Predictive mapping of Meadow Pipit density using integrated remote sensing data and an atlas of vascular plants dataset

Abstract

Capsule The highest densities of Meadow Pipits in Central Europe are found in lowland and upland wet meadows.

Aims To create a large-scale predictive model of Meadow Pipit density.

Methods We analysed factors affecting the density of the Meadows Pipit in Poland using data from 777 × 1 km study plots and a set of 22 environmental variables, including agriculture intensification and habitat-specific plant species as classifiers of meadow types. Predictors were selected using variation inflation factor, then related to species density data using generalized additive models.

Results The best-supported model included 11 variables and was clearly better (Akaike information criterion weight = 0.47) than other models. The density of the Meadow Pipit reaches its highest levels on large areas of extensively used wet meadows as well as pastures where livestock graze and which show high photosynthetic activity in April.

Conclusion Some aspects of the environment that were not identified from remote sensing data were vital for determining relatively high density. Conservation efforts for preserving Meadow Pipit populations should focus on maintaining wet meadows and extensively grazed pastures. Given the results, the Meadows Pipit may be classified as a good indicator of traditional agriculture.

Changes in agricultural land use patterns, driven by the Common Agricultural Policy, have been identified as primary factors affecting species distribution at a regional and local scale (Pereira et al. 2004, Holzkämper & Seppelt 2007, Giordano et al. 2010). European grassland birds have already suffered long-term population decline as a result of changes in land use. Further declines are forecast, because agricultural intensification is the simplest way to increase efficiency in agricultural production (Newton 2004, Butler et al. 2010, Setchfield et al. 2012). As part of this process, conservationists pay a lot of attention to developing methods for collecting and analysing data from both small and large areas so they can monitor the population health of many species (Franco & Sutherland 2004, Posillico et al. 2004, Balbotín 2005, Gregory et al. 2007, Franklin 2010). Despite numerous techniques for predictive mapping of species distribution (Carpenter et al. 1993, Elith et al. 2006, Araújo & New 2007, Giordano et al. 2010), e.g. maxent (Phillips 2005); bioclim (Austin et al. 1994); saga (http://www.saga-gis.org/en/index.html); garp (Anderson et al. 2002); and enfa (Hirzel et al. 2002), procedures tend to be similar, i.e. extrapolating data from a small- or meso-geographical scale to a large geographical scale. This technique has been used for predicting species distribution on different scales, for example, wetland amphibians in South Africa (Tarrant & Armstrong 2013), gastropods in South Australia (Fordham et al. 2013), mammals in North America (Badgley & Fox 2000), and bird species richness in Central Europe (Kosicki & Chylarecki 2012a).

In many cases, data from geographic information systems (gis) alone are used as predictors, such as corine land cover (CLC), ndvi dataset, wordclim, etc. (Guisan & Zimmermann 2000, Giordano et al. 2010, Kosicki & Chylarecki 2012a, 2012b). Such data are only suitable for some species that have habitat requirements with sharp boundaries in the landscape, e.g. the Common Quail Coturnix coturnix (Sardà-Palomera & Vieites 2011) and the Ortolan Bunting Emberiza hortulana (Kosicki & Chylarecki 2012b). However, other species may respond to more detailed habitats, which are not defined in remote sensing data (Akçakaya 2001, Giordano et al. 2010). This may be the case of the Meadow Pipit Anthus pratensis, a declining grassland bird (Pearce-Higgins & Grant 2006, Chylarecki & Jawińska 2007) that inhabits large areas of moorland and wet meadows (Moss et al. 1979, Cramp 1988, Brown & Stillman 1993).

The Meadow Pipit is a migratory bird that winters in Southern Europe and North Africa and breeds across Eurasia. Its population in Poland has decreased by 20% in the last ten years (Chylarecki & Jawińska 2007), and a similar tendency has been recorded in other European countries, including Scotland, England, and France (Birdlife International 2013).

Previous studies on habitat selection conducted on a small spatial scale showed a connection between the density of Meadow Pipits and lowland open grassy areas with dense low vegetation cover, such as meadows, wet marshy meadows, and extensive meadows (Cramp 1988, Vanhinsbergh & Chamberlain 2001), whereas high grazing pressure in upland negatively affected egg size and the species’ abundance (Evans et al. 2005).

Here, we construct habitat suitability models for Meadow Pipit density in Poland using remote sensing data, the atlas of vascular plants (Zając & Zając 2001) and data from the ‘Common Bird Monitoring Scheme’ (Gregory et al. 2007, Kosicki & Chylarecki 2012a, 2012b). The habitats identified by remote sensing were reclassified into finer categories according to where they were relative to plant atlas records, i.e. extensively and intensively used meadows, dry meadows and peat meadows as well as upland and lowland pastures, etc. In this way, we were able to capture more detailed relationships between Meadow Pipit occurrence and environmental conditions on a large spatial scale (McPherson & Jetz 2007). Our models help to identify areas critical not only for Meadow Pipit populations but also for other endangered grassland species with similar habitat requirements, such as the Lapwing (Sabatier et al. 2012) and the White stork (Kosicki 2010), the numbers of which have drastically decreased in recent years.

METHODS

Bird data

The data were derived from the Common Breeding Birds Monitoring Scheme (Chylarecki & Jawińska 2007) and gathered in Poland from 2000 to 2010 in 777 grid cells measuring 1 km² (Fig. 1). Each cell was chosen at random out of 311 664 squares, and it was surveyed twice a year by volunteers. During the eleven-year period, each square was inspected on average (± sd) 5.4 ± 3.0 times. The first visit took place between 10 April and 15 May and the second between 16 May and 30 June, between 06:00 hours and 07:00 hours. The route of the bird census consisted of two parallel 1 km × 1 km sections (transects), along an east–west or north–south axis. Each transect was divided into five 200 m sections, in which birds were noted in three distance categories (< 25 m, 25–100 m and > 100 m).

Figure 1. Location of the study plots across Poland.

Environmental data

All remote sensing data (Supplementary Table A1) were converted into grass gis file format (Neteler & Mitasova 2008) and re-projected to a coordinate system EPSG4284 projection (http://spatialreference.org/ref/epsg/4284/). Altitude data, which come from the digital evaluation model dataset (GTOPO30), were originally provided by the US Geological Survey's EROS Data Center (http://eros.usgs.gov). Climate data were derived from the worldclim database (www.worldclim.org), which is a set of global climate layers (climate grids) with spatial resolution of 1 km². All grid cells were characterized by six variables, such as Annual Mean Temperature (temperature), Mean Temperature of Warmest Quarter (mtwaq), Mean Temperature of Coldest Quarter (mtcq), Annual Precipitation (precipitation), Precipitation of Warmest Quarter (pwaq), and Precipitation of Coldest Quarter (pcq).

Thirty-six land cover types are recognized in the clc database (based on remote sensing with basic spatial units of 100 m × 100 m), created in 1986–96, 2000, 2003, and 2006 from Landsat TM and SPOT satellites (http://www.eea.europa.eu/publications/COR0-landcover). In Poland, ten types of habitats were distinguished on average in each grid cell. The most frequent were environments classified as arable fields, forests (including coniferous, deciduous, and mixed forest), meadows, complex cultivation patterns, agricultural areas with natural vegetation, urban areas, shrubs, water bodies, and inland marshes. Urban areas were excluded from the analysis because more accurate data were available. We used images of lights at night (human) instead of urban land cover because they measure the distribution of urban cover more accurately. This kind of data is highly correlated with industrial activities (Small et al. 2005, Doll et al. 2007) (Supplementary Table A1).

Meadows (Supplementary Table A1) were categorized on the basis of the Distribution Atlas of Vascular Plants in Poland (Zając & Zając 2001). Maps showing the occurrence of 20 plant species were digitized and arranged in layers. The sum of habitat-specific plant species at the same point reflected the type of meadow. According to plant species specific to a given habitat, six meadow types were distinguished: (1) extensively used lowland meadows, which were very wet in spring and dry in summer, mowed once a year, and not fertilized. This type was chosen on the basis of four plant species, such as Lathyrus palustris, Dactylorhiza majalis, Trollius europeus, Ostericum palustre, Viola persicifolia, Gentiana pneumonanthe, Silaum silaus, and Iris sibirica; (2) dry meadows, characterized by a low level of ground water and no flooding, included Dianthus arenarius, Pulsatilla pratensis, Aster amellus, Adonis vernalis, and Orchis militaris; (3) mountain meadows were classified on the basis of Phyteum orbiculare and Poa chaixa; (4) meadows adjacent to pine forests, which were unpredictable with regards to water conditions and likely to be flooded, included plants such as Nardus stricta and Pedicularis sylvatica; (5) intensively used meadows characterized by a high level of fertilizing and frequent mowing included Poa trivialis and Festuca pratensis and (6) peat meadows, which were very damp, mainly due to rainfall, included Scheuchzeria palustris as the most typical species.

We obtained the normalized differences vegetation index (ndvi, original resolution of 1 km²). The data were derived from the spot dataset (http://free.vgt.vito.be/). ndvi is an index of green vegetation, expressed as a mean monthly value (from March to June), calculated from three measurements taken every ten days.

Data on farming intensification were provided by the Agricultural Census 2002 (http://www.stat.gov.pl/bdl/app/strona.html?p_name=indeks). The variables combine the number of tractors (tractors) and number of cows (cows) per farm. However, the spatial resolution of these data is inconsistent with other class data, because the data are collected within larger administrative units. Therefore, in each of the 16 administrative units of Poland, tractors and cows were expressed as the number of pieces of agricultural equipment per hectare of an open habitat in each grid cell (1 km²) (i.e. sum of the areas: arable fields, meadows, complex cultivation patterns, agricultural areas with natural vegetation), and the number of cows per meadow area in each grid cell, respectively.

Data processing and analysis

All statistical analyses were performed using the statistical package r (R Development Core Team 2010). The mean density of Meadows Pipit as the number of birds in each grid cell was 4.5 individuals/km² (95% confidence limits (CL) = 3.4–5.7, n = 777). However, due to the fact that species density in cells depends on the transect's length and the bird's distance from the observer (Krebs 1999), the Hayne estimator of species density was calculated for each cell according to the following equation (Hayne 1949, Krebs 1999):where D_H is Hayne's estimator of density, n is the number of animals seen, L is the length of transect, and r_i is the sighting distance to each animal i.

The Pearson correlation coefficients and the variation inflation factor (VIF, using the HH library in r; Heiberger 2013) were used to assess the relationship between all predictors (Supplementary Table A2). A VIF of 5 and above indicated a multi-collinearity problem (O'Brien 2007). In the present study, the VIF ranged from 1.00 to 7.79, therefore predictors with a VIF ≥ 5 were excluded from the analysis (Supplementary Table A2).

We used generalized additive models (GAMs) to fit resource selection functions (Hastie & Tibshirani 1990). The response was the Hayne estimator of Meadow Pipit density. Twenty-two variables selected by their VIF were used as predictors (Supplementary Table A2).The most parsimonious model was selected using Akaike information criterion (AIC) (mgcv library in r; Wood 2013) with the lowest AIC and consequently the highest Akaike weight (Burnham & Anderson 2002). We analysed all possible models (2ⁿ , where n is the number of variables), using MuMIn library in r (Bartoń 2013, Hastie & Tibshirani 1990). Although this approach is criticized, the analysis of all possible models is often used when there is not enough a priori information to develop a small set of models (Whittingham et al. 2005, Reino et al. 2010). To reduce the possibility of finding spurious models, we used only dominant uncorrelated habitat variables selected by their VIF from the initial large set of variables (see also Reino et al. 2010). The probability of including a variable in the best parsimonious model was estimated as relative importance (RI) by summing the Akaike weights of all candidate models in which the variable was included (Burnham & Anderson, 2002, Reino et al. 2010). In order to allow some complexity in the functions, while avoiding over-fitting the data, we defined the basic dimension, i.e. k = 4 (Santana et al. 2012). Furthermore, scatter plots and regression diagnostics were used to inspect the shape of the fitted curves. This procedure identifies problems resulting from outliers and influential points (Legendre & Legendre 1998, Santana et al. 2012). A Gaussian distribution of errors and the identity link function were used. As a measure of deviance reduction, we used the D² coefficient, which is equivalent to R², well known from least-squares estimation (Weisberg 1980). We then created a predictive distribution map of the densities of Meadow Pipits from these models. The correlation coefficient between the predicted versus the observed densities (log-transformed) was used as a measure of error prediction (Hastie et al. 2008, Kuczyński et al. 2009).

RESULTS

Population size

Breeding populations of the Meadow Pipit were recorded in 39.6% of grid cells. The mean density expressed as a Hayne estimator was 1.4 (95% CL: 0.7–1.8) individuals/km², while the mean Hayne estimator on occupied plots was 4.2 (95% CL: 3.9–5.5) individuals/km².

Habitat use

Out of all analysed models, only six gained support using information-theoretic criteria, showing AIC weights > 0 (Table 1). The most parsimonious GAM (Table 1) fit to the Hayne estimator of Meadow Pipit density included a smooth GAM fit to extensively used lowland meadows, intensively used meadows, precipitation and latitude and a linear fit to mountain meadows, peat meadows, arable fields, forests, cows, tractors, and ndvi in April. The D² coefficient of this model was 0.64 and it was more than two times better for describing the variation in Meadow Pipit density than the second best model (evidence ratio 2.27) in our candidate set. The most important predictor based on RI was extensively used lowland meadows (RI = 1, F = 4.7, P < 0.001, Fig. 2a), which had a non-linear function and showed that the highest density of Meadow Pipits depended on meadows which were wet in spring and dry in summer, mowed once a year, and not fertilized. The second ariable, with a linear function showed that the highest density of the Meadow Pipit was also found in mountain meadows (RI = 1, F = 3.6, P < 0.001, Fig. 2b). The third significant variable showed that the highest density was in cells with large areas of peat meadows (RI = 0.939, F = 2.3, P < 0.01 Fig. 2c). The fourth variable was arable fields (RI = 0.928, F = 8.2, P < 0.0001, Fig. 2d) with a linear function, showing that Meadow Pipits avoided large areas of arable fields. Forest was yet another important predictor (RI = 0.841, F = 7.3, P < 0.0001, Fig. 2e), also with a linear function, which showed that the species avoided large areas of all types of forests (deciduous, coniferous, and mixed). An equally important variable (RI = 0.841, F = 4.0, P < 0.001, Fig. 2f) was intensively used meadows proving that the species also avoided meadows which were fertilized and frequently mowed. The seventh major variable was precipitation, showing that the highest density depended on a high level of rainfall (RI = 0.767, F = 3.7, P < 0.001, Fig. 2g). The next two factors included in the top model were cows (RI = 0.707, F = 3.2, P < 0.001, Fig. 2h) and tractors (RI = 0.707, F = 3.0, P < 0.001, Fig. 2i). Both were characterized by a linear shape of the response function revealing that Meadow Pipits avoid areas with a high level of farmland mechanization and at the same time preferred areas with a high number of livestock. The ninth variable was ndvi in April (RI = 0.701, F = 5.2, P < 0.0001, Fig. 2j) with the highest density of Meadow Pipits noted in areas with a high level of green vegetation at the beginning of spring. The final variable was latitude (RI = 0.690, F = 3.8, P < 0.001, Fig. 2k), demonstrating that the density of Meadow Pipits increased from the south to the north of Poland.

Figure 2. GAM fits for Meadow Pipit density. Each plot represents a variable's response shape, independent of the other variables, in relation to Meadows Pipit density. The y-axis in each case represents Meadow Pipit density for a smoothed environmental variable and the estimated degrees of freedom are given in parentheses. The shaded areas represent standard errors of the estimated curves. The values fitted are partial residuals. (a) extensive.m – extensively used lowland meadows, (b) mountaim.m – mountain meadows, (c) peat.m – peat meadows, (d) arable fields – area of fields according with clc, (e) forests – area of three types of forest (coniferous, deciduous and mixed) according with clc, (f) intense.m – intensively used meadows, (g) precipitation – annual precipitation according with WorldClim database, (h) cows – number of cows per meadow area, (i) tractors – agricultural equipment per hectare of an open habitat, (j) ndvi.april – green vegetation in April, and (k) latitude – geographical gradient from south (low value) to north (high value).

Table 1. A comparison of general additive models (null and six top-ranked models, where AIC weight >0).

No.	Model	LogLink	AIC	AICc	ΔAICc	AIC weight	D^2
1	Null	−1623.1	3250.2	3250.3	281.5	0
2	s(extensive.m) + L(montaim.m) + L(peat.m) + L(fields) + L(forests) + s(intense.m) + s(precipitation) + L(cows) + L(tractors) + s(ndvi.april) + s(latitude)	−1437.4	2963.2	2968.7	0	0.479	0.641
3	s(extensive.m) + L(montaim.m) + L(peat.m) + L(fields) + L(forests) + s(intense.m) + s(precipitation) + s(ndvi.april) + s(latitude) + s(temperature)	−1438.7	2964.9	2970.3	1.68	0.211	0.610
4	s(extensive.m) + L(montaim.m) + s(peat.m) + L(fields) + s(forests) + s(intense.m) + L(cows) + L(tractors) + s(altitude) + s(human)	−1441.5	2966.1	2970.9	2.28	0.151	0.587
5	s(extensive.m) + s(montaim.m) + s(peat.m) + L(fields) + s(altitude) + s(temperature) + s(human) + s(ndvi.may)	−1438.7	2966.5	2972.0	3.38	0.087	0.321
6	s(extensive.m) + L(montaim.m) + s(precipitation) + L(cows) + L(tractors) + s(ndvi.may) + s(temperature)	−1438.0	2966.8	2972.6	3.93	0.066	0.021
7	s(extensive.m) + s(montaim.m) + s(peat.m) + L(forests) + s(intense.m) + s(precipitation) + L(cows) + L(tractors) + s(ndvi.april)	−1440.9	2970.7	2976.3	7.68	0.011	0.053

(s) – reflects spline, while (L) – linear. The most parsimonious model given in bold. extensive.m, extensively used lowland meadows; montaim.m, mountain meadows; peat.m, peat meadows; fields, arable fields, intense.m, intensively used meadows, precipitation, annual precipitation, human, urban land cover.

On the basis of the GAMs, we created predictive maps of the occurrence of Meadow Pipit density (Fig. 3). The correlation coefficient between the predicted versus the observed density was 0.84 (df = 777, P < 0.0001).

Figure 3. Predicted density of the breeding population of Meadow Pipits in Poland. The map illustrates the joint influence of both spatial and environmental variation on density.

DISCUSSION

On the basis of data from the Polish Common Breeding Birds Monitoring Scheme, Atlas of Vascular Plants (Zając & Zając 2001) and remote sensing data, we created predictive maps of Meadow Pipit density. Our analyses focused on detailed explanations of relationships between species density, habitat, vegetation variables, and climate conditions on a meso-geographical scale (Heikkinen et al. 2004), as well as their extrapolation to a large geographical scale. In general, our results were consistent with other studies (Vanhinsbergh & Chamberlain 2001, Pearce-Higgins & Grant 2006) but several patterns have not been described before. Meadow Pipits reach the highest density on large areas of extensively used meadows (i.e. with a low number of tractors per hectare), peat meadows, and lowland and upland pastures (i.e. with high density of cows), provided that these areas are located in regions with a relatively high level of rainfall. At the same time, Meadow Pipits avoid dry areas as well as pastures, meadows, and arable fields that are intensively used (i.e. with a high number of tractors per hectare).

The preferred habitats reflect the foraging requirements of Meadow Pipits (Douglas et al. 2008). Extensively used wet meadows and pastures, i.e. areas with traditional farming, have high vegetation productivity, seed and insect abundance, and thus provide food resources for many birds (Marone 1992, Verner & Purcell 1999). Meadow Pipits feed on a wide range of invertebrates, particularly tipulidae larvae (Walton 1979, Cramp 1988, Douglas et al. 2008), which in turn rely on wet and stable hydrological conditions such as bogs and wet meadows. On the other hand, in upland pastures, where tipulid density is lower, cattle grazing can lead to a greater diversity and abundance of arthropods that are attracted to the dung (e.g. Diptera spp.). This additional factor may positively affect density in pastures. Some larvae of Tipulid species, such as Tipula oleracea, are plant pests, and are controlled by draining meadows or the use of pesticides (Boczek 2001). Such anthropogenic endeavours, as well as summer droughts, may contribute to the decline of the Meadow Pipit population in intensively used meadows (Renwick et al. 2012).

As far as climate components are concerned, only precipitation affected the high density of the species. Its effect was linear and positive: in areas with higher levels of rainfall, a higher density of Meadows Pipits was recorded. Many studies show that open habitat birds are more sensitive to changes in moisture conditions than species that reside in forests (Marone 1992, Verner & Purcell 1999, Kosicki & Chylarecki 2012a). We suspect that the influence of rainfall on Meadow Pipit density is indirect. Apart from extreme climatic situations, rainfall provides uniform humidity in ‘wet meadows’, e.g. in late spring, when the ground water level is low, thus providing a diversity of potential food.

Our findings also show that Meadow Pipits choose areas with high levels of green vegetation (ndvi) in April, although this has no effect in May and June. Green vegetation levels recorded in April reflect both the condition of the meadows after the winter and water retention in the meadows and/or pastures. The index's high value at the beginning of spring indicates meadows/pastures where flooding is absent and where grass growth has already started. These conditions provide optimal nesting habitats for early spring migrant species (Tryjanowski et al. 2005). In the months that follow, i.e. May and June, grass vegetation is higher, so food is not as easily accessible and detectable for Meadow Pipits (Butler & Gillings 2004, Douglas et al. 2008, Vandenberghe et al. 2009), and foraging mobility is restricted (Devereux et al. 2006, Douglas et al. 2008). We suspect that lowland meadows, which are mowed once or twice per year, provide low vegetation; in lowland and upland pastures, low vegetation is also achieved by grazing. However, the latter relationship is not linear, because severe grazing pressure, especially with regards to sheep, is blamed for being deleterious for vegetation and birds in many upland regions of the UK (Fuller & Gough 1999). Unfortunately, we were unable to conduct a detailed investigation of the effects of grazing intensity, and the richness of specific groups of invertebrate species, because no data were available on a Poland-wide scale.

We also observed that the geographical gradient of Meadow Pipit density in Poland depends on latitude. The highest density occurs in the northern and north-east parts of the country, while the lowest occurs in the southern and south-east parts. A similar pattern can be found in the Polish Breeding Birds Atlas (Sikora et al. 2007), where a heuristic comparison of species occurrence maps shows that this species prefers areas in the northern and eastern parts of Poland and avoids the southeast and central regions. We show that this geographical density gradient of the Meadow Pipit can be attributed to habitat distribution and differences in weather conditions. The high density of Meadow Pipits in lowlands depends on stable or predictable hydrological conditions in meadows, while in the uplands, extensively used pastures tend to play an important role (Buchanan et al. 2006). Alternatively, the geographical gradient could be attributed to habitat modification that was not explored in this study (e.g. fertilizer use, water pollution, wind farms, etc). An earlier study showed that farmland in the western and central parts of Poland is much more intensive (the highest percentage of fertilizer and herbicide use) than in the northern and eastern parts, where traditional farming still occurs (Kosicki & Chylarecki 2012a).

Our model of species density, which is based on habitat, vegetation, and climate variables, is yet another step in predicting species distribution according to different environmental scenarios. By employing remote sensing data and maps of vascular plant distribution, we have not limited the study to climate and habitat (Sardà-Palomera & Vieites 2011, Kosicki & Chylarecki 2012b). Meadow Pipit density relies strongly on an adequate, wet meadow habitat, which in turn depends on precipitation and non-natural processes, i.e. farmland use intensity. Our findings suggest that drainage, and the consequent changes in the vegetation structure caused by changing water conditions, may have a direct effect on this species. Having demonstrated associations between ‘wet meadows’, bird abundance and detailed measures of vegetation, the results of this study have succeeded in providing a quantitative basis for assessing how recent habitat changes may have affected bird abundance.

SUPPLEMENTAL DATA

Supplementary online Tables A1 and A2 can be accessed at http://dx.doi.org/10.1080/00063657.2013.849656.

ACKNOWLEDGEMENTS

We wish to express our gratitude to observers who collected data in the field. The full list of their names can be found at http://main3.amu.edu.pl/~kubako/tsr/. We want to thank Justyna Grześkowiak and Sylwia Wilcox for their linguistic assistance and two anonymous reviewers for comments, which improved the manuscript.

FUNDING

The study was supported by MNiSW [grant N N304 025936].