Northeast China Plain: 48, 51%; greater white-fronted geese in the Northeast China Plain: 66, 34%; tundra bean geese in the Yangtze River Floodplain: 84, 16%; greater white-fronted geese in the Yangtze River Floodplain: 82, 18%). For each species on each day, the distance to the roost was calculated as the distance from a specific GPS point recorded during the day to the center location of GPS points recorded in the previous night. We then built a mixed-effect resource selection function model , estimated by the generalized linear mixed model with a logit link, to predict bird presence for each species at each region separately. Specifically, we used human pressure, food resource (land cover) type (farmland and wetland/grass, with farmland as the baseline), distance to roosts, and their interaction terms as the fixed factors, and bird ID and year as random factors. As we mainly focused on the fixed effects, the Akaike information criterion (AIC) was used to rank the top fixed-effect model first, and random factors were added afterwards. Response curves of factors with a significant effect on bird presence were constructed by fixing all variables other than the variable of interest constant at their median values, and making predictions at regular intervals over the range of the given variable. The response curves of the interaction terms were calculated by fixing all variables other than the two independent factors (1 and 2) in the specific interaction term constant at their median values, and making predictions for factor 1 at regular intervals over the non-outlier range of factor 2. To facilitate comparing among models, we also produced the predictions for factor 1 at a fixed value of factor 2 (e.g., with a fixed distance to the roost). The 95% confidence interval (CI) was calculated using the Wald confidence interval based on the sampling distribution of the Wald statistic in repeated samples. Only the response curve of the interaction term was plotted if both the interaction and the single term were significant. Analyses were carried out in R 3.5.3 , using ‘glmmTMB’ and ‘effects’ packages.