In recent decades, farmland birds have had to contend with dramatic habitat changes, in terms of both quality and quantity (Donald et al. 2001; Reif et al. 2008). Many passerine bird species breeding in farmland have declined due to the loss and degradation of their breeding and foraging habitats (Fuller et al. 1995; Murphy 2003). In contrast, some species that winter in farmland, such as geese, have been increasing in number since they began feeding on agricultural crops, thereby improving their foraging conditions (Fox et al. 2005; Gauthier et al. 2005).
The Greater White-fronted Goose Anser albifrons is one such species in Japan. The wintering population of this migratory species is currently increasing and depends on rice fields for food (Amano 2009; Fujioka et al. 2010). Prior to the early 1960s, these geese occurred throughout most of Japan (Miyabayashi 1994). However, the population size dropped from an estimated 150,000 in 1940s to about 3,000 in the early 1970s (Yokota et al. 1982). The dramatic decline in distribution and numbers came as a result of destruction of wetland habitats and excessive hunting (Yokota et al. 1982). The population has subsequently recovered to approximately 170,000 (Ministry of the Environment, Japan 2012) after a ban on their hunting was implemented in 1971 (Miyabayashi 1994), post-harvest food availability improved in the rice fields (Shimada 2002), and as an increasing number of habitats were protected legally, such as under the Ramsar Convention (Amano 2009). However, the number of wintering and stopover sites remains limited. The increasing numbers of geese visiting a limited number of sites has caused not only severe agricultural damage (Amano et al. 2004a; Shimada & Mizota 2009), but also raises the risk of the spread of epidemic diseases (Kurechi 1997).
For long-term population conservation, identifying the areas that are likely to be used by the geese at a population level such as the whole population wintering in Japan is necessary (Wisz et al. 2008) because it provides important information on locations suitable for habitat management and restoration. At a local scale, Greater White-fronted Geese prefer to use foraging habitats near their roosts (Ely 1992; Elphick 2008). However, few studies have focused on the habitat selection of geese at a larger spatial scale. Wisz et al. (2008) found that at the flyway level Pinkfooted Geese Anser brachyrhynchus selected foraging sites in cropland and grassland at low elevations near coasts. Investigating factors affecting the distribution of the geese should help when setting priorities for habitat management and restoration.
The purpose of this study was to: (1) investigate the factors associated with the distribution of Greater White-fronted Geese in Japan, and (2) create a potential distribution map for this species in Japan. This study aims to provide basic information for the habitat restoration and management.
METHODS
1) Data collection
The following explanatory variables were used in the model: minimum temperature; average elevation; proportion of rice field area, urban area, and lake area; distance to lakes; maximum snow depth; and latitude and longitude, combined by principle component analysis. WorldClim data (resolution: 5 arc-minutes; Hijmans et al. 2005), which provides global climate layers (e.g. temperature and precipitation) with high resolution (from 30 arc-seconds to 10 arc-minutes), were used to create minimum temperature layers for five winter months (November to February) and elevation layers. Rice fields are the main foraging areas for the geese in Japan (Amano 2009), and the geese are known to avoid man-made structures including houses and roads (Amano et al. 2004b). Rice field (km2) and urban (km2) areas for each grid cell were extracted from Natural Environmental Information GIS (Environmental Agency of Japan 1997), and the proportions of rice field and urban areas to total land area were calculated. The rice field area data was collected in 1997, thus it was older than the goose data collected in 2008–2012. Although the rice field area in Japan has been declining, we used the available data set because the rates of decline were similar across regions (Ministry of Agriculture, Forestry and Fisheries 2012). The urban area was defined as man-made, including residential, industrial and commercial zones. Since the geese usually roost on lakes in Japan (Miyabayashi 1994; Shimada 2009), we calculated the proportion of lake area (km2) to total land area and the distance to lakes (km). We extracted these data from Natural Environmental Information GIS (Environmental Agency of Japan 1997) and Numeric Map 25000 (spatial database; Geospatial Information Authority of Japan 2003). The maximum snow depth in each cell was obtained from the Mesh Climatic Data 2000 (Japan Meteorological Agency 2002). The data were collected from 1971 to 2000, thus it too was older than the goose data. However, we used these data because, although snow depth fluctuates between years, potentially causing population fluctuations in waterbirds, no trend in mean snow depth was observed in a recent decade in Japan (Ueta et al. 2007). Because the geese breed northeast of Japan in north-eastern Siberia (Takekawa et al. 2000), they prefer to winter in more northerly locations (Hestbeck et al. 1991). Therefore, the latitude and longitude for the center of each cell were combined with the principle component analysis, and the PCI scores (=latlong PC1) were used as explanatory variables. All variables were resampled to the spatial resolution of 0.0962° (about 10 km ×ばつ 10 km), and 5,162 grid cells were created because the geese travelled a typical distance of 10 km from their roosts to foraging areas while wintering in Japan (Takekawa et al. 2000).
The localities where geese were recorded were extracted from five-year data (2008–2012) of the Nationwide Waterfowl Survey, organized by the Ministry of the Environment in Japan (Ministry of the Environment, Japan 2008, 2009, 2010, 2011, 2012). A one-day survey was conducted in mid-January every year at 8,631–9,084 sites and thus targeted at wintering, not staging, populations. The count data were collected either at roosts or foraging sites. However, assuming that when geese are observed on a particular lake (i.e., at roost), they are also present in surrounding foraging sites, we combined both types of data to use as much information as possible in the following analyses. Greater White-fronted Geese were observed at 128 sites during five years. These observations were aggregated to 88 grid cells (resolution of 0.0962°) and used as presence-only data for the response variable (Fig. 1). Survey sites were set up at all sites where swans, geese, and ducks had been observed in the past surveys or by other conservation groups. The records for each site were also resampled using the same grid cells as the explanatory variables. All the explanatory variable values, latitudes, and longitudes were extracted for the 88 grid cells, and they were grouped into 50 clusters by cluster analysis performed using R (ver. 2.13.1, R Development Core Team 2011) to avoid using similar environmental data in nearby locations. Every location of each cluster was selected, and the 50 localities were used as response variables.
Fig. 1.
Wintering sites and numbers of Greater White-fronted Geese that were used in the analyses, in Japan in 2008–2012.
2) Alleviating spatial autocorrelation
Spatial autocorrelation is known to violate the assumption of independently and identically distributed errors of most standard statistical procedures and increase type I errors (Dormann et al. 2007). Eigenvector-based spatial filtering is one of the methods to resolve spatial autocorrelation in species distribution modeling (Griffith 2010). This method embeds spatially dependent variables into independent variables to remove spatial dependence from the models. We used the software package Spatial Analysis in Macroecology (SAM, Version 4.0; Rangel et al. 2010) to create spatial filters by using spatial eigenvector mapping. We chose the truncation distance in the geographical distance matrix for spatial eigenvector mapping to be 25% of the maximum distance. Because of the computational limitations of SAM, calculations could not be carried out for more than 4,000 sites. Thus, spatial filters were created at a slightly coarse resolution of 0.15°. We used the inverse distance-weighted method with ArcGIS 9.3.1 (ESRI Inc., Redlands, CA, USA) to interpolate the spatial filters to the same resolution as the environmental variables (0.0962°). Several strategies have been proposed for the selection of spatial filters to be included as independent variables in species distribution models (SDMs; e.g., Dormann et al. 2007; Blach-Overgaard et al. 2010; Griffith 2010). As reported by Dormann et al. (2007), we adopted a method of selecting spatial filters that minimized the spatial autocorrelation of the model. That is, we selected the combination that enabled Moran's I of model residuals to be lower than 0.01 with the smallest number of spatial filters.
To check multicollinearity, we calculated the correlation matrix for all explanatory variables. The significance levels of the Pearson's product-moment correlation were adjusted using the Bonferroni correction (Quinn & Keough 2002). We eliminated the significantly correlated variables from the subsequent analysis. These statistical analyses were carried out using R 2.13.1 (R Development Core Team 2011).
3) Species distribution modeling
We used the maximum entropy approach (MaxEnt version 3.3.3e; Phillips et al. 2006) to develop distribution models from presence-only data, because detection probabilities are expected to be low with one-day surveys and the recorded absence might actually represent a failure to detect the species. The maximum entropy approach is known to be better than other approaches (Elith et al. 2006) and is capable of providing highly accurate estimates even with small sample sizes (Hernandez et al. 2006). We used 75% of the locations to compute 10 randomly chosen replicates for model training; the remaining 25% of the locations were used for model testing, and each replicate of the model was iterated 10,000 times.
The method of selecting background data, in other words, pseudo-absence data, is known to improve the model reliability for MaxEnt (Phillips et al. 2009; Elith et al. 2011). All surveyed grid cells (n=2,226) were used as the background datasets, and the results were extrapolated to all grid cells (n=5,162) in Japan.
We evaluated the resulting model with receiver operating characteristic (ROC) curves by calculating the area under the curve (AUC); the threshold independent index ranged from 0.5 (random accuracy) to 1.0 (perfect discrimination). Models with AUC values above 0.75 are considered potentially useful (Elith 2002).
For comparison, we also applied a model without spatial filters. We used mean AUC values to compare the models with and without spatial filters. The median potential suitable index of the model with high AUC value was used to produce a potential map. To compare the strength of spatial autocorrelation, we also calculated global Moran's I coefficients for residuals of the estimated habitat suitability indices in the model.
RESULTS
Of the 775 spatial filters that were extracted, five were selected to minimize Moran's I value for the residuals. For the 12 explanatory variables, elevation was correlated with the proportion of the rice field area. Minimum temperature was correlated with distance to lakes, maximum snow depth, and latlong PC1. Latlong PC1 was correlated with the proportion of rice field area, distance to lakes, and maximum snow depth. Therefore, elevation, minimum temperature, and latlong PC1 were excluded from the following analysis.
The AUC was 0.76 (0.07 SD) and 0.80 (0.06 SD) for the models with and without spatial filters, respectively. However, there was significant spatial autocorrelation left in the residuals of the model without spatial filters (Moran's I=0.08, P = 0.001), not in the residuals of the model with spatial filters (Moran's I=0.02, P = 0.2). Therefore, the species' potential distribution was mapped using the model with environmental variables and spatial filters.
The top predictors of the distribution of the geese were (in order of importance): the proportion of rice field area, urban area, and lake area and maximum snow depth (Table 1). In particular, the contribution of rice field area was greatest and was positively correlated with the potential suitability indices (Fig. 2a). The proportion of urban area showed a positive relationship with the low suitability indices; however, it reached a plateau when the suitability indices were high (Fig. 2b). The suitability indices were highest at intermediate levels (15–25%) of the lake area (Fig. 2c). The maximum snow depth showed a negative effect on the suitability indices (Fig. 2d).
Table 1.
Contributions of explanatory variables to the model estimated by MaxEnt. Contribution values were in a range between 0 and 100, and they indicate the importance of the variables for the model. The contribution values for five spatial filters were added as the sum of spatial filters.
t01_117.gif
Fig. 2.
Relationships between habitat suitability indices and the four most effective environmental variables: (a) proportion of rice field area, (b) proportion of urban area, (c) proportion of lake area, and (d) maximum snow depth. Solid lines represent average values, and dotted lines represent 95% confidence intervals.
Suitable habitats for the Greater White-fronted Goose tended to be concentrated on the plains along the Sea of Japan and the Pacific coasts, along the Inland Sea (Setonaikai), in northern Kyushu, and in central Hokkaido (Fig. 3). Areas with a high suitability index, but without geese, included the southern part of the Sendai Plain in Miyagi Prefecture, Lake Teganuma in northern Chiba Prefecture, and Lake Kasumigaura in southern Ibaraki Prefecture.
DISCUSSION
The model predicted the distribution of Greater White-fronted Geese with an adequate AUC value. Despite the coarse resolution for creating the spatial filters, the use of spatial filters alleviated spatial autocorrelation in the model.
The most important environmental factors affecting distribution were the proportions of rice field area, urban area, and lake area, and maximum snow depth. Habitat suitability was high in plains with rice fields. This result suggests that rice field is an important foraging habitat for geese. Goose habitat was originally wetland (Ackerman et al. 2006); however, in Japan, owing to the reclamation of wetlands and plains for the development for rice fields and urban areas, wetland area has decreased by 49% since 1868 (Geospatial Information Authority of Japan 2000).
The importance of farmland as the habitat of geese has been reported in numerous earlier studies of the same species (e.g. Ackerman et al. 2006) and of Pink-footed Goose (e.g. Wisz et al. 2008). The extent of the urban area also had a partially positive relationship with habitat suitability, but only at lower proportions (less than 20%). This presumably reflects the effects of several factors that were not directly considered in this study, such as higher detectability and larger number of survey sites in urban areas, or spatial congruence of urban areas and the lowlands that geese usually prefer as habitat. Wisz et al. (2008) reported that Pink-footed Geese preferred habitats close to the sea because they usually roost along seacoasts. Similarly, the partially positive effect of the proportion of lake area on the habitat suitability of Greater White-fronted Geese seems to confirm that they prefer lakes as roost sites in Japan (Miyabayashi 1994). However, the size of lakes that the geese prefer, remains to be investigated. We cannot dismiss the possibility that the positive effect of lake area was slightly overestimated by including observations at roosts.
Fig. 3.
Map of the estimated habitat suitability for Greater White-fronted Geese wintering in Japan. The grid cells were set at about 10 km ×ばつ 10 km.
Although more than 80% of the wintering Greater White-fronted Goose population in Japan is concentrated around lakes Izunuma and Kabukurinuma in northern Miyagi Prefecture (e.g., Shimada 2002), this study revealed that more suitable wintering sites are distributed in other regions. One reason for the difference between the predicted suitability and actual distribution may be the effect of other factors that were not considered in this study. For example, longterm fidelity to wintering sites seems to be one of the primary factors affecting habitat selection by geese (Hestbeck et al. 1991). Other factors that might affect the winter food supply include: the distribution of wheat fields, another habitat type used by geese in some regions (Shimada & Mizota 2009), the amount of ratoon rice (rice regenerated after harvesting), which is known to be an important food resource for geese (Huang & Isobe 2007), and the extent of autumn plowing in rice fields, which dramatically reduces the amont of post-harvest waste grain remaining available to birds (Amano et al. 2007).
Another reason for the difference between the predicted suitability and actual distribution of wintering sites is that the geese need suitable roosts, such as lakes, as well as suitable foraging habitats, such as rice fields, in their wintering areas (Kurechi 2006) and this study did not consider factors affecting the quality of roosting sites. Consequently, the suitability index was predicted to be high for some previous wintering sites that are not visited by geese anymore. Such areas include Lake Teganuma in Chiba Prefecture, where the quality of roosts, not foraging sites, has been degraded because of development (Austin & Kuroda 1953). In contrast, although other previous roosting lakes, such as Shinhama Game Preserve, Lake Shinobazu and moats near Hanzomon in Tokyo, still exist, our study accurately predicted the currently low suitability of these areas because fields surrounding these lakes were lost during urban development (Kuroda 1939, Austin & Kuroda 1953). Such degradation of roosts and/or foraging habitat seems to have caused the population to become concentrated in particular wintering areas such as in Miyagi Prefecture.
Wintering sites for Bean Goose Anser fabalis, such as Lake Biwa in Shiga Prefecture, Lake Fukushimagata in Niigata Prefecture and Lake Kasumigaura in Ibaraki Prefecture, were also evaluated as highly suitable sites for Greater White-fronted Goose. Similarly, major wintering sites for two swan species (Tundra Swan Cygnus columbianus and Whooper Swan C. cygnus), along the Sea of Japan coast were also evaluated as highly suitable sites. Although Greater White-fronted Goose sometimes winter at these sites, their numbers there are not large. The analysis in this study mainly focused on foraging habitats, but roost site selection might be also affected by interspecific relationships. In order to test this hypothesis, it is necessary to investigate the micro-habitat selection of roost sites by each waterfowl species.
The results of this study provide useful information for the habitat restoration and management of Greater White-fronted Geese. Managing or restoring suitable roost sites in suitable habitats predicted by this study would increase the number of wintering sites available to the Greater White-fronted Goose in Japan. Particularly, in suitable habitats where only small numbers of geese were observed wintering (e.g. in Hokkaido, Toyama and Okayama Prefectures) and those near existing large wintering sites (e.g. in southern Miyagi Prefecture), the improvement of roost sites will allow the geese to expand their wintering areas leading to the successful restoration of their habitats.