2.1. Study area and spatial scales
The study area covers the whole of Finland. In the first chapter, the spatial unit of investigation was the forestry board district. Out of 19 districts, 18 were subject to the analysis (number 14 was excluded due to scarce data in several years, Fig. 2A, I). In the second chapter, two broad spatial scales were used.
First, the data were calculated for 50
× 50 km grid cells (N = 131) which cover the whole of Finland (Fig. 2B, II). Secondly, to incorporate the assumed regional differences into the models, Finland was divided into five study regions: southwest, southeast, west, east and north (Fig.
2B, II). The spatial scales in chapters I and II can be described as regional and national (Table 1).
Moving towards the landscape-level, in the last two chapters, lekking area data from three study locations (southwest, central and north Finland) were used (Fig. 2C, III, IV). Landscape structure was determined around the lekking sites using two radii. First, to create a spatial scale covering the whole lekking area, a circle with a radius of 1000 m around the middle point of the lek (covering 314 ha, Fig. 2D) was chosen (Wegge & Larsen 1987).
The second spatial scale was created using a circle with a radius of 3000 m, representing the landscape context in which the lekking areas are embedded (covering 2827 ha, Fig. 2D, III, IV).
In the study area, forests are dominated by Scots pine and Norway spruce (Picea abies L.), with some birches (Betula spp.) and other deciduous trees. Southwest Finland is under the heaviest human impact, with large areas being reserved for cultivated fields and human settlement (II, III). In the eastern and central areas (east, southeast and central Finland), the area under cultivation is smaller compared to other parts of the country, whereas the total area of water bodies is dramatically larger (II, III). North Finland is probably the most distinctive region, with the lowest overall productivity, the oldest average age of the forest, and the lowest degree of human impact (II).
There, the matrix (i.e. non-forest area) consists mainly of open bogs and areas clear-cut for regeneration (III). The different study locations made it possible to have different habitat and matrix types along the south-north axis of Finland, also depicting the decreasing overall productivity towards north and the variation in the degree of human impact.
Fig. 2 Study area and the spatial scales used in the thesis. A) 19 forestry board districts covering Finland (I). B) 50 × 50 km grid cells covering Finland and the five study regions (SW = southwest, SE = southeast, W = west, E = east and N = north Finland, II). The small black dots are wildlife triangles from which the species data was obtained (see Species data below). C) The three study locations (from the bottom: southwest, central and north Finland, III, IV). Dots and crosses represent persisting and ceased lekking sites, respectively. D) Two spatial scales (1000 and 3000 m radii around the lekking centre) are shown in the magnification of a lekking site from southwest Finland.
2.2. Species data
In the first two chapters, we used Capercaillie abundance data. Before the year 1989, the abundance estimates (individuals per km2) were based on tetraonid route
censuses (Rajala 1974, Lindén &
Rajala 1981, I). Census routes were located on the best grouse habitats, such as edges of forests, and the census data comprised of the relative densities of young and adult birds (selected as the dependent
variable in chapter I), estimates of the percentage of hens with a brood, and estimates of brood sizes. The routes were counted annually in August by thousands of volunteer hunters. About 500–800 routes were annually counted, the total route length varying between 20 000–30 000 km/year (Lindén & Rajala 1981).
From 1989 onwards, data on Finnish wildlife abundances have been collected using the wildlife triangle scheme (see Lindén et al. 1996). The wildlife triangle network consists of 1650 triangles, from which 800–900 are counted twice a year, in winter (January–March) and in late summer (August), mainly by volunteer hunters. The network covers Finland in a regionally representative way.
These census routes are equilateral triangles with 4 km sides (total
length 12 km), which are randomly located in forest. For chapter II, we used Moose abundance which is estimated in winter by counting snow tracks crossing the census line (tracks/10 km/day), and Capercaillie abundance which is based on grouse counts during August, using the same triangles (individuals/km2 of forest land).
For the last two chapters we used Capercaillie lekking site data (N = 381). The data were first collected in 1970–1992, and the same sites were resurveyed mostly in 2000–2005.
The data were collected by the Finnish state enterprise Metsähallitus, by the Finnish Game and Fisheries Research Institute (especially P. Valkeajärvi and his team), and by questionnaires and
Table 1 Descriptions of the spatial and temporal scales, and the data used in the chapters I–IV. See details in the text. CC= Capercaillie, NFI = National Forest Inventory.
Study
characteristics Chapter I Chapter II Chapter III Chapter IV Spatial scale 18 forestry
boards
50×50 km grids
& 5 regions
1000 m &
3000 m radii
1000 m &
3000 m radii - description Regional National, regional Landscape Landscape Temporal scale 1965–1988 1989–1996 (–07) 1970s/2000s 1970s/2000s
- description Past Recent Past vs. recent Past vs. recent Species data CC abundance CC & Moose
abundance Lekking area
persistence Lekking area persistence - data source Route cencuses Wildlife triangles Several sources Several sources Land-use data Forests < 40 &
< 80 yrs old Forest & human impact data
Forests
> 60 m3ha–1
Forests
> 152 m3ha–1 - data source NFIs 3, 5–8 Several sources NFI 9 NFI 9
interviews from local game district, land-owners, and hunters. The sites were visited one or more times during the lekking season in March–
May, and the presence of birds, snow tracks or fresh excrements were all interpreted as an occupied lekking area. The lekking areas were located on a digital map, and the occupancy data were classified according to the distance between the leks in the old and new surveys.
The old and new leks that were 1000 m away from each other were classified as persisting (i.e. the centre of the lek had moved inside the radius of 1000 m between years but the lek area had remained largely the same, Rolstad & Wegge 1989c, d), whereas the old leks that were > 1000 m away from the new leks were classified as ceased (III, IV).
2.3. Land-use data
Data on forest cover and age structure were obtained from National Forest Inventories (NFIs, see e.g. Tomppo et al. 2008, I, II, III, IV). For chapter I, we used data from five inventories, which cover years 1951–1953 (NFI 3), 1964–1970 (NFI 5), 1971–1976 (NFI 6), 1977–1982 (NFI 7) and 1986–
1994 (NFI 8). Annual values for forest age structure variables (see Table 1) were obtained through linear interpolation, using the first year of an inventory as the main data point from which the interpolation was carried out. For
chapter II, we used data from the 8th NFI (1986–1994) which was the first inventory that combined information on Landsat TM 5 satellite images and ground reference plots (Tomppo et al. 2008). The proportions of predominant tree species and the age and development classes were calculated for each municipality, and subsequently as averages for each 50 × 50 km grid cell, by using the relative proportions of the municipalities as weights. The total proportion of forest land (TPF, average growth of 1 m3/ha/yr), unproductive forest area (average growth < 1 m3/ha/yr, relative to TPF), the proportions of forest under 40 years and over 60 years, and the average age of forest were selected as explanatory variables into the models.
For chapters III and IV, we used data from the 9th NFI. The satellite images from southwest, central and north Finland originate from 1998, 1996 and 2002–2003, respectively.
Using pixels that correspond to 25 × 25 m land area, we calculated forest cover (proportion of forest cover of the total area); MPS, mean patch size (ha); PD, patch density (number of patches per 100 ha) and TE, total edge (m) between forest and non-forest patches, for both spatial scales (1000 and 3000 m radii). For chapter III, forest was defined to include all pixels with > 60 m3ha–1 of timber, which refers to forests from the age class 61–80 years upwards in north Finland, and from 41–50 years
upwards in central and southwest Finland. For chapter IV, forest included pixels with > 152 m3ha–1 of timber, i.e. forests from the age class 51–70 years upwards in central and southwest Finland, and the most stocked mature forests in north Finland (Peltola 2003).
For chapter II, we also extracted data on the amount of total settlement (number of people) and on the number of people living in scattered hamlets (i.e. outside of population centres according to Finnish environmental administration in 1990). These were used as explanatory variables in the analysis, describing human impact.
2.4. Statistical analyses
Several statistical and analytical methods were used in the thesis. In chapter I, we modelled the Capercaillie population dynamics on a logarithmic scale with second order vector autoregressive models. We used the proportion of forests < 40 and < 80 years of age with a 7 years lag as explanatory variables. The lag was chosen to describe the population dynamic effects of lekking population destruction, i.e. the expected time until 90% of the adult males had died (log0.71[1 – 0.90]
6.72, calculated using parameters from Lindén 1981a). As an alternative to the forest variables, the decline was modelled as an undistinguished exponential declining trend, using the year of investigation as an explanatory variable. We
allowed geographical gradients in the population density by including the coordinates of the forestry board districts as explanatory variables in the models. Interactions between other explanatory variables and the coordinates were also allowed.
We used two approaches to address spatial synchrony in the process errors: 1) no correlation, 2) the process errors were assumed to be a 50–50% mixture of spatially correlated noise that decreased in correlation with distance, and compound symmetrical noise (correlation between all sites equal).
The models were fitted using maximum likelihood estimation, and standard errors of the parameters were calculated using parametric bootstrapping. Models with different combinations of the explanatory variables and error structures were compared with an information theoretical approach, according to the Akaike information criterion (AIC) and Akaike weights ( , Burnham & Anderson 2002).
For chapter II, we first calculated Spearman’s rank correlation coefficients separately for 130 grid cells throughout Finland, using the species-specific average abundances of Capercaillie and Moose in wildlife triangles over the years 1989–2007.
One grid cell had to be excluded from the calculations because of the lack of data. Secondly, we continued our analyses with a set of linear regression models for five separate regions (Fig. 2B). We analyzed how the association between abundance
of Capercaillie and Moose (averaged over the years 1989–1996 to temporally coincide with our land use data), represented by a regression slope, changed when the explanatory variables were included alone or as combinations in the stepwise regression models. The criterion of inclusion and exclusion of variables was always kept at P 0.05 andP> 0.05, respectively.
For chapters III and IV, we used a correlation-based principal component analysis for both spatial scales separately, to form a fragmentation index from mean forest patch size (MPS), patch density of forest patches (PD) and total forest — non-forest edge (TE).
In chapter III, the first principal component (PC 1) was selected to describe forest fragmentation. PC 1 embodied the three simultaneous effects of fragmentation: as PC 1 increased, MPS decreased, and PD and TE increased (Trzcinski et al.
1999, III). The correlation between PC 1 and forest cover (r = –0.70, P
< 0.0001) was removed by using a linear regression, and the residuals were used as an independent measure of fine-grain forest fragmentation. For chapter IV, the analysis did not produce a well-functioning measure of forest fragmentation. Regarding the first principal component, all indices (MPS, PD and TE) increased with increasing values of PC 1, whereas the second component represented MPS (see IV). Thus, neither of the components was selected for further modelling in chapter IV.
In both chapters III and IV, we analyzed the data using logistic regression models, where persisting versus ceased leks (binomial distribution, logit link function) was treated as a dependent variable.
Forest cover (III, IV), fine-grain forest fragmentation (only in III), and their interaction (only in III), were used in different combinations as explanatory variables in the models. In chapter III, also a class variable describing the study location and all relevant interactions with it were included in the models, whereas in chapter IV, separate models for each study location were built. Model selection in chapter III was performed by using AIC and Akaike weights ( ).