Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Determinants of distribution and abundance in the clouded apollo butterfly: a landscape ecological approach

Determinants of distribution and abundance in the clouded apollo butterfly: a landscape... Luoto. M.. Kuussaari, M.. Riui. H.. Salininen. J. and von Bonsdorff. T. 2001. Deterniinant.s of distribution and abundance in the clouded apollo butterfly: a landscape ecological approach. - Ecography 24: 601-617. Recent studies on ihe determinants of distribution and abiindance of animals at landscape level have emphasized the usefulness of the metapopulalion approach, in which patch area and habitat connectivity have often proved to explain satisfactorily existing patch occupancy patterns. A dilTerent approach is needed to study ihe common situation in which suitable habitat is difficuU to determine or does not occur in well-defined habitat patches. We applied a landscape ecological approach to study the determinants of distribution and abundance of the threatened clouded apollo Parnassius mnemosyiie butterfly within an area of 6 km^ of agricultural landscape in south-western Finland. The relative role ol' 24 environmental variable.s potentially atTecting the distribution and abundance of the butterfly was studied using a spatial grid system with 2408 grid squares of Q.25 ha. of which 349 were occupied by the clouded apollo. Both the probability of butterfly presence and abundance in a 0.25 ha square increased with the presence of the larval host plant Corydalis solida, the cover of semi-natural grassland, the amount of solar radiation and spalial autocorrelation in butterfly occurrence. Additionally, butterfly abundance increased with overall mean patch size and decreased with maximum slope angle and wind speed. Two advantages of the employment of a spalial grid system included the avoidance of a subjective definition of suitable habitat patches and an evaluation of ihe relative signiflcance of dilTeren! components of habitat quality at the same time with habitat availability and connectivity. The large variation in habitat quality was influenced by the abundance of the larval host plant and adult nectar sources but also by cliniatological. topographical and structural factors. The application of a spatial grid system as used here has potential for a wide use in studies on landscape-level distribution and abundance patterns in species with complex habitat requirements and habitat availability patterns. M. Luoio (miskaJuotoQiiwh.fi). Finnish Environment Inst., CIS and Remote Sensing Unit, P.O. Box 140 (Kesdkatu 6). FIN-0Q25! Hehinki, Finland. - M. Ktmssari and J. Salminen, Finnish Environment Inst., Mature and Land Use Dir.. P.O. Box 140 (Ke.sakatu 6). FfN-00251 Helsinki. Finland. - H- Rita. Dept of Ecohgv and Sy.stematics. Div. of Population Biology. P.O Box 17 (Arkadiankatu 7). FIN-00014'Vmi. of Helsinki. Finland. - T. von Bon.sdorff. The Finnish Mu.wum of Natural History. Div. of Mycology. P.O. Box 7 (Hameentie 153). FIN-00014 Univ.ofHehinki. Finland. Which factors explain the distribution and the abundance of a species has continued to be a central question in ecological research since the pioneering work of Andrewartha and Birch (1954). The development of spatial ecology during the last ten years has shifted the focus of the question from local and geographic scales Accepted 16 January 2001 Copyright © ECOGRAPHY 2001 ISSN 0906-7590 Printed in Ireland ~ all rights reserved ECOGRAPHY 24:5 (20011 to the intermediate, "'landscape" scale (Gilpin and Hanski 1991. Forman 1995, Hanski and Gilpin 1997, Tiltnan and Kareiva 1997). Empirical studies on distribution and abundance of species al landscape (or regional) scale can be categorized into two kinds of approaches (Wiens 1997, Hanski 1999). metapopulation approach in which the landscape is divided into two types of habitat, suitable patches and unsuitable matrix, and landscape ecological approach in which the complexity of real landscapes is more explicitly considered. Metapopuiation approaeh has been especially popular among butterfly ecologists (Thomas and Hanski 1997). Empirical butterfly studies have shown that for speeies which typically occur in relatively discrete habitat patches, existing patch occupancy patterns can be satisfactorily predicted by using only two variables, patch area and habitat connectivity (e.g. Thomas and Harrison 1992. Thomas et al. 1992, Hanski et al. 1995, Wahlberg et al. 1996; for an exception, see Dennis and Eales 1997, 1999). In the case of the Glanville fritillary Melitaea cinxia which occurs in networks of welldeflned patches of dry meadows in the Aland islands. Moilanen and Hanski {199f^) showed that the accuracy of the model predictions increased only slightly when the variation in loeal habitat quality and landscape structure were taken into account. For many species suitable habitat occurs in less distinct habitat patches and defining suitable habitat is difficulf because of continuous variation in habitat quality. In sueh cases metapopuiation approach is less useful. Thomas and Kunin {1999) suggested the use of spatial grid system in studying the population structure of sueh species. At the geographie scale, grid based distribution maps and environmental variables have been used to distinguish faetors affecting speeies range limits {Turner et al. 1987, Hill et al. 1999). Grid systems have also been used to prediet occurrences and to estimate declines at smaller geographic scales (e.g. Thomas and Abery 1995, Kunin 1998, Dennis and Hardy 1999. Cowley et al. 1999. Dennis et al. 1999, Leon-Cortes et al. 1999). A fine-scale spatial grid system has been commonly used to study butterfly movements and population structure at a loeal scale of one or a few closely located habitat patches (e.g. Singer 1972. Arnold 1983, Gall 1984. Vaisanen et al. 1994). Sometimes environmental variables measured from the grids have also been used to explain butterfly abundance {e.g. Thomas 1985. Vaisanen et al. 1994. Fiseher et al. 1999) and patterns of adult phenology (Weiss and Weiss 1998). Examples of extending the fine-scale grid based approach to the landscape level in studies of butterflies seem to be largely lacking {see exceptions in Ravenscroft 1994, Thomas et al. 1999). One advantage of the use of a spatial grid system is the possibility to utilize GIS data as a source of potentially important explanatory variables. Environmental variables like the oecurrence of various kinds of habitats and vegetation types as well as the patterns of variation in topography and climate are becoming inereasingly available from aerial photographs and satellite images via rapidly developing GIS techniques. We used a landscape ecological approach to study the factors affeeting distribution and abundance of the threatened clouded apollo butterfly Panuissius mnemosyne L.; Papilionidae in a system of 50 x 50 m^ grid squares covering 6 km^ of agricultural landscape in south-western Finland. Our aim was to evaluate the relative role of several lypes of potentially important environmental variables for the occurrence and the abundance of the butterfly. A total of 24 environmental variables were flrst grouped into seven variable families (food resources, habitat composition, habitat management, habitat structure, topography, climatology and habitat conneetivity), whieh were then used in a stepwise manner by adding one family at a time to construct multivariate models to explain butterfly distribution and abundance. Because of the high number of environmental variables, we put special emphasis on the logic of model building. The variable families, whieh were known to be biologically necessary (food resources and habitat composition), were brought into the model first, whereas the other fatnilies were entered into the model in the order of iheir statistical significance. Material and methods Clouded apollo Pai-nassius mnemosyne The clouded apollo is a threatened species in Finland as well as in most areas where it presently occurs in western and northern Europe (Somerma 1997, Aagaard et al. 1997, Kotiranta et al. 1998, Konvicka and Kuras 1999). The decline of the clouded apollo has been attributed to the ceasing of traditional management regimes, grazing and mowing, of semi-natural grasslands and coppiced woodlands {Vaisanen and Somerma 1985, Kotiranta et al. 1998). Several European studies have focused on the spatial and genetic population structure (Aagaard el al. 1997, Meglecz et al. 1997. 1999. Konvicka and Kuras 1999) as well as on genetic and morphometric differentiation between geographically separated populations (Vaisanen et al. 1991. Deseimon and Napolitano 1993), but detailed environmental requirements of the butterfly and its preimaginal stages have remained relatively poorly known. At any one location the larva of the clouded apollo seems to be monophagous on one of the species in the genus Corvdaiis {Fumariaeeae). but the particular species used varies geographically: C. solida in Finland (Somerma 1997 and our own observations), C intermedia in Norway (Aagaard and Hanssen 1989), C .solida and C. cava in the Czech Republic {Konvicka and Kuras 1999) and in Hungary (Meglecz et al. 1999). Females lay 1-3 eggs at a time in the ground vegetation when the host plants have already senescenced {Konvicka and Kuras 1999). ECOGRAPHY li-.i (2001) Study area ' ' The study area, ca 6 km" is situated in Somero, Hantala in south-western Finland. The upper course of the river Rekijoki flows in the study area across broad clay plains (Fig. t). Large parts of the study area are eharacterized by broad and flat expanses of intensively cultivated clay soils. Near the river Rekijoki, V-shaped valleys with steep meadow slopes dominate the landscape. River valleys are characterized by highly diverse types of land use and large semi-natural meadows and pastures, which are remnants from a long period of less intensive, traditional agriculture. At present Rekijoki is the largest area of mesic grasslands that currently exists in Finland (Kontula et al. 2000). A detailed description of the area is given by Luoto (2000) and IContula et al. (2000). To summarize the results of the survey we produced two variables measuring the distribution and abundance of the butterfly. Distribution was measured as a binary variable, presence ( > 0 observations) or absence (0 observations) of butterflies in each grid square during the whole flight season. Abundance was measured as the total number of butterflies observed during all visits divided by the number of visits in each grid square. Environmental variables Seven groups of environmental variables were recorded for each 0.25 ha grid square (n = 2408): 1) food resources. 2) habitat composition, 3) habitat management, 4) habitat structure. 5) topography, 6) climatology and 7) habitat connectivity. The measurement of the environmental variables and species distribution and abundance is dependent on the scale on which the measurements are made (Turner 1989. Wiens 1989, Forman 1995). Therefore a spatial grid system with a constant grid size was chosen to reduce errors of scale dependence of explanatory and response variables (see Turner 1989). Food resources Food resources of the larvae and adult butterflies were mapped in the spring and summer 1999, respectively. Average density of larval host plant (Corydalis .\olida) shoots m " - was estimated from each grid square in the field between 22 April and 20 May. In distribution modelling we also used a binomial variable of Corvdalls (0 ^ absence, 1 = presence). The abundance of the nectar plants (Amhriseus syhestris. Geranium sylvalicum. TrifoUum spp., Vicia eraeea, V. sepium. Ranunculus spp., Cainpamila patula) used by P. mnemosyne (our own observations) was recorded from 567 grid squares of semi-natural grassland (70'>^i of the study squares with any grassland within the square were .surveyed). The abundance of nectar plants was calculated using a scale from 0 to 10 (0 = no nectar plants, 10 ^nectar plants extremely abundant) between 15 and 23 June. Due to the hmited spatial coverage of nectar plant data it was only used in the analyses of the abundance of P. mnemosyne. Habitat composilkm The occurrence of different kinds of habitats in the study area (Fig. 1) was deflned from aerial low altitude blaek and white photographs (1:7500 scale) and from topographic maps. A total of 8 different land cover types were identified and assigned. The habitat patches were digitized and the derived habitat map was verifled in the fleld. Habitat composition was measured using GIS program Arc/Tnfo. vers. 7.2.1. The amount of different habitats for each 0.25 ha grid square was calculated in m^ from polygon coverages by a frequency function (Anon. I991J. 603 Butterfly survey In summer 1999 we surveyed the oeeurrence and the abundance of the clouded apollo in a total of 2408 grid squares, 50 x 50 m in size, within a 6 km- study area. We also used mark-release-recapture method to collect data on butterfly movements. Results on the movements will be reported in a separate paper. Butterflies were marked with an individual number and recaptured daily, weather permitting, between 6 June and 1 July. The location of eaeh eapture was marked on to a copy of a high-resolution aerial photograph (scale 1:3750). We divided the study area into eight sub-areas. One of the sub-areas was visited daily and the other seven sub-areas were visited Hve to eight times during the adult flight season. Fig. 1. A map on the occurrence of different kinds of habitats in tbe study area. Due to the cartographic reasons forest habitats are reclassified as one habitat class. ECOGRAPHY 24:5 (2001) Habitat structure Measures of habitat structure were calcuhtted for each 0.25 ha square with a 50 m buffer zone surrounding the square usitig the geostatistical progratn Fragstats, ver. 2.0 (McGarical and Marks 1994). The cell size used in Fragstats was 1 x 1 m, and the core area distance and border zone were defined as 0 m. Four measures of habitat structure were produced based on the habitat map in Fig. 1: mean patch size. Shannon's diversity index, mean shape index and area weighted mean shape index (for formulas see McGarical and Marks 1994). Topography Six topographical variables were calculated from each grid square: 1) mean altitude. 2) the lowest point. 3) the highest point, 4) relative altitude (highest - lowest point), 5) mean slope angle and 6) maximum slope angle. Variables were calculated using Arc/Info Grid from a digital elevation model (Anon. 1997). CHimitology An estimate of maximum theoretical solar radiation for each grid square was produced using a computer model of clear sky insolation and exposure of different slopes. In the following calculation, variation in thermal conditions across the study area is a direct result of solar radiation on different slope exposures (Griffiths 1985): Solar radiation = [cos(altitude) x sin(slope angle) xeos(ground direction — sun direction) + sin(altitude) xcos(stope angle)] Habitat connectivity Habitat connectivity was trteasured only for one habitat lype, semi-natural grassland, which was expected to be the single most important habitat type for the clouded apollo. Connectivity (S) to semi-natural grasslands in the neighbourhood of grid square i was measured following Hanski (1994): where A| is the area of semi-natural grassland (in ha) in square, and d^j is the distance between squares i and j (in km). For each study square, connectivity to seminatural grasslands in all other study squares was measttred up till two km distance. Two km was set as the upper limit, because for longer distances no accurate habitat maps were available. The exponential function gives more weight for those squares, which are close to the focal square. Value 1 was used for paratneter ex, and it was selected based on mark-recaptut-e data on the movements of the clouded apollo (Kuussaari and Luoto ttnpubl.). Adjusting variables Spatial autocorrelation The probability of occurrence of P. mnemosyne in one grid square is unlikely to be independent of whether the species occurs in a neighbouring square. This will generate spatial autocorrelation that cannot be modelled satisfactorily by environmental covariates (Augustin et al. 1996). In this study, we attempted to minimize the effect of spatial autocorrelation of data by using two adjusting spatial autocorrelation variables. For the analysis on the factors affecting the presence and absence of butterflies the adjusting variable was calculated as the number of occupied squares among the eight neighbouring squares of each sttidy square. For the analysis on butterfly abundances the adjusting variable was calculated as the average number of captures divided by the number of visits among the neighbours for each grid square. In the case of the grid sqttares at the borders of the study area we calculated the average of available neighbouring squares. Study effort The probability of observing P. mnemosyne in a 0.25 ha grid square is affected by the nutnber of visits, i.e. study effort (Dennis et al. 1999). Therefore the study area was divided into two parts differing in their sampling effort: "the core sub-area" and "the other sub-areas'". The cover of the core sub-area was 375 grid squares (16%) and cover of other areas was 2033 grid squares (84'V(i). Sampling effort (the total amount of visits in an average grid square) was ca 15 times in the core sub-area ECOGRAPHY 24:S |2(Kll) (1) In this calculation altitude is the angle of sun above horizon. Sun direction is the direction of the sun from the south. It increases positively to west and negatively to east (south ^ 0 . east = - 9 0 . west = +90). Slope angle is slope angle of the ground (flat area = 0. vertical = 90). Ground direction is the downgrade direction of the ground from the south. It increases positively to west and negatively to cast (south = 0. east = - 90, west =^ + 90). Solar radiation was calculated as the average of three moments during clouded apollo's daily activity period: 9.00, 12.00 and 15.00. The average wind speed for each grid square during the butterfly's daily activity period (8.00-17.00) was produced by the Finnish Meteorological Institute based on a field survey and topographical and habitat maps. Estimation of the average wind speed in June was done using a four-class scale: 1) <0.5 m s " '; 2) 0.5- 1.5 m s " ' ; 3) 1.5-2.3 m s " ' : and 4) > 2.3 m s " ' . Management Data on eattle grazing, the primary management method of the semi-natural grasslands in the study area were collected during the fieldwork in summer 1999. In the analyses, we used cattle grazing as a binary variable (0 = no grazing. 1 = grazitig) for each grid square. 604 Adjusting i HHbltal com;osltior htebitat connectivity TopogiafSiy StructuG CNmatQtogy Management Bemainlng fcur groups * — Ramalning three groups nainire two groups Remalnlrg one gmup Fig, 2, The logic of model building in explaining the patterns of distribution (logistic regression) and abundance (GLM regression) of the clouded apollo. Adjusting variables were entered into the models first, food resources secondly and habitat composition variables at the third step. Next the most predictive of the remaining five variable families was added; topography, habitat structure, climatology, habitat connectivity and management. This procedure was repeated, and the next predictive variable family added until no more statistically significant (p<O.Ol) variables remained, tial autocorrelation as an adjusting variable and the other without spatial autocorrelation. Model building for both distribution and abundance was started by entering the adjusting variables, study effort and spatial autocorrelation (see summary of the logic of model building in Fig. 2), Adjusting variables were kept in the model regardless of their statistical significance. Next, variables according to their biological importance were selected for inclusion. Thereafter, the remaining variables were entered in the order of their statistical significance. The statistical significance after including (or excltiding) a variable in a model was calculated. This procedure was repeated until no more important variables remained. At eaeh step, variables entered in previous steps were tested for redundancy (see Yee and Mitchell 1991. Tonteri 1994). Distribution modelling First, all explanatory variables were described by comparing descriptive statistics between the grids in which P. nmemosyne was present versus absent. We used univariate analysis to assess individual model variables independently from each other and to obtain information on each variables' role before proceeding with model development and testing, a procedure that is also recommended by Pereira and Itami (1991). Next, multiple logistic regression, a form of generalized linear modelling (Rita and Ranta 1993. Trexler and Travis 1993. Sokal and Rohlf 1995) was used to analyse the relationship between environmental variables and distribution of P. mnemosxne in 2408 grid squares (see Brito et al. 1999, Carroll et al. 1999), using the SPSS-PC 7.0 software paekage. and ca 6 times in the other sub-areas during the adult flight season. In the analyses on butterfly distribution and abundance the variable study effort was taken into account as a binary variable: in the core subarea effort = 1. and in the other sub-areas effort = 0, Statistical analysis Logic of model building Recent papers have criticized automatic stepwise procedures, as they are not able to select the most influential from a set of variables (James and McCulloch 1990, Bustamante 1997, MaeNally 2000). Therefore, to compile models for distribution and abundance of P. mnemosyne, the explanatory variable groups were selected to the model in a priori defmed manner taking into account the existing knowledge on the biology of the focal species (Fig. 2). We used a combination of forward selection and backwise elimination procedures with a strict \% significance level as the criterion. Food resources, the most primary need of P. nmemosyne. were ranked first, and habitat composition secondly due to the biological importance of the type of habitat. Habitat structure, habitat management, topography, climatology and habitat connectivity variables were ranked equally important and they were entered into the multivariate models in the order of their statistical significance. Two kinds of models were constructed to explain butterfiy distribution and abundance: one using spaECOGRAPHY 24:.^ (200!) Abundance modelling The abundance of clouded apollo (captures/visits) in each grid square was related to the environmental factors using generalized linear models (GLM) (Austin et al. 1984, McCulIagh and Nelder 1989, Nichoils 1991. Crawley 1993), as implemented in the statistical program GLIM 3,77 (Payne 1986), All survey grid squares where P. mncniosvne was found to be present in summer 1999 were included in the analyses. In addition, also those grid squares were chosen to the analysis where the logistic regression model estimated the probability of P. nmemosyne presence over 0.5. With this selection procedure the abundance analyses were focused on the observed distribution area and also on those grid squares where the environmental factors seemed to be suitable for occurrence. The total amount of grid squares in the analyses was 412. A Poisson distribution of errors was assumed and abundance of P. nmemosyne was 605 related to explanatory variables via a logarithmic link function (Mclntyre and Lavorel 1994). First, all the 24 explanatory variables were related separately to the response variable, using data from the 412 grid squares. The statistical significance of the change in deviance after including (or excluding) an explanatory variable in the model was determined by an F-ratio test (McCullagh and Nelder 1989, Crawley 1993) with a 1% significance level as the criterion. Further, we tested also the non-linear relationship between explanatory variables and abundance by including the quadratic terms of continuous variables into the model (see Margules et al. 1987, Austin et al. 1996). The derived models were evaluated by plotting residuals against the fitted values and the explanatory variables, and by plotting residuals in a normal probability plot as suggested by Nicholls (1991). tree saplings, which were situated adjacent to seminatural grasslands, had low or moderate incidence. The eastern part of the study area formed an exception to the distribution oi P. mnemosyne: regardless of relatively large cover of semi-natural grassland and presence of Corydalis solida no P. mnemosyne were found (see Fig. 3a, b). Distribution Comparison of descriptive statistics of various environmental variables between occupied and unoccupied grid squares showed that there were large statistical differences between those two groups of grid squares (Table I). Occupied squares were located in areas where larval host plant was abundant, the covers of semi-natural grassland and deciduous forest were higher and habitat structure was more heterogeneous than in unoccupied grid squares. In addition, occupied grid squares were characterized by steep slopes and low wind speed. Before the entry of explanatory variables the overall correct classification rate was 85.51%, because in 2059 of 2408 grid squares P. mnemosyne was absent (Table 2). The distribution modelling was started by fitting the "adjusting variables" in the model (spatial autocorrelation. p<O.OOI and effort p = 0.50). Together these variables predicted 207 of 349 (59.3%) P. mnemosyne grid squares correctly and the overall correct classification rate was 91.7%. Next the binomial Corydalis variable ( p < 0.001) was fitted of the food resources group. After this, model building was continued by fitting the best explanatory variable of the habitat composition group, namely the cover of seminatural grassland (p<0.()01). No other variables of habitat composition group could be fitted to the model. After this inclusion the overall correct classification rate was 93.9%. In the preceding steps only radiation (p < 0.002) of the climatological variable group could be fitted to the model. Finally, the overall classification of predicted and observed grid squares was 93.8 and 75.4'^!) of the presence grid squares were correctly classified. Coefficients and standard errors are given in Table 3. There was a statistically significant ( p < 0.001) spatial autocorrelation in the presence grid squares. When the distribution modelling was done without the spatial autocorrelation variable also habitat connectivity and minimum altitude could be fitted to the model in addition to the previous three variables (see Tables 2 and 3). In this case the overall correct classification rate was 92.7%. Predicted probability of occurrence of P. mnemosyne calculated by logistic regression with spatial autocorrelation are shown in Fig. 3d. and without spatial autocorrelation in Fig. 3e. tICOGRAPHY 24:5 (2001) Characteristics of hotspots To define the characteristics of high abundance grid squares, descriptive statistics for selected environmental variables of hotspot and other populated grid squares were calculated and statistically tested using Mann-Whitney U-test. Following Prendergast et al. (1993) we defined hotspots as the top 5% of the grid squares, ranked by the abundance of the butterfiy (captures/visits). Predictive capacity of different types of environmental variables To determine the predictive capacity of different environrnental variable families in estimating distribution and abundance of P. mnemosyne we related the variable families separately to the response variable using logistic and GLM regression modelling. The overall correct classification rate (distribution model), change of deviance (abundance model) and the number of significant variables were calculated for all factor groups. The predictive capacity of each factor group was tested with adjusting variables included. Results A total of 349 (14.5%) of 0.25 ha study squares were found to contain P. mnemosyne. The distribution of P. mnemosyne was clearly associated with the river valleys having extensive cover of semi-natural grasslands (Fig. 3a). high density of Corydalis .solida (Fig. 3b) and high amount of solar radiation (Fig. 3c). Butterflies were absent or showed very low incidence in agricultural fields, built-up areas and coniferous forests (Table I). Mixed and deciduous forests, fellings and 606 ative) and three topography variables, showing often a quadratic relationship between abundance. Of the adIn the squares occupied by the butterfly abundance justing variables spatial autocorrelation was signifivaried between 0.05 and 9.50 captures/visit (mean = 1.05, SD = 1.28; Fig. 3). In 104 of 349 occupied squares cantly related to the abundance of P. mnemosyne. The construction of the multivariate GLM model to (29.8%) the abundance was over one capture/visit. When the effect of environmental variables was tested e.xplain butterfly abundance is shown step by step in singly, 16 of the 22 explanatory variables were signifi- Table 5. The set of variables, which together best cantly (p<0.01) related to the abundance of P. explained the abundance of P. mnemosyne in the multimnemosyne (Table 4). The variables accounting for the variate GLM model with adjusting variables, indicates largest change in deviance were the abundance of nec- that the abundance increased with the amount of Corytar plants (positive effect), the cover of semi-natural dalis., cover of semi-natural grassland, radiation and grassland (positive), the cover of agricultural area (neg- mean patch size, and decreased with maximum slope Abundance li-tWspotgrtdsquare Chxidai apotlo abundance f • 0 (capOjtes/viail) C _ 0.05.0.99 1.00-1,99 2.00-9.50 Fig. 3. (a) Observed abundance (captures/visits) of clouded apollo Parnassius mnemosyne in the 50 x 50 m grid squares and the occurrence of semi-natural grasslands. The hotspots (the top 5% of the squares with the highest butterfly abundance) are marked by a star symbol (b) Densily (shoots m " -) of larval host plant CorydalLs solida. and (c) solar radiation in the study area. Predicted distribution of P. mnemosvne according to the logistic regression models presented in Table 3: (d) with spatial autocorrelation, and (e) without spatial autocorrelation. PredWai pt«aiflity o( occun»x» (viiih spatial aita»(Te4aikm O00(W)20 rrt]021050 aHO |BO Profcled prtiabily rt occunwce (iwthixit spaHal aulDcoffBlHtion) I'lo.21-11.50 wio.5i-o.ao •io.ei-1.00 ECOGRAPHV 24:5 (21)01) Table I. A list of environmental variables calculated for eaeli of the 50x50 m grid squares in the study area and used as explanatory variables of the distribution and abundance of P. itmemo.syne butterfly. Descriptive staiisties (mean and SD) for ihe two types of grid squares in the study area; I) P. mnemosyne present (n = 359), and 2) P. mnemosyne absent (n = 2059), P-vatues were derived from Mann-Wliitney U-test. Variable families and variables P. mnemo.nyne absent mean + SD 2.21 +5.56 6.46 ± 1.84 1380+1079 150 + 397 101 +372 169 + 506 404 + 817 102 + 403 12.0 + 55 178+470 0.27 ± 0.45 1.47 0.50 1.63 0.76 + + + + 0.25 0.46 0.28 0.45 P. mnemosyne present mean ± SD 9.35 + 8.18 6.55 ±1.94 342 + 656 -25.618 -0.456 -15.948 -28.121 -18.444 -10.838 -2.078 -1.213 - 16.754 -7.259 -21.438 -19.800 -18.477 -6.118 -16.858 - 24.228 -23..Ml -17.678 -24.258 -21.188 -21.879 -4.716 - 10.583 -18.544 0.000 0.585 0.000 0.000 0.000 0.000 0.038 0.225 0.000 0.000 0.000 0.000 0.000 0,000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 Food resources Corydalis sollda (host plant of larvae, shoots m~-) Nectar plants (only squares with semi-iiiit. grassl., 10-class scale) Habitat compo.sition (m') Agricultural area in the grid square Semi-natural grassland in the grid square Deciduous forest In tlie grid square Mixed forest in (lie grid square ConiferoLLS forest in the grid square Felling or tree sapling in the grid square River in the grid square Built area in the grid square Management (only squares witli semi-nat. grassl.) Management of semi-natural grassland (()/!) Habitat structure Area weighted mean shape index Mean pateh size (ha) Mean shape index Shannon's diversity index Topography Altitude min (m a.s.l.) Altitude ma.\ (m a.s,l,) Altitude relative (m a.s.l.) Altitude mean (m a.s.l.) Slope angle mean (°) Slope angle max (") Climatology Radiation (0-100) Wind speed (4-dass scale) Habitat connectivity Habitat conneetivity of semi-natuial grassland 1151+758 373 + 588 314 + 536 142 + 392 79.9 + 358 67.9+ 116 29.7 ± 180 0.47 ±0.50 1.78+0.22 0.22 + 0.09 1.70 + 0.14 1.20 ±0.28 66.0 + 71.7 + 5.67 + 68.4 + 8.07 + 7.60 6.84 3.41 7.27 3.61 80.5 + 8.09 82.6 + 6.90 2.18-K 2.76 81.3+ 7.41 2.91+3.50 4.05 ± 4.54 69.9 + 2.83 2.58+1.18 2.99 + 0.58 11.0 ±3.96 70.8 + 4.60 1.85 ±0.89 3-59 + 0.33 angle and wind speed (Table 6). In addition, the quadratic function of mixed forest showed also statistical significance in the abundance model. The flnal models of abundance of P. mnemosyne with coefficients and standard errors are given in Table 6. The model explained 35.8'>ii of the variation in abundance of P. mnemosyne (total deviance change 182.1 out of 509.1). There was a statistically significant (p < 0.001) spatial autocorrelation in the abundance of P. mnemosyne grid squares. When the abundance modelling was done without the spatial autocorrelation variable the number of explanatory variables increased remarkably. Wind speed, slope maximum and mean patch size, which were included in the autocorrelation model, got replaced by quadratic term of deciduous forest, mean altitude, quadratic term of mean slope angle and quadratic tertn of Shannon's diversity index (Tables 5 and 6). distribution pattern in the study area (hotspots are indicated by a star sytnbol in Fig. 3). Mann-Whitney U-test showed statistical difference in several environmental variables between hotspot and other grid squares (Table 7), Hotspots were characterized by high amount of nectar plants and larval host plant. Further characteristics of hotspot grid squares included large cover of semi-natural grassland and deciduous forest, and low mean altitude. Predictive capacity of different environmental variables Predictive capacity of different variable families for distribution and abundance of P. mnemosyne showed clear differences (Table 8). Food resources (overall correct classification 93.52'Mi) and habitat composition (93.36%.) groups were the single most powerful explanatory variable groups for distribution, whereas habitat structure and topography were the poorest variable families in the predictive power. The most powerful groups of abundance modelling were habitat composiECOGRAPHY 24:^ (Mil) Characteristics of hotspots The hotspoEs (the top 5"A, of the squares with the highest butterfly abundance) had a relatively dispersed tion (33.35% change of deviance), topography (25,39%) and food resources (23.82%). The number of significant variables varied between 0 (management and habitat connectivity in the abundance model) and 4 (habitat composition in the abundance model) within the variable families. determinants of distribution and abundance. Additional factors contributing to the occurrence of butterflies included variables associated with topography, climatology, habitat structure, habitat connectivity and spatial autocorrelation in butterfly occurrence. Factors affecting distribution Discussion Employment of a spatial grid system proved to be a useful way to estimate simultaneously the role of many potentially significant environmental variables in determining the distribution and abundance of the clouded apollo. Availability of semi-natural grassland together with the necessary food resources for the larvae and adult butterflies turned out to be the most significant Parnassius nmemosyne butterflies occur in senii-natura! grasslands on the slopes of the river valleys. The set of environmental factors, which together best explained the presence and absence of/', mneniosyne in a multiple logistic regression, indicates that the probability of P. nmemosyne occurring in a square increases with the presence of the larval host plant Corydalis solida, cover of semi-natural grassland and solar radiation. In the multiple logistic regression model there was no statisti- Table 2. Summary of the logistic regression mode! building procedures (forward selection and backwise deletion of explanatory variables) for distribution of P. mnemosyne. Both forward selection and backwise deletion of variables were employed in the model building, using a strict criterion (p<0,01) for their inclusion or exclusion. The model is built using P. nmemosyne distribution (presence/absence) and environmental data from the 2408 grid squares in the study area. Model building procedures with and without a spatial autocorrelation variable are presented. Model with spatial autocorrelation Model No terms added I (adjusting variables) + Spatial autocorrelation + Effort II (Food resources) + Corydalis (0/1) (larval host plant) III (Habitat composition) + Semi-natural grassland IV (Climatology) 4- Radiation * * p < 0 , 0 1 : ***p<O.OOI, Model without spatial autocorrelation Model No terms added I (adjusting variables) + Effort II (Food resources) + Corydalis (0/1) (larval host plant) III (Habitat composition) + Semi-natural grassland + Deciduous forest + Mixed forest IV (Habitat connectivity) + Habitat connectivity — Deciduous forest — Mixed forest VI (Topography) -1- Altitude minimum VII (Climatology) -1- Radiation B SE % of total correct classificat. 85,51 0,8469 4.7183 0.1546 0.0487 0.0460 2.1664 0,0251 0,0334 -0,0897 0,0815 0.1367 0.3106 0.0115 0.0129 0.0141 0.2325 0.0147 0.0151 0.0130 0.0200 85,51 87,79 91,20 91,28 91.57 92,32 92,28 92,23 92.36 92,73 0,1351 0.3389 0,2990 0,0787 0,0657 0,2063 0,0214 0,0380 -0,1519 0,0858 *** ++* SE % of total correct classificat. 85,51 1.0077 -0,0081 2,7846 0,1091 0,0645 0.0443 0.2347 0.3402 0,0143 0,0209 91,65 91,65 93.52 93.94 93.81 0,5088 0,0000 0,1806 0,1675 0,0615 *** n,s. •** •** • ** *** *++ ** *** n.s. 11.S. *** ++* p<0.01; ***p<0,001. ECOGRAPHY 24:5 (2001) Table 3. Variables in final multiple logistic regression models (see Table 2) for the P. mnemosyne distribution showing coeffieients, standard errors and significance values. Models with and without a spatial autocorrelation variable are presented. Model wilh spatial autocorrelation Model Constant Spatial autocorrelation Effort Corydali.s (O.'l) (larval host plant) Semi-natural grassland Radiation **p<0.0]; ***p<O.0()l. Model without spatial autocorrelation Model Constant Effort Corvdalis (0/1) (larval host plant) Semi-natural grassland Habitat connectivity Altitude minimum Radiation **p<0.01; ***p<0.001. B -12.168 1.487 2.944 0.115 2.172 -0.102 0.105 SE 1.871 0.260 0.354 0.014 0.242 0.013 0.020 Exp(B( 4.424 18.999 1.122 1.122 0.903 I.Ill R 0.124 0.184 0.188 0.188 -0.166 0.110 P ** **+ *** *** **« *** B -10.528 6.8K4 0.851 2^71 0.101 0.065 SE 1.522 0.049 0.278 0.377 0.0! 5 0.02 i Exp(B) 2.341 1.990 13.080 1.106 l.0(S7 R 0.313 0.061 0.150 0,151 0.0615 P *** •*• ++* *** **• ** cal significance between habitat connectivity, habitat management, habitat structure and topographical variables, and distribution pattern of P. mnemosyne. It should be noted, however, that our measures of spatial autocorrelation and connectivity to surrounding seminalural grasslands were strongly correlated (Pearson correlation coefficient. R ^ 0.48. p<().01). When the model was constructed without spatial autocorrelation as an adjusting variable, the effect of habitat connectivity on the grid square occupancy became highly significant (Table 3). Due to the limited spatial coverage of nectar pianl data no analyses on the role of nectar plants for the distribution of P. mnemosyne were performed. The univariate analysis revealed clear differences in topographical and habitat structural variables between squares wilh and without P. mnemosyne (Table 1). However, these variables were of minor importance after the biologically primary variables (food resources and habitat composition) had been included to the distribution model (Table 3). The sequence of selecting variables used here probably obscured the fact that some variation in distribution could have been explained by food resource and habitat composition variables as well as by topographical and habitat structural variables. Topographical and structural variables are intercorrelated with food resources and habitat composition variables, because semi-natural grasslatids are eoncentrated in steep river valleys with low minimum altitude and small mean pateh size (see MacNally 2000). The clouded apollo was absent in the eastern part of the study area and showed very low and scattered incidence in the .study area at the north-eastern edge (Fig. 3). Those areas had high or moderate density of larval host plant and solar radiation and relatively large 610 cover of semi-natural grassland. They differed from the clouded apollo distribution area mainly by more fragmented pattern in the occurrence of semi-natural grassland (i.e. low connectivity) (Figs I and 3). The predicted probability of oeeurrence in both logistic regression models gradually decreased towards north-east (Fig. 3d-e). Factors affecting abundance The GLM modelling procedure and analysis of predictive capacity of different factor groups showed that habitat composition and food resources variables were the most important factors explaining the abundance of P. mnemosyne. The abundance increased with cover of semi-natural grasslands and density of Corydalis. After quadratic function of semi-natural grasslands was fitted in the model, the abundance of nectar resources was no longer significant in the model. This suggests that the density of nectar plants was generally sufficient for adult butterflies in the study area. Nevertheless, the density of P. mnemosyne was especially high in the grid squares with high atnount of nectar plants (Table 7). Mixed forest patches, whieh were included as a statistically significant variable in the model for abundance, have certain eeologieal features for abundance. It seems that small forest patches in the semi-natural grasslands did not decrease butterfly abundance - on the contrary they increased it. This can be an implication of high larval host plant density and possibly high larval densily in those areas due to high oviposition rate or high survival of larvae. ECOCiRAPllY 24:.i a Climatological conditions infiuence activity patterns and growth of insects in all life stages (Taylor 1981. Turner et al. 1987, Weiss et al. 1988, 1993, Weiss and Weiss 1998). For the clouded apollo climatological variables, like wind speed and solar radiation, were important predictors of abundance. Wind speed showed a negative relationship with butterfly abundance, whereas the correlation of butterfly abundance with solar radiation was positive. As reported by Dover et al. (1997), high wind speed hinders fiying of butterflies and sheltered areas seem to be very important for them in open agricultural landscapes. Variation in thermal conditions across the study area is a direct result of solar radiation on different slope exposures (Griffiths 1985). South-facing slopes with high solar radiation have certain special ecological features for the clouded apollo, which occurs in the study area in the northern margin of its European distribution. In spring the lar- vae need to complete their development before their host plants senesce. Slopes with high radiation are the warmest on average and there is less evaporation caused by wind (Mansikkaniemi and Laitinen 1990). This suggests that developmental conditions for larvae tnay be more favourable on sheltered south-facing slopes than in other areas. In addition, when the amount of radiation increases, the potential duration of daily activity period and potential time for reproduction (e.g. number of eggs) of adult butterflies increases. Many studies have indicated the importance of habitat and topographical heterogeneity in explaining species' spatial patterns (Weiss et al. 1988. 1993, Verboom et al. 1991, Gustafson and Gardner 1996, Moilanen and Hanski 1998). In this study, mean patch size showed a positive relationship with the abundance. This can be explained by the fact that the abundance of P. mncmosyne was related highly and positively to the Table 4. The change in the deviance from a model f"or abundance of P. mnemosyne wilh no terms to that containing each environmental variable on ils own, A Poisson error distribution for the abundance of /'. mnemosyne is assumed and the response variable is linked to the set predictor variables via a logarithmic function. Only those linear and quadratic functions of Ihe variitbles that account for a statistically significant (_p<0.01) change in ihe deviance are shown, except linear non-significant terms are shown which have a significant quadratic function. Moreover, only those quadratic functions where the change from linear to quadratic model is significant are listed. Direction of ihe effect is presented by + and - symbols. Variable No terms added Food resources Nectar planls (aduit food plants) Corydiilis {larval host plant) Habitat composition Semi-natural grassland Semi-natural grassland-|-semi-natural grassland^ Agricultural area Coniferous trees River Built area Deciduous forest Deciduous forest 4-deciduous forestMixed forest Mixed forest-I-mixed forestHabitat structure Area weighted mean shape index Area weighted mean shape index-I-area weighted msiTopography Altitude mean Altitude maximum Altitude minimum Slope maximum Slope maximum-I-slope maximumSlope mean Slope mean + slope mean^ Climatology Wind speed Wind speed+wind speed^ Management No significant change in the deviance Habitat connectivity No significant change in the deviance Adjusting variables Spatial autocorrelation Spatial autocorrelation + spatial autocorrelation^ Effort (0\l) p<0.01; ***p<O.OOL ECOGRAPHV 24:5 (20011 Deviance 509.07 DF 411 Direclion of the effect 410 410 410 409 410 410 410 410 410 409 410 409 410 409 410 410 410 4fO 409 410 409 410 409 36.62 12.67 38.92 60.06 37.86 9.38 8.43 7.36 0.33 25.17 0.03 18.70 23. S 5 31.23 46.24 41,13 40.75 2.49 11.70 1.26 19.99 18.54 .14.40 +** *** *** ,^^ ** «* ** n.s. *** n.s. *** *** *** ««* •«* + + -h n.s. ** n.s. «** «« * Table 5. Summary of the multiple regression model building procedures (forward selection and backwise deletion of explanatory variables) for abundance of P. mnemosyne. Both forward selection and backwise deletion or variables were employed in the model building, using a strict criterion (p<O.OI) for their inclusion or exclusion. The model employs a Poisson distribution of error assumption and a log link, and is built using P. mnemosyne abundance and environmental data from the 412 grid squares in the study area. Model building procedures with and wiihout a spatial autocorrelation variable are presented. Model building procedure with spatial autocorrelation Model No terms added 1 (adjusting variables) + Spatial autocorrelation + Spatial autocorrelation+ Effort (core area/other area) II (Food resources) + Nectar plants (adult food plants) + Corydalis (larval host plant) III (Habitat composition) + Semi-natural grassland + Semi-natural grassland— Nectar plants (adulf food plants) + Mixed forest + Mixed forest" IV (Climatology) + Wind speed — Semi-nalural grassland" + Radiation " (Topography) V + Slope maximum VI (Habitat structure) + Mean patch size **p<0.01; ***p<0.001; Model building procedure without spatial autocorrelation Model No terms added I (adjusting variables) + Effort (core area/other area) II (Food resources) + Nectar plants (adult food plants) + Corydalis (larvai host plant) III (Habitat composilion) -I- Semi-natural grassland + Semi-natural grassland" + Agricultural land + Coniferous forest + Deciduous forest + Deciduous forest" — Semi-natural grassland' — Coniferous forest + Mixed forest + Mixed forest— Agricultural land IV (Topography) + Altitude mean + Slope mean + Slope mean^ V (Climate) + Radiation V! (Habitat structure) + Shannon's diversity index + Shannon's diversity index• * p < 0 . 0 1 ; ***p<0.001. Deviance 509.07 507.88 465.07 444.20 407.50 395.26 387.98 379.42 379.40 372.51 410 409 408 407 406 405 404 403 402 403 404 DF Change in deviance Deviance 509.07 439.26 414.30 411.85 402.13 388.43 355.30 342.78 347,87 347.11 337.76 328.47 331.62 325.31 318.89 313.48 410 409 408 407 406 405 404 405 404 403 402 403 402 401 400 DF Change in deviance 411 69.80 24.96 2.45 9.72 13.70 33.126 12.53 5.09 0.76 9.34 9.29 3.15 6.31 6.42 5.41 65.16 24.64 2.43 9.84 14.32 37.76 14.77 5.93 0.88 11.14 11.37 3.83 7.80 S.07 6,90 +** *** n.s. «>« if** * * * * • * n^. n.s. • «* *** *** *» '*« 4U 1.18 42.71 20.97 36.70 12.24 7.28 8.55 0.02 6.89 5.68 4.70 3.01 17.16 0.62 11.27 1.37 5.84 9.60 1.01 6.81 0.^. 37.55 19.26 36.66 25.14 7.60 9.11 0;Q2 7:44 5.61 4.96 3.19 19.04 0.68 12.88 1.56 6.78 11.151 1.21 8C27 • ^ . • *• i*. n.s. yn.ii 382.46 379.46 362.30 362.91 351.65 350.28 1,UAA 334.85 333.83 327.02 «* I1.S. ri.s. n.s. #** n.s. «»* n.s. ** ECOOBAPHY 24:5 (2001) Table 6, Variables in fmal multiple GLM regression models for the P. mnemusyne abundance showing coefficients and SE (see Table 5). Models with and without a spatial autocorrelation variable are presented. Model with spatial autocorrelation Variable c, intercept. Spatial autocorrelation Spatial autocorrelation-^ Effort (core area/other area) Corydalis (larval host plant) Semi-natural grassland Mixed forest Mixed forestWind speed Radiation Slope maximum Mean paEch size Coefficient -4,323 1.854 -0.5488 0.3860 0.02765 0.05713 0.1030 -0.006786 -0.3772 0.03958 -0.03679 1.765 SE 0,9391 0,3255 0.1405 0,1486 0,00675 0,00897 0,03127 0,00205 0,08333 0,01274 0,0!576 0,7293 Model without spatial autocorrelation Variable c, intercept. Effort (core area/other area) Nectar plants (adult food plants Corydalis (larval host plant) Semi-natural grassland Deciduous forest Deciduous forest^ Mixed forest Mixed forestAltitude mean Slope mean Slope mean^ Radiation Shannon's diversity index Shannon's diversity indexCoefficient -6.544 0.1290 0.1456 0.02978 0.05694 0.1198 -0.004773 0.1336 -0.006564 -0.02204 0.1623 -0.01177 0.03791 3.932 -i.575 SE 1,531 0,1456 0,03617 0,007136 0,01252 0,02270 0,001490 0,03192 0,001998 0,008844 0,08557 0,004965 0,01329 1,514 0,6331 cover of semi-natural grasslands and in the most optimal areas for abundance the grid squares had a high mean patch size because most of the grid squares were covered by semi-natural grasslands. Relationship between maximum slope angle and abundance was tiegative, which is partly due to the fact that landslides are frequent on the steepest slopes. In areas of landslide scarps low amount of nectar plants and especially Corydah's is characteristic. Habitat loss caused by overgrowth after abandonment of the traditional use and deteriorating habitat quality of semi-natural grasslands are the main threats of decline of P. mnemosyne (Vaisanen and Somerma 1985, Kotiranta et al. 1998). In this study, grassland management by cattle grazing did not show any statistical significance for abundance or distribution of P. mnemosyne. The accuracy of the management variable may influence this result, because only crude binomial data of grazing were available. For example, high intensity of grazing may decrease butterfly abundance by reducing the abundance of nectar flowers and larval host plants. Although the regression models uncovered statistieally significant relationships between the explanatory variables, and the butterfly distribution and abundance, their goodness was relatively modest. More than half of the variation in the abundance remained unexplained, whereas the overall correct classification rate of logistic regression was 93.8%, and 75.4% of the occupied grid squares (sensitivity of the model) were correetly classified. One potential reason for the modest fit of the models may be the scale itself (see Levin 1992). The distribution of a butterfly can be difficult to predict since it is often a combination of large-scale biogeographical variation and fine-scale eeological variation (see Wiens 1989, Levin 1992). The different life stages of the butterfly react on very different scales to their environment. For example, very small scale (metres to centimetres) microclimate factors can have pronounced effect on larval growth rate (Ravenscroft 1994), whereas the population density of adult butterflies is often affected by large scale habitat patterns (Hanski 1994, 1999). The area of grid square (0.25 ha) can be too small for this kind of modelling approach. On the other hand, a large amount of the environmental variation in the heterogeneous landscape would be lost at a coarser scale. An- Table 7, Environmental characteristics of "hotspots" of P. mnemosyne vs other occupied grid squares. Hotspot grid squares were defined as the top 5% grid squares (here 17 grid squares out of 349), ranked by the abundance (captures/visits), P-values were derived from Mann-Whitney U-test (n, = 17, n^ = 332). Variable Nectar plants (0-10) Corydalis (shoots m~-) Cover of semi-natural grassland (m-) Cover of mixed forest (m~) Cover of deciduous forest (m^) Mean altitude (m) Radiation (%) Wind speed (1-4) Mean patch size (m^) Management, grazing (O/I) Habitat connectivity Hotspot - grid squares mean + SD 8.06+ 1.30 13.34 + 8.03 1531 +560 263 + 396 594 + 528 64.3 + 5.8 71,5+4.8 1,47 + 0,51 2118+473 0,76+0,44 3,59 + 0,35 Other - grid squares mean + SD 6,47+ 1,94 9,i5 + 8,l4 1132 + 762 317 + 543 373 + 589 68.6 + 7,3 70.7 + 4.6 1.87 + 0,90 2217 + 867 0.70 + 0.46 3.48 + 0.33 z. -3,527 -2,391 -2,110 -0.340 -2.785 -2.417 -0.678 -1.641 -0.255 -0.554 -1.252 P < 0.001 0,018 0.036 0,685 0,005 0,016 0,497 0,101 0,799 0,580 0,211 ECOGRAPHY 24:5 (2001] Table 8. Predictive capacity of different variable families on their own for P. mnemosyne distribution and abundance when modelled together with the adjusting variables. Factor group Correctly classified (%) No variables added Food resources Habitat composition Management Habitat connectivity Climatology Habitat structure Topography 85.51 93.52 93.36 92.77 92.73 92.57 92.57 92.48 Distribution Number of sig. variables (p<0.01) Change of deviance ('14) Abundance Number of sia. variables (p<O.(H) 1 2': i1 1 2,- other reason for the modest fit of the tnodels may be the hick of historical data of the distribution and managetnciit of semi-natural grasslands in the area. The environmental variables affecting species distribution patterns can be difficull lo separate in a study with only a single area. The modelling of this study could be taken in a methodological sense one step further using predictive modelling, where the data are divided into a training set and into an independent test set (see Le Due et al. 1992). Hotspot characteristics Hotspots of P. mnemosyne differed statistically significantly (p<0.01) in only two environmental variables from other occupied grid squares (Table 7). namely in the amount of nectar planls and the cover of deciduous forest. Interestingly, neither factor became included in the multivariate abundatice model. Although the difference v^fas statistically weaker, the hotspot squares were positively related to the cover of setni-natural grassland (p < 0.036) and abundance of larval host plants ( p < 0.018). Surprisingly, habitat management or climatological variables did not show any statistical significance between hotspots and other grid squares. For example, the northernmost hotspot area was not managed, it is situated on north-east facing slope (low radiation), and it is obviously isolated from the main distribution area. Landscape ecology and metapopulation studies In this study, we chose to use landscape ecological approach instead of metapopulation approach because of several reasons. First, suitable habitat was impossible to define before the field study. It was known that the butterflies tend to occur on semi-natural grassland, but it was unclear whether the deciduous forests with high density of the larval host plant would be an essential part of suitable habitat. Secondly, the seemingly suit- able habitat did not occur in distinct patches, but instead formed a highly well connected and complex corridor-like system along the branches of the river Rekijoki (Fig. 1). Thirdly, within semi-natural grassland the occurrence of the larval host plant Corydalis solida varied from complete absence and scattered occurrence to continuous high density in large areas, which suggested that there might be large continuous variation in habitat quality among the suitable areas for reproduction (Fig. 3b). Fourthly, it seemed possible that other factors like abundance of nectar resources. habitat management by cattle grazing and topographical factors could substantially affect habitat quality. The results showed that potentially suitable habitat for the clouded apollo was determined by the occurrence of semi-natural grassland together with the larval host plant and adult nectar sources, but that there was large continuous and complex variation in habitat quality depending on several factors. An advantage of the use of spatial grid system was that the relative significance of difTerent components of habitat quality could be evaluated at the same time with the area of suitable habitat, habitat connectivity and spatial autocorrelation in butterfly occurrence. Even though suitable habitat in our study system did not occur in discrete habitat patches, the effects of the amount of suitable habitat and habitat connectivity were evident as in many metapopulation studies on butterflies (Hanski and Kuussaari 1995. Thomas and Hanski 1997) and other organisms (Hanski 1994. 1999). The area selected for this study represents the core of a wider clouded apollo occurrence in Ihe Somero. Hiintala region (Kuussaari and Luoto unpubl.). Although this system is highly well connected in its core area, there are more isolated areas of suitable habitat and butterfly occurrence at the edges. The population structure has features of "a patchy population" in the middle part of its distribution area and "a metapopulation" at the margins (see Harrison and Taylor 1997). The use of a spatial grid system helps to avoid difficulties of subjectivity in defining patches of suitable habitat when analysing patterns of distribution and tCOGR.APHY 24:5 (21)UI) abundance in this kind of population structure (e.g. Thomas and Kunin 1999, Tischendorf and Fahrig 2000). It has often been emphasized that the applicability of the metapopulation approach is restricted to cases in which the landscape can be meaningfully divided into suitable and unsuitable habitat (e,g. Hanski and Kuussaari 1995, Hanski and Simberloff 1997). Nevertheless, empirical studies on the factors affecting distribution and abundance at the landscape scale ( 5 100 km-^) of the many species with more complex habitat availability patterns and habitat requirements are rather scarce (Wiens 1997). A landscape ecological approach based on a spatial grid system may prove to be a widely applicable tool for studying these kinds of common situations. Acknowledgements - We thank Kari Haapaia, Julia Poyry and Heikki Tuominen for their help in the field in 1999, Roger L, H. Deniii.s, Risto Heikkinen. Marko Nieminen, Juha Poyry and Raimo Virkkala gave helpful comments on Ihc manuscript. We thank Antii Ponttinen for useful hints in spatial adjustment, Riitta Teiniianta and Tuuli Toivonen helped with several GIS problems, Jussi Kaukoranta produced the wind speed estimations in the Finnish Meteorological Institute, This study was financially supported by ihe Finnish Biodiversity Research Programme (FIBRE). The Finnish Ministry of the Environment and Forest and Park Service. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Ecography Wiley

Determinants of distribution and abundance in the clouded apollo butterfly: a landscape ecological approach

Download PDF
18 pages

Loading next page...
 
/lp/wiley/determinants-of-distribution-and-abundance-in-the-clouded-apollo-HbwXdZDJL5

References (82)

Publisher
Wiley
Copyright
Copyright © 2001 Wiley Subscription Services, Inc., A Wiley Company
ISSN
0906-7590
eISSN
1600-0587
DOI
10.1111/j.1600-0587.2001.tb00494.x
Publisher site
See Article on Publisher Site

Abstract

Luoto. M.. Kuussaari, M.. Riui. H.. Salininen. J. and von Bonsdorff. T. 2001. Deterniinant.s of distribution and abundance in the clouded apollo butterfly: a landscape ecological approach. - Ecography 24: 601-617. Recent studies on ihe determinants of distribution and abiindance of animals at landscape level have emphasized the usefulness of the metapopulalion approach, in which patch area and habitat connectivity have often proved to explain satisfactorily existing patch occupancy patterns. A dilTerent approach is needed to study ihe common situation in which suitable habitat is difficuU to determine or does not occur in well-defined habitat patches. We applied a landscape ecological approach to study the determinants of distribution and abundance of the threatened clouded apollo Parnassius mnemosyiie butterfly within an area of 6 km^ of agricultural landscape in south-western Finland. The relative role ol' 24 environmental variable.s potentially atTecting the distribution and abundance of the butterfly was studied using a spatial grid system with 2408 grid squares of Q.25 ha. of which 349 were occupied by the clouded apollo. Both the probability of butterfly presence and abundance in a 0.25 ha square increased with the presence of the larval host plant Corydalis solida, the cover of semi-natural grassland, the amount of solar radiation and spalial autocorrelation in butterfly occurrence. Additionally, butterfly abundance increased with overall mean patch size and decreased with maximum slope angle and wind speed. Two advantages of the employment of a spalial grid system included the avoidance of a subjective definition of suitable habitat patches and an evaluation of ihe relative signiflcance of dilTeren! components of habitat quality at the same time with habitat availability and connectivity. The large variation in habitat quality was influenced by the abundance of the larval host plant and adult nectar sources but also by cliniatological. topographical and structural factors. The application of a spatial grid system as used here has potential for a wide use in studies on landscape-level distribution and abundance patterns in species with complex habitat requirements and habitat availability patterns. M. Luoio (miskaJuotoQiiwh.fi). Finnish Environment Inst., CIS and Remote Sensing Unit, P.O. Box 140 (Kesdkatu 6). FIN-0Q25! Hehinki, Finland. - M. Ktmssari and J. Salminen, Finnish Environment Inst., Mature and Land Use Dir.. P.O. Box 140 (Ke.sakatu 6). FfN-00251 Helsinki. Finland. - H- Rita. Dept of Ecohgv and Sy.stematics. Div. of Population Biology. P.O Box 17 (Arkadiankatu 7). FIN-00014'Vmi. of Helsinki. Finland. - T. von Bon.sdorff. The Finnish Mu.wum of Natural History. Div. of Mycology. P.O. Box 7 (Hameentie 153). FIN-00014 Univ.ofHehinki. Finland. Which factors explain the distribution and the abundance of a species has continued to be a central question in ecological research since the pioneering work of Andrewartha and Birch (1954). The development of spatial ecology during the last ten years has shifted the focus of the question from local and geographic scales Accepted 16 January 2001 Copyright © ECOGRAPHY 2001 ISSN 0906-7590 Printed in Ireland ~ all rights reserved ECOGRAPHY 24:5 (20011 to the intermediate, "'landscape" scale (Gilpin and Hanski 1991. Forman 1995, Hanski and Gilpin 1997, Tiltnan and Kareiva 1997). Empirical studies on distribution and abundance of species al landscape (or regional) scale can be categorized into two kinds of approaches (Wiens 1997, Hanski 1999). metapopulation approach in which the landscape is divided into two types of habitat, suitable patches and unsuitable matrix, and landscape ecological approach in which the complexity of real landscapes is more explicitly considered. Metapopuiation approaeh has been especially popular among butterfly ecologists (Thomas and Hanski 1997). Empirical butterfly studies have shown that for speeies which typically occur in relatively discrete habitat patches, existing patch occupancy patterns can be satisfactorily predicted by using only two variables, patch area and habitat connectivity (e.g. Thomas and Harrison 1992. Thomas et al. 1992, Hanski et al. 1995, Wahlberg et al. 1996; for an exception, see Dennis and Eales 1997, 1999). In the case of the Glanville fritillary Melitaea cinxia which occurs in networks of welldeflned patches of dry meadows in the Aland islands. Moilanen and Hanski {199f^) showed that the accuracy of the model predictions increased only slightly when the variation in loeal habitat quality and landscape structure were taken into account. For many species suitable habitat occurs in less distinct habitat patches and defining suitable habitat is difficulf because of continuous variation in habitat quality. In sueh cases metapopuiation approach is less useful. Thomas and Kunin {1999) suggested the use of spatial grid system in studying the population structure of sueh species. At the geographie scale, grid based distribution maps and environmental variables have been used to distinguish faetors affecting speeies range limits {Turner et al. 1987, Hill et al. 1999). Grid systems have also been used to prediet occurrences and to estimate declines at smaller geographic scales (e.g. Thomas and Abery 1995, Kunin 1998, Dennis and Hardy 1999. Cowley et al. 1999. Dennis et al. 1999, Leon-Cortes et al. 1999). A fine-scale spatial grid system has been commonly used to study butterfly movements and population structure at a loeal scale of one or a few closely located habitat patches (e.g. Singer 1972. Arnold 1983, Gall 1984. Vaisanen et al. 1994). Sometimes environmental variables measured from the grids have also been used to explain butterfly abundance {e.g. Thomas 1985. Vaisanen et al. 1994. Fiseher et al. 1999) and patterns of adult phenology (Weiss and Weiss 1998). Examples of extending the fine-scale grid based approach to the landscape level in studies of butterflies seem to be largely lacking {see exceptions in Ravenscroft 1994, Thomas et al. 1999). One advantage of the use of a spatial grid system is the possibility to utilize GIS data as a source of potentially important explanatory variables. Environmental variables like the oecurrence of various kinds of habitats and vegetation types as well as the patterns of variation in topography and climate are becoming inereasingly available from aerial photographs and satellite images via rapidly developing GIS techniques. We used a landscape ecological approach to study the factors affeeting distribution and abundance of the threatened clouded apollo butterfly Panuissius mnemosyne L.; Papilionidae in a system of 50 x 50 m^ grid squares covering 6 km^ of agricultural landscape in south-western Finland. Our aim was to evaluate the relative role of several lypes of potentially important environmental variables for the occurrence and the abundance of the butterfly. A total of 24 environmental variables were flrst grouped into seven variable families (food resources, habitat composition, habitat management, habitat structure, topography, climatology and habitat conneetivity), whieh were then used in a stepwise manner by adding one family at a time to construct multivariate models to explain butterfly distribution and abundance. Because of the high number of environmental variables, we put special emphasis on the logic of model building. The variable families, whieh were known to be biologically necessary (food resources and habitat composition), were brought into the model first, whereas the other fatnilies were entered into the model in the order of iheir statistical significance. Material and methods Clouded apollo Pai-nassius mnemosyne The clouded apollo is a threatened species in Finland as well as in most areas where it presently occurs in western and northern Europe (Somerma 1997, Aagaard et al. 1997, Kotiranta et al. 1998, Konvicka and Kuras 1999). The decline of the clouded apollo has been attributed to the ceasing of traditional management regimes, grazing and mowing, of semi-natural grasslands and coppiced woodlands {Vaisanen and Somerma 1985, Kotiranta et al. 1998). Several European studies have focused on the spatial and genetic population structure (Aagaard el al. 1997, Meglecz et al. 1997. 1999. Konvicka and Kuras 1999) as well as on genetic and morphometric differentiation between geographically separated populations (Vaisanen et al. 1991. Deseimon and Napolitano 1993), but detailed environmental requirements of the butterfly and its preimaginal stages have remained relatively poorly known. At any one location the larva of the clouded apollo seems to be monophagous on one of the species in the genus Corvdaiis {Fumariaeeae). but the particular species used varies geographically: C. solida in Finland (Somerma 1997 and our own observations), C intermedia in Norway (Aagaard and Hanssen 1989), C .solida and C. cava in the Czech Republic {Konvicka and Kuras 1999) and in Hungary (Meglecz et al. 1999). Females lay 1-3 eggs at a time in the ground vegetation when the host plants have already senescenced {Konvicka and Kuras 1999). ECOGRAPHY li-.i (2001) Study area ' ' The study area, ca 6 km" is situated in Somero, Hantala in south-western Finland. The upper course of the river Rekijoki flows in the study area across broad clay plains (Fig. t). Large parts of the study area are eharacterized by broad and flat expanses of intensively cultivated clay soils. Near the river Rekijoki, V-shaped valleys with steep meadow slopes dominate the landscape. River valleys are characterized by highly diverse types of land use and large semi-natural meadows and pastures, which are remnants from a long period of less intensive, traditional agriculture. At present Rekijoki is the largest area of mesic grasslands that currently exists in Finland (Kontula et al. 2000). A detailed description of the area is given by Luoto (2000) and IContula et al. (2000). To summarize the results of the survey we produced two variables measuring the distribution and abundance of the butterfly. Distribution was measured as a binary variable, presence ( > 0 observations) or absence (0 observations) of butterflies in each grid square during the whole flight season. Abundance was measured as the total number of butterflies observed during all visits divided by the number of visits in each grid square. Environmental variables Seven groups of environmental variables were recorded for each 0.25 ha grid square (n = 2408): 1) food resources. 2) habitat composition, 3) habitat management, 4) habitat structure. 5) topography, 6) climatology and 7) habitat connectivity. The measurement of the environmental variables and species distribution and abundance is dependent on the scale on which the measurements are made (Turner 1989. Wiens 1989, Forman 1995). Therefore a spatial grid system with a constant grid size was chosen to reduce errors of scale dependence of explanatory and response variables (see Turner 1989). Food resources Food resources of the larvae and adult butterflies were mapped in the spring and summer 1999, respectively. Average density of larval host plant (Corydalis .\olida) shoots m " - was estimated from each grid square in the field between 22 April and 20 May. In distribution modelling we also used a binomial variable of Corvdalls (0 ^ absence, 1 = presence). The abundance of the nectar plants (Amhriseus syhestris. Geranium sylvalicum. TrifoUum spp., Vicia eraeea, V. sepium. Ranunculus spp., Cainpamila patula) used by P. mnemosyne (our own observations) was recorded from 567 grid squares of semi-natural grassland (70'>^i of the study squares with any grassland within the square were .surveyed). The abundance of nectar plants was calculated using a scale from 0 to 10 (0 = no nectar plants, 10 ^nectar plants extremely abundant) between 15 and 23 June. Due to the hmited spatial coverage of nectar plant data it was only used in the analyses of the abundance of P. mnemosyne. Habitat composilkm The occurrence of different kinds of habitats in the study area (Fig. 1) was deflned from aerial low altitude blaek and white photographs (1:7500 scale) and from topographic maps. A total of 8 different land cover types were identified and assigned. The habitat patches were digitized and the derived habitat map was verifled in the fleld. Habitat composition was measured using GIS program Arc/Tnfo. vers. 7.2.1. The amount of different habitats for each 0.25 ha grid square was calculated in m^ from polygon coverages by a frequency function (Anon. I991J. 603 Butterfly survey In summer 1999 we surveyed the oeeurrence and the abundance of the clouded apollo in a total of 2408 grid squares, 50 x 50 m in size, within a 6 km- study area. We also used mark-release-recapture method to collect data on butterfly movements. Results on the movements will be reported in a separate paper. Butterflies were marked with an individual number and recaptured daily, weather permitting, between 6 June and 1 July. The location of eaeh eapture was marked on to a copy of a high-resolution aerial photograph (scale 1:3750). We divided the study area into eight sub-areas. One of the sub-areas was visited daily and the other seven sub-areas were visited Hve to eight times during the adult flight season. Fig. 1. A map on the occurrence of different kinds of habitats in tbe study area. Due to the cartographic reasons forest habitats are reclassified as one habitat class. ECOGRAPHY 24:5 (2001) Habitat structure Measures of habitat structure were calcuhtted for each 0.25 ha square with a 50 m buffer zone surrounding the square usitig the geostatistical progratn Fragstats, ver. 2.0 (McGarical and Marks 1994). The cell size used in Fragstats was 1 x 1 m, and the core area distance and border zone were defined as 0 m. Four measures of habitat structure were produced based on the habitat map in Fig. 1: mean patch size. Shannon's diversity index, mean shape index and area weighted mean shape index (for formulas see McGarical and Marks 1994). Topography Six topographical variables were calculated from each grid square: 1) mean altitude. 2) the lowest point. 3) the highest point, 4) relative altitude (highest - lowest point), 5) mean slope angle and 6) maximum slope angle. Variables were calculated using Arc/Info Grid from a digital elevation model (Anon. 1997). CHimitology An estimate of maximum theoretical solar radiation for each grid square was produced using a computer model of clear sky insolation and exposure of different slopes. In the following calculation, variation in thermal conditions across the study area is a direct result of solar radiation on different slope exposures (Griffiths 1985): Solar radiation = [cos(altitude) x sin(slope angle) xeos(ground direction — sun direction) + sin(altitude) xcos(stope angle)] Habitat connectivity Habitat connectivity was trteasured only for one habitat lype, semi-natural grassland, which was expected to be the single most important habitat type for the clouded apollo. Connectivity (S) to semi-natural grasslands in the neighbourhood of grid square i was measured following Hanski (1994): where A| is the area of semi-natural grassland (in ha) in square, and d^j is the distance between squares i and j (in km). For each study square, connectivity to seminatural grasslands in all other study squares was measttred up till two km distance. Two km was set as the upper limit, because for longer distances no accurate habitat maps were available. The exponential function gives more weight for those squares, which are close to the focal square. Value 1 was used for paratneter ex, and it was selected based on mark-recaptut-e data on the movements of the clouded apollo (Kuussaari and Luoto ttnpubl.). Adjusting variables Spatial autocorrelation The probability of occurrence of P. mnemosyne in one grid square is unlikely to be independent of whether the species occurs in a neighbouring square. This will generate spatial autocorrelation that cannot be modelled satisfactorily by environmental covariates (Augustin et al. 1996). In this study, we attempted to minimize the effect of spatial autocorrelation of data by using two adjusting spatial autocorrelation variables. For the analysis on the factors affecting the presence and absence of butterflies the adjusting variable was calculated as the number of occupied squares among the eight neighbouring squares of each sttidy square. For the analysis on butterfly abundances the adjusting variable was calculated as the average number of captures divided by the number of visits among the neighbours for each grid square. In the case of the grid sqttares at the borders of the study area we calculated the average of available neighbouring squares. Study effort The probability of observing P. mnemosyne in a 0.25 ha grid square is affected by the nutnber of visits, i.e. study effort (Dennis et al. 1999). Therefore the study area was divided into two parts differing in their sampling effort: "the core sub-area" and "the other sub-areas'". The cover of the core sub-area was 375 grid squares (16%) and cover of other areas was 2033 grid squares (84'V(i). Sampling effort (the total amount of visits in an average grid square) was ca 15 times in the core sub-area ECOGRAPHY 24:S |2(Kll) (1) In this calculation altitude is the angle of sun above horizon. Sun direction is the direction of the sun from the south. It increases positively to west and negatively to east (south ^ 0 . east = - 9 0 . west = +90). Slope angle is slope angle of the ground (flat area = 0. vertical = 90). Ground direction is the downgrade direction of the ground from the south. It increases positively to west and negatively to cast (south = 0. east = - 90, west =^ + 90). Solar radiation was calculated as the average of three moments during clouded apollo's daily activity period: 9.00, 12.00 and 15.00. The average wind speed for each grid square during the butterfly's daily activity period (8.00-17.00) was produced by the Finnish Meteorological Institute based on a field survey and topographical and habitat maps. Estimation of the average wind speed in June was done using a four-class scale: 1) <0.5 m s " '; 2) 0.5- 1.5 m s " ' ; 3) 1.5-2.3 m s " ' : and 4) > 2.3 m s " ' . Management Data on eattle grazing, the primary management method of the semi-natural grasslands in the study area were collected during the fieldwork in summer 1999. In the analyses, we used cattle grazing as a binary variable (0 = no grazing. 1 = grazitig) for each grid square. 604 Adjusting i HHbltal com;osltior htebitat connectivity TopogiafSiy StructuG CNmatQtogy Management Bemainlng fcur groups * — Ramalning three groups nainire two groups Remalnlrg one gmup Fig, 2, The logic of model building in explaining the patterns of distribution (logistic regression) and abundance (GLM regression) of the clouded apollo. Adjusting variables were entered into the models first, food resources secondly and habitat composition variables at the third step. Next the most predictive of the remaining five variable families was added; topography, habitat structure, climatology, habitat connectivity and management. This procedure was repeated, and the next predictive variable family added until no more statistically significant (p<O.Ol) variables remained, tial autocorrelation as an adjusting variable and the other without spatial autocorrelation. Model building for both distribution and abundance was started by entering the adjusting variables, study effort and spatial autocorrelation (see summary of the logic of model building in Fig. 2), Adjusting variables were kept in the model regardless of their statistical significance. Next, variables according to their biological importance were selected for inclusion. Thereafter, the remaining variables were entered in the order of their statistical significance. The statistical significance after including (or excltiding) a variable in a model was calculated. This procedure was repeated until no more important variables remained. At eaeh step, variables entered in previous steps were tested for redundancy (see Yee and Mitchell 1991. Tonteri 1994). Distribution modelling First, all explanatory variables were described by comparing descriptive statistics between the grids in which P. nmemosyne was present versus absent. We used univariate analysis to assess individual model variables independently from each other and to obtain information on each variables' role before proceeding with model development and testing, a procedure that is also recommended by Pereira and Itami (1991). Next, multiple logistic regression, a form of generalized linear modelling (Rita and Ranta 1993. Trexler and Travis 1993. Sokal and Rohlf 1995) was used to analyse the relationship between environmental variables and distribution of P. mnemosxne in 2408 grid squares (see Brito et al. 1999, Carroll et al. 1999), using the SPSS-PC 7.0 software paekage. and ca 6 times in the other sub-areas during the adult flight season. In the analyses on butterfly distribution and abundance the variable study effort was taken into account as a binary variable: in the core subarea effort = 1. and in the other sub-areas effort = 0, Statistical analysis Logic of model building Recent papers have criticized automatic stepwise procedures, as they are not able to select the most influential from a set of variables (James and McCulloch 1990, Bustamante 1997, MaeNally 2000). Therefore, to compile models for distribution and abundance of P. mnemosyne, the explanatory variable groups were selected to the model in a priori defmed manner taking into account the existing knowledge on the biology of the focal species (Fig. 2). We used a combination of forward selection and backwise elimination procedures with a strict \% significance level as the criterion. Food resources, the most primary need of P. nmemosyne. were ranked first, and habitat composition secondly due to the biological importance of the type of habitat. Habitat structure, habitat management, topography, climatology and habitat connectivity variables were ranked equally important and they were entered into the multivariate models in the order of their statistical significance. Two kinds of models were constructed to explain butterfiy distribution and abundance: one using spaECOGRAPHY 24:.^ (200!) Abundance modelling The abundance of clouded apollo (captures/visits) in each grid square was related to the environmental factors using generalized linear models (GLM) (Austin et al. 1984, McCulIagh and Nelder 1989, Nichoils 1991. Crawley 1993), as implemented in the statistical program GLIM 3,77 (Payne 1986), All survey grid squares where P. mncniosvne was found to be present in summer 1999 were included in the analyses. In addition, also those grid squares were chosen to the analysis where the logistic regression model estimated the probability of P. nmemosyne presence over 0.5. With this selection procedure the abundance analyses were focused on the observed distribution area and also on those grid squares where the environmental factors seemed to be suitable for occurrence. The total amount of grid squares in the analyses was 412. A Poisson distribution of errors was assumed and abundance of P. nmemosyne was 605 related to explanatory variables via a logarithmic link function (Mclntyre and Lavorel 1994). First, all the 24 explanatory variables were related separately to the response variable, using data from the 412 grid squares. The statistical significance of the change in deviance after including (or excluding) an explanatory variable in the model was determined by an F-ratio test (McCullagh and Nelder 1989, Crawley 1993) with a 1% significance level as the criterion. Further, we tested also the non-linear relationship between explanatory variables and abundance by including the quadratic terms of continuous variables into the model (see Margules et al. 1987, Austin et al. 1996). The derived models were evaluated by plotting residuals against the fitted values and the explanatory variables, and by plotting residuals in a normal probability plot as suggested by Nicholls (1991). tree saplings, which were situated adjacent to seminatural grasslands, had low or moderate incidence. The eastern part of the study area formed an exception to the distribution oi P. mnemosyne: regardless of relatively large cover of semi-natural grassland and presence of Corydalis solida no P. mnemosyne were found (see Fig. 3a, b). Distribution Comparison of descriptive statistics of various environmental variables between occupied and unoccupied grid squares showed that there were large statistical differences between those two groups of grid squares (Table I). Occupied squares were located in areas where larval host plant was abundant, the covers of semi-natural grassland and deciduous forest were higher and habitat structure was more heterogeneous than in unoccupied grid squares. In addition, occupied grid squares were characterized by steep slopes and low wind speed. Before the entry of explanatory variables the overall correct classification rate was 85.51%, because in 2059 of 2408 grid squares P. mnemosyne was absent (Table 2). The distribution modelling was started by fitting the "adjusting variables" in the model (spatial autocorrelation. p<O.OOI and effort p = 0.50). Together these variables predicted 207 of 349 (59.3%) P. mnemosyne grid squares correctly and the overall correct classification rate was 91.7%. Next the binomial Corydalis variable ( p < 0.001) was fitted of the food resources group. After this, model building was continued by fitting the best explanatory variable of the habitat composition group, namely the cover of seminatural grassland (p<0.()01). No other variables of habitat composition group could be fitted to the model. After this inclusion the overall correct classification rate was 93.9%. In the preceding steps only radiation (p < 0.002) of the climatological variable group could be fitted to the model. Finally, the overall classification of predicted and observed grid squares was 93.8 and 75.4'^!) of the presence grid squares were correctly classified. Coefficients and standard errors are given in Table 3. There was a statistically significant ( p < 0.001) spatial autocorrelation in the presence grid squares. When the distribution modelling was done without the spatial autocorrelation variable also habitat connectivity and minimum altitude could be fitted to the model in addition to the previous three variables (see Tables 2 and 3). In this case the overall correct classification rate was 92.7%. Predicted probability of occurrence of P. mnemosyne calculated by logistic regression with spatial autocorrelation are shown in Fig. 3d. and without spatial autocorrelation in Fig. 3e. tICOGRAPHY 24:5 (2001) Characteristics of hotspots To define the characteristics of high abundance grid squares, descriptive statistics for selected environmental variables of hotspot and other populated grid squares were calculated and statistically tested using Mann-Whitney U-test. Following Prendergast et al. (1993) we defined hotspots as the top 5% of the grid squares, ranked by the abundance of the butterfiy (captures/visits). Predictive capacity of different types of environmental variables To determine the predictive capacity of different environrnental variable families in estimating distribution and abundance of P. mnemosyne we related the variable families separately to the response variable using logistic and GLM regression modelling. The overall correct classification rate (distribution model), change of deviance (abundance model) and the number of significant variables were calculated for all factor groups. The predictive capacity of each factor group was tested with adjusting variables included. Results A total of 349 (14.5%) of 0.25 ha study squares were found to contain P. mnemosyne. The distribution of P. mnemosyne was clearly associated with the river valleys having extensive cover of semi-natural grasslands (Fig. 3a). high density of Corydalis .solida (Fig. 3b) and high amount of solar radiation (Fig. 3c). Butterflies were absent or showed very low incidence in agricultural fields, built-up areas and coniferous forests (Table I). Mixed and deciduous forests, fellings and 606 ative) and three topography variables, showing often a quadratic relationship between abundance. Of the adIn the squares occupied by the butterfly abundance justing variables spatial autocorrelation was signifivaried between 0.05 and 9.50 captures/visit (mean = 1.05, SD = 1.28; Fig. 3). In 104 of 349 occupied squares cantly related to the abundance of P. mnemosyne. The construction of the multivariate GLM model to (29.8%) the abundance was over one capture/visit. When the effect of environmental variables was tested e.xplain butterfly abundance is shown step by step in singly, 16 of the 22 explanatory variables were signifi- Table 5. The set of variables, which together best cantly (p<0.01) related to the abundance of P. explained the abundance of P. mnemosyne in the multimnemosyne (Table 4). The variables accounting for the variate GLM model with adjusting variables, indicates largest change in deviance were the abundance of nec- that the abundance increased with the amount of Corytar plants (positive effect), the cover of semi-natural dalis., cover of semi-natural grassland, radiation and grassland (positive), the cover of agricultural area (neg- mean patch size, and decreased with maximum slope Abundance li-tWspotgrtdsquare Chxidai apotlo abundance f • 0 (capOjtes/viail) C _ 0.05.0.99 1.00-1,99 2.00-9.50 Fig. 3. (a) Observed abundance (captures/visits) of clouded apollo Parnassius mnemosyne in the 50 x 50 m grid squares and the occurrence of semi-natural grasslands. The hotspots (the top 5% of the squares with the highest butterfly abundance) are marked by a star symbol (b) Densily (shoots m " -) of larval host plant CorydalLs solida. and (c) solar radiation in the study area. Predicted distribution of P. mnemosvne according to the logistic regression models presented in Table 3: (d) with spatial autocorrelation, and (e) without spatial autocorrelation. PredWai pt«aiflity o( occun»x» (viiih spatial aita»(Te4aikm O00(W)20 rrt]021050 aHO |BO Profcled prtiabily rt occunwce (iwthixit spaHal aulDcoffBlHtion) I'lo.21-11.50 wio.5i-o.ao •io.ei-1.00 ECOGRAPHV 24:5 (21)01) Table I. A list of environmental variables calculated for eaeli of the 50x50 m grid squares in the study area and used as explanatory variables of the distribution and abundance of P. itmemo.syne butterfly. Descriptive staiisties (mean and SD) for ihe two types of grid squares in the study area; I) P. mnemosyne present (n = 359), and 2) P. mnemosyne absent (n = 2059), P-vatues were derived from Mann-Wliitney U-test. Variable families and variables P. mnemo.nyne absent mean + SD 2.21 +5.56 6.46 ± 1.84 1380+1079 150 + 397 101 +372 169 + 506 404 + 817 102 + 403 12.0 + 55 178+470 0.27 ± 0.45 1.47 0.50 1.63 0.76 + + + + 0.25 0.46 0.28 0.45 P. mnemosyne present mean ± SD 9.35 + 8.18 6.55 ±1.94 342 + 656 -25.618 -0.456 -15.948 -28.121 -18.444 -10.838 -2.078 -1.213 - 16.754 -7.259 -21.438 -19.800 -18.477 -6.118 -16.858 - 24.228 -23..Ml -17.678 -24.258 -21.188 -21.879 -4.716 - 10.583 -18.544 0.000 0.585 0.000 0.000 0.000 0.000 0.038 0.225 0.000 0.000 0.000 0.000 0.000 0,000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 Food resources Corydalis sollda (host plant of larvae, shoots m~-) Nectar plants (only squares with semi-iiiit. grassl., 10-class scale) Habitat compo.sition (m') Agricultural area in the grid square Semi-natural grassland in the grid square Deciduous forest In tlie grid square Mixed forest in (lie grid square ConiferoLLS forest in the grid square Felling or tree sapling in the grid square River in the grid square Built area in the grid square Management (only squares witli semi-nat. grassl.) Management of semi-natural grassland (()/!) Habitat structure Area weighted mean shape index Mean pateh size (ha) Mean shape index Shannon's diversity index Topography Altitude min (m a.s.l.) Altitude ma.\ (m a.s,l,) Altitude relative (m a.s.l.) Altitude mean (m a.s.l.) Slope angle mean (°) Slope angle max (") Climatology Radiation (0-100) Wind speed (4-dass scale) Habitat connectivity Habitat conneetivity of semi-natuial grassland 1151+758 373 + 588 314 + 536 142 + 392 79.9 + 358 67.9+ 116 29.7 ± 180 0.47 ±0.50 1.78+0.22 0.22 + 0.09 1.70 + 0.14 1.20 ±0.28 66.0 + 71.7 + 5.67 + 68.4 + 8.07 + 7.60 6.84 3.41 7.27 3.61 80.5 + 8.09 82.6 + 6.90 2.18-K 2.76 81.3+ 7.41 2.91+3.50 4.05 ± 4.54 69.9 + 2.83 2.58+1.18 2.99 + 0.58 11.0 ±3.96 70.8 + 4.60 1.85 ±0.89 3-59 + 0.33 angle and wind speed (Table 6). In addition, the quadratic function of mixed forest showed also statistical significance in the abundance model. The flnal models of abundance of P. mnemosyne with coefficients and standard errors are given in Table 6. The model explained 35.8'>ii of the variation in abundance of P. mnemosyne (total deviance change 182.1 out of 509.1). There was a statistically significant (p < 0.001) spatial autocorrelation in the abundance of P. mnemosyne grid squares. When the abundance modelling was done without the spatial autocorrelation variable the number of explanatory variables increased remarkably. Wind speed, slope maximum and mean patch size, which were included in the autocorrelation model, got replaced by quadratic term of deciduous forest, mean altitude, quadratic term of mean slope angle and quadratic tertn of Shannon's diversity index (Tables 5 and 6). distribution pattern in the study area (hotspots are indicated by a star sytnbol in Fig. 3). Mann-Whitney U-test showed statistical difference in several environmental variables between hotspot and other grid squares (Table 7), Hotspots were characterized by high amount of nectar plants and larval host plant. Further characteristics of hotspot grid squares included large cover of semi-natural grassland and deciduous forest, and low mean altitude. Predictive capacity of different environmental variables Predictive capacity of different variable families for distribution and abundance of P. mnemosyne showed clear differences (Table 8). Food resources (overall correct classification 93.52'Mi) and habitat composition (93.36%.) groups were the single most powerful explanatory variable groups for distribution, whereas habitat structure and topography were the poorest variable families in the predictive power. The most powerful groups of abundance modelling were habitat composiECOGRAPHY 24:^ (Mil) Characteristics of hotspots The hotspoEs (the top 5"A, of the squares with the highest butterfly abundance) had a relatively dispersed tion (33.35% change of deviance), topography (25,39%) and food resources (23.82%). The number of significant variables varied between 0 (management and habitat connectivity in the abundance model) and 4 (habitat composition in the abundance model) within the variable families. determinants of distribution and abundance. Additional factors contributing to the occurrence of butterflies included variables associated with topography, climatology, habitat structure, habitat connectivity and spatial autocorrelation in butterfly occurrence. Factors affecting distribution Discussion Employment of a spatial grid system proved to be a useful way to estimate simultaneously the role of many potentially significant environmental variables in determining the distribution and abundance of the clouded apollo. Availability of semi-natural grassland together with the necessary food resources for the larvae and adult butterflies turned out to be the most significant Parnassius nmemosyne butterflies occur in senii-natura! grasslands on the slopes of the river valleys. The set of environmental factors, which together best explained the presence and absence of/', mneniosyne in a multiple logistic regression, indicates that the probability of P. nmemosyne occurring in a square increases with the presence of the larval host plant Corydalis solida, cover of semi-natural grassland and solar radiation. In the multiple logistic regression model there was no statisti- Table 2. Summary of the logistic regression mode! building procedures (forward selection and backwise deletion of explanatory variables) for distribution of P. mnemosyne. Both forward selection and backwise deletion of variables were employed in the model building, using a strict criterion (p<0,01) for their inclusion or exclusion. The model is built using P. nmemosyne distribution (presence/absence) and environmental data from the 2408 grid squares in the study area. Model building procedures with and without a spatial autocorrelation variable are presented. Model with spatial autocorrelation Model No terms added I (adjusting variables) + Spatial autocorrelation + Effort II (Food resources) + Corydalis (0/1) (larval host plant) III (Habitat composition) + Semi-natural grassland IV (Climatology) 4- Radiation * * p < 0 , 0 1 : ***p<O.OOI, Model without spatial autocorrelation Model No terms added I (adjusting variables) + Effort II (Food resources) + Corydalis (0/1) (larval host plant) III (Habitat composition) + Semi-natural grassland + Deciduous forest + Mixed forest IV (Habitat connectivity) + Habitat connectivity — Deciduous forest — Mixed forest VI (Topography) -1- Altitude minimum VII (Climatology) -1- Radiation B SE % of total correct classificat. 85,51 0,8469 4.7183 0.1546 0.0487 0.0460 2.1664 0,0251 0,0334 -0,0897 0,0815 0.1367 0.3106 0.0115 0.0129 0.0141 0.2325 0.0147 0.0151 0.0130 0.0200 85,51 87,79 91,20 91,28 91.57 92,32 92,28 92,23 92.36 92,73 0,1351 0.3389 0,2990 0,0787 0,0657 0,2063 0,0214 0,0380 -0,1519 0,0858 *** ++* SE % of total correct classificat. 85,51 1.0077 -0,0081 2,7846 0,1091 0,0645 0.0443 0.2347 0.3402 0,0143 0,0209 91,65 91,65 93.52 93.94 93.81 0,5088 0,0000 0,1806 0,1675 0,0615 *** n,s. •** •** • ** *** *++ ** *** n.s. 11.S. *** ++* p<0.01; ***p<0,001. ECOGRAPHY 24:5 (2001) Table 3. Variables in final multiple logistic regression models (see Table 2) for the P. mnemosyne distribution showing coeffieients, standard errors and significance values. Models with and without a spatial autocorrelation variable are presented. Model wilh spatial autocorrelation Model Constant Spatial autocorrelation Effort Corydali.s (O.'l) (larval host plant) Semi-natural grassland Radiation **p<0.0]; ***p<O.0()l. Model without spatial autocorrelation Model Constant Effort Corvdalis (0/1) (larval host plant) Semi-natural grassland Habitat connectivity Altitude minimum Radiation **p<0.01; ***p<0.001. B -12.168 1.487 2.944 0.115 2.172 -0.102 0.105 SE 1.871 0.260 0.354 0.014 0.242 0.013 0.020 Exp(B( 4.424 18.999 1.122 1.122 0.903 I.Ill R 0.124 0.184 0.188 0.188 -0.166 0.110 P ** **+ *** *** **« *** B -10.528 6.8K4 0.851 2^71 0.101 0.065 SE 1.522 0.049 0.278 0.377 0.0! 5 0.02 i Exp(B) 2.341 1.990 13.080 1.106 l.0(S7 R 0.313 0.061 0.150 0,151 0.0615 P *** •*• ++* *** **• ** cal significance between habitat connectivity, habitat management, habitat structure and topographical variables, and distribution pattern of P. mnemosyne. It should be noted, however, that our measures of spatial autocorrelation and connectivity to surrounding seminalural grasslands were strongly correlated (Pearson correlation coefficient. R ^ 0.48. p<().01). When the model was constructed without spatial autocorrelation as an adjusting variable, the effect of habitat connectivity on the grid square occupancy became highly significant (Table 3). Due to the limited spatial coverage of nectar pianl data no analyses on the role of nectar plants for the distribution of P. mnemosyne were performed. The univariate analysis revealed clear differences in topographical and habitat structural variables between squares wilh and without P. mnemosyne (Table 1). However, these variables were of minor importance after the biologically primary variables (food resources and habitat composition) had been included to the distribution model (Table 3). The sequence of selecting variables used here probably obscured the fact that some variation in distribution could have been explained by food resource and habitat composition variables as well as by topographical and habitat structural variables. Topographical and structural variables are intercorrelated with food resources and habitat composition variables, because semi-natural grasslatids are eoncentrated in steep river valleys with low minimum altitude and small mean pateh size (see MacNally 2000). The clouded apollo was absent in the eastern part of the study area and showed very low and scattered incidence in the .study area at the north-eastern edge (Fig. 3). Those areas had high or moderate density of larval host plant and solar radiation and relatively large 610 cover of semi-natural grassland. They differed from the clouded apollo distribution area mainly by more fragmented pattern in the occurrence of semi-natural grassland (i.e. low connectivity) (Figs I and 3). The predicted probability of oeeurrence in both logistic regression models gradually decreased towards north-east (Fig. 3d-e). Factors affecting abundance The GLM modelling procedure and analysis of predictive capacity of different factor groups showed that habitat composition and food resources variables were the most important factors explaining the abundance of P. mnemosyne. The abundance increased with cover of semi-natural grasslands and density of Corydalis. After quadratic function of semi-natural grasslands was fitted in the model, the abundance of nectar resources was no longer significant in the model. This suggests that the density of nectar plants was generally sufficient for adult butterflies in the study area. Nevertheless, the density of P. mnemosyne was especially high in the grid squares with high atnount of nectar plants (Table 7). Mixed forest patches, whieh were included as a statistically significant variable in the model for abundance, have certain eeologieal features for abundance. It seems that small forest patches in the semi-natural grasslands did not decrease butterfly abundance - on the contrary they increased it. This can be an implication of high larval host plant density and possibly high larval densily in those areas due to high oviposition rate or high survival of larvae. ECOCiRAPllY 24:.i a Climatological conditions infiuence activity patterns and growth of insects in all life stages (Taylor 1981. Turner et al. 1987, Weiss et al. 1988, 1993, Weiss and Weiss 1998). For the clouded apollo climatological variables, like wind speed and solar radiation, were important predictors of abundance. Wind speed showed a negative relationship with butterfly abundance, whereas the correlation of butterfly abundance with solar radiation was positive. As reported by Dover et al. (1997), high wind speed hinders fiying of butterflies and sheltered areas seem to be very important for them in open agricultural landscapes. Variation in thermal conditions across the study area is a direct result of solar radiation on different slope exposures (Griffiths 1985). South-facing slopes with high solar radiation have certain special ecological features for the clouded apollo, which occurs in the study area in the northern margin of its European distribution. In spring the lar- vae need to complete their development before their host plants senesce. Slopes with high radiation are the warmest on average and there is less evaporation caused by wind (Mansikkaniemi and Laitinen 1990). This suggests that developmental conditions for larvae tnay be more favourable on sheltered south-facing slopes than in other areas. In addition, when the amount of radiation increases, the potential duration of daily activity period and potential time for reproduction (e.g. number of eggs) of adult butterflies increases. Many studies have indicated the importance of habitat and topographical heterogeneity in explaining species' spatial patterns (Weiss et al. 1988. 1993, Verboom et al. 1991, Gustafson and Gardner 1996, Moilanen and Hanski 1998). In this study, mean patch size showed a positive relationship with the abundance. This can be explained by the fact that the abundance of P. mncmosyne was related highly and positively to the Table 4. The change in the deviance from a model f"or abundance of P. mnemosyne wilh no terms to that containing each environmental variable on ils own, A Poisson error distribution for the abundance of /'. mnemosyne is assumed and the response variable is linked to the set predictor variables via a logarithmic function. Only those linear and quadratic functions of Ihe variitbles that account for a statistically significant (_p<0.01) change in ihe deviance are shown, except linear non-significant terms are shown which have a significant quadratic function. Moreover, only those quadratic functions where the change from linear to quadratic model is significant are listed. Direction of ihe effect is presented by + and - symbols. Variable No terms added Food resources Nectar planls (aduit food plants) Corydiilis {larval host plant) Habitat composition Semi-natural grassland Semi-natural grassland-|-semi-natural grassland^ Agricultural area Coniferous trees River Built area Deciduous forest Deciduous forest 4-deciduous forestMixed forest Mixed forest-I-mixed forestHabitat structure Area weighted mean shape index Area weighted mean shape index-I-area weighted msiTopography Altitude mean Altitude maximum Altitude minimum Slope maximum Slope maximum-I-slope maximumSlope mean Slope mean + slope mean^ Climatology Wind speed Wind speed+wind speed^ Management No significant change in the deviance Habitat connectivity No significant change in the deviance Adjusting variables Spatial autocorrelation Spatial autocorrelation + spatial autocorrelation^ Effort (0\l) p<0.01; ***p<O.OOL ECOGRAPHV 24:5 (20011 Deviance 509.07 DF 411 Direclion of the effect 410 410 410 409 410 410 410 410 410 409 410 409 410 409 410 410 410 4fO 409 410 409 410 409 36.62 12.67 38.92 60.06 37.86 9.38 8.43 7.36 0.33 25.17 0.03 18.70 23. S 5 31.23 46.24 41,13 40.75 2.49 11.70 1.26 19.99 18.54 .14.40 +** *** *** ,^^ ** «* ** n.s. *** n.s. *** *** *** ««* •«* + + -h n.s. ** n.s. «** «« * Table 5. Summary of the multiple regression model building procedures (forward selection and backwise deletion of explanatory variables) for abundance of P. mnemosyne. Both forward selection and backwise deletion or variables were employed in the model building, using a strict criterion (p<O.OI) for their inclusion or exclusion. The model employs a Poisson distribution of error assumption and a log link, and is built using P. mnemosyne abundance and environmental data from the 412 grid squares in the study area. Model building procedures with and wiihout a spatial autocorrelation variable are presented. Model building procedure with spatial autocorrelation Model No terms added 1 (adjusting variables) + Spatial autocorrelation + Spatial autocorrelation+ Effort (core area/other area) II (Food resources) + Nectar plants (adult food plants) + Corydalis (larval host plant) III (Habitat composition) + Semi-natural grassland + Semi-natural grassland— Nectar plants (adulf food plants) + Mixed forest + Mixed forest" IV (Climatology) + Wind speed — Semi-nalural grassland" + Radiation " (Topography) V + Slope maximum VI (Habitat structure) + Mean patch size **p<0.01; ***p<0.001; Model building procedure without spatial autocorrelation Model No terms added I (adjusting variables) + Effort (core area/other area) II (Food resources) + Nectar plants (adult food plants) + Corydalis (larvai host plant) III (Habitat composilion) -I- Semi-natural grassland + Semi-natural grassland" + Agricultural land + Coniferous forest + Deciduous forest + Deciduous forest" — Semi-natural grassland' — Coniferous forest + Mixed forest + Mixed forest— Agricultural land IV (Topography) + Altitude mean + Slope mean + Slope mean^ V (Climate) + Radiation V! (Habitat structure) + Shannon's diversity index + Shannon's diversity index• * p < 0 . 0 1 ; ***p<0.001. Deviance 509.07 507.88 465.07 444.20 407.50 395.26 387.98 379.42 379.40 372.51 410 409 408 407 406 405 404 403 402 403 404 DF Change in deviance Deviance 509.07 439.26 414.30 411.85 402.13 388.43 355.30 342.78 347,87 347.11 337.76 328.47 331.62 325.31 318.89 313.48 410 409 408 407 406 405 404 405 404 403 402 403 402 401 400 DF Change in deviance 411 69.80 24.96 2.45 9.72 13.70 33.126 12.53 5.09 0.76 9.34 9.29 3.15 6.31 6.42 5.41 65.16 24.64 2.43 9.84 14.32 37.76 14.77 5.93 0.88 11.14 11.37 3.83 7.80 S.07 6,90 +** *** n.s. «>« if** * * * * • * n^. n.s. • «* *** *** *» '*« 4U 1.18 42.71 20.97 36.70 12.24 7.28 8.55 0.02 6.89 5.68 4.70 3.01 17.16 0.62 11.27 1.37 5.84 9.60 1.01 6.81 0.^. 37.55 19.26 36.66 25.14 7.60 9.11 0;Q2 7:44 5.61 4.96 3.19 19.04 0.68 12.88 1.56 6.78 11.151 1.21 8C27 • ^ . • *• i*. n.s. yn.ii 382.46 379.46 362.30 362.91 351.65 350.28 1,UAA 334.85 333.83 327.02 «* I1.S. ri.s. n.s. #** n.s. «»* n.s. ** ECOOBAPHY 24:5 (2001) Table 6, Variables in fmal multiple GLM regression models for the P. mnemusyne abundance showing coefficients and SE (see Table 5). Models with and without a spatial autocorrelation variable are presented. Model with spatial autocorrelation Variable c, intercept. Spatial autocorrelation Spatial autocorrelation-^ Effort (core area/other area) Corydalis (larval host plant) Semi-natural grassland Mixed forest Mixed forestWind speed Radiation Slope maximum Mean paEch size Coefficient -4,323 1.854 -0.5488 0.3860 0.02765 0.05713 0.1030 -0.006786 -0.3772 0.03958 -0.03679 1.765 SE 0,9391 0,3255 0.1405 0,1486 0,00675 0,00897 0,03127 0,00205 0,08333 0,01274 0,0!576 0,7293 Model without spatial autocorrelation Variable c, intercept. Effort (core area/other area) Nectar plants (adult food plants Corydalis (larval host plant) Semi-natural grassland Deciduous forest Deciduous forest^ Mixed forest Mixed forestAltitude mean Slope mean Slope mean^ Radiation Shannon's diversity index Shannon's diversity indexCoefficient -6.544 0.1290 0.1456 0.02978 0.05694 0.1198 -0.004773 0.1336 -0.006564 -0.02204 0.1623 -0.01177 0.03791 3.932 -i.575 SE 1,531 0,1456 0,03617 0,007136 0,01252 0,02270 0,001490 0,03192 0,001998 0,008844 0,08557 0,004965 0,01329 1,514 0,6331 cover of semi-natural grasslands and in the most optimal areas for abundance the grid squares had a high mean patch size because most of the grid squares were covered by semi-natural grasslands. Relationship between maximum slope angle and abundance was tiegative, which is partly due to the fact that landslides are frequent on the steepest slopes. In areas of landslide scarps low amount of nectar plants and especially Corydah's is characteristic. Habitat loss caused by overgrowth after abandonment of the traditional use and deteriorating habitat quality of semi-natural grasslands are the main threats of decline of P. mnemosyne (Vaisanen and Somerma 1985, Kotiranta et al. 1998). In this study, grassland management by cattle grazing did not show any statistical significance for abundance or distribution of P. mnemosyne. The accuracy of the management variable may influence this result, because only crude binomial data of grazing were available. For example, high intensity of grazing may decrease butterfly abundance by reducing the abundance of nectar flowers and larval host plants. Although the regression models uncovered statistieally significant relationships between the explanatory variables, and the butterfly distribution and abundance, their goodness was relatively modest. More than half of the variation in the abundance remained unexplained, whereas the overall correct classification rate of logistic regression was 93.8%, and 75.4% of the occupied grid squares (sensitivity of the model) were correetly classified. One potential reason for the modest fit of the models may be the scale itself (see Levin 1992). The distribution of a butterfly can be difficult to predict since it is often a combination of large-scale biogeographical variation and fine-scale eeological variation (see Wiens 1989, Levin 1992). The different life stages of the butterfly react on very different scales to their environment. For example, very small scale (metres to centimetres) microclimate factors can have pronounced effect on larval growth rate (Ravenscroft 1994), whereas the population density of adult butterflies is often affected by large scale habitat patterns (Hanski 1994, 1999). The area of grid square (0.25 ha) can be too small for this kind of modelling approach. On the other hand, a large amount of the environmental variation in the heterogeneous landscape would be lost at a coarser scale. An- Table 7, Environmental characteristics of "hotspots" of P. mnemosyne vs other occupied grid squares. Hotspot grid squares were defined as the top 5% grid squares (here 17 grid squares out of 349), ranked by the abundance (captures/visits), P-values were derived from Mann-Whitney U-test (n, = 17, n^ = 332). Variable Nectar plants (0-10) Corydalis (shoots m~-) Cover of semi-natural grassland (m-) Cover of mixed forest (m~) Cover of deciduous forest (m^) Mean altitude (m) Radiation (%) Wind speed (1-4) Mean patch size (m^) Management, grazing (O/I) Habitat connectivity Hotspot - grid squares mean + SD 8.06+ 1.30 13.34 + 8.03 1531 +560 263 + 396 594 + 528 64.3 + 5.8 71,5+4.8 1,47 + 0,51 2118+473 0,76+0,44 3,59 + 0,35 Other - grid squares mean + SD 6,47+ 1,94 9,i5 + 8,l4 1132 + 762 317 + 543 373 + 589 68.6 + 7,3 70.7 + 4.6 1.87 + 0,90 2217 + 867 0.70 + 0.46 3.48 + 0.33 z. -3,527 -2,391 -2,110 -0.340 -2.785 -2.417 -0.678 -1.641 -0.255 -0.554 -1.252 P < 0.001 0,018 0.036 0,685 0,005 0,016 0,497 0,101 0,799 0,580 0,211 ECOGRAPHY 24:5 (2001] Table 8. Predictive capacity of different variable families on their own for P. mnemosyne distribution and abundance when modelled together with the adjusting variables. Factor group Correctly classified (%) No variables added Food resources Habitat composition Management Habitat connectivity Climatology Habitat structure Topography 85.51 93.52 93.36 92.77 92.73 92.57 92.57 92.48 Distribution Number of sig. variables (p<0.01) Change of deviance ('14) Abundance Number of sia. variables (p<O.(H) 1 2': i1 1 2,- other reason for the modest fit of the tnodels may be the hick of historical data of the distribution and managetnciit of semi-natural grasslands in the area. The environmental variables affecting species distribution patterns can be difficull lo separate in a study with only a single area. The modelling of this study could be taken in a methodological sense one step further using predictive modelling, where the data are divided into a training set and into an independent test set (see Le Due et al. 1992). Hotspot characteristics Hotspots of P. mnemosyne differed statistically significantly (p<0.01) in only two environmental variables from other occupied grid squares (Table 7). namely in the amount of nectar planls and the cover of deciduous forest. Interestingly, neither factor became included in the multivariate abundatice model. Although the difference v^fas statistically weaker, the hotspot squares were positively related to the cover of setni-natural grassland (p < 0.036) and abundance of larval host plants ( p < 0.018). Surprisingly, habitat management or climatological variables did not show any statistical significance between hotspots and other grid squares. For example, the northernmost hotspot area was not managed, it is situated on north-east facing slope (low radiation), and it is obviously isolated from the main distribution area. Landscape ecology and metapopulation studies In this study, we chose to use landscape ecological approach instead of metapopulation approach because of several reasons. First, suitable habitat was impossible to define before the field study. It was known that the butterflies tend to occur on semi-natural grassland, but it was unclear whether the deciduous forests with high density of the larval host plant would be an essential part of suitable habitat. Secondly, the seemingly suit- able habitat did not occur in distinct patches, but instead formed a highly well connected and complex corridor-like system along the branches of the river Rekijoki (Fig. 1). Thirdly, within semi-natural grassland the occurrence of the larval host plant Corydalis solida varied from complete absence and scattered occurrence to continuous high density in large areas, which suggested that there might be large continuous variation in habitat quality among the suitable areas for reproduction (Fig. 3b). Fourthly, it seemed possible that other factors like abundance of nectar resources. habitat management by cattle grazing and topographical factors could substantially affect habitat quality. The results showed that potentially suitable habitat for the clouded apollo was determined by the occurrence of semi-natural grassland together with the larval host plant and adult nectar sources, but that there was large continuous and complex variation in habitat quality depending on several factors. An advantage of the use of spatial grid system was that the relative significance of difTerent components of habitat quality could be evaluated at the same time with the area of suitable habitat, habitat connectivity and spatial autocorrelation in butterfly occurrence. Even though suitable habitat in our study system did not occur in discrete habitat patches, the effects of the amount of suitable habitat and habitat connectivity were evident as in many metapopulation studies on butterflies (Hanski and Kuussaari 1995. Thomas and Hanski 1997) and other organisms (Hanski 1994. 1999). The area selected for this study represents the core of a wider clouded apollo occurrence in Ihe Somero. Hiintala region (Kuussaari and Luoto unpubl.). Although this system is highly well connected in its core area, there are more isolated areas of suitable habitat and butterfly occurrence at the edges. The population structure has features of "a patchy population" in the middle part of its distribution area and "a metapopulation" at the margins (see Harrison and Taylor 1997). The use of a spatial grid system helps to avoid difficulties of subjectivity in defining patches of suitable habitat when analysing patterns of distribution and tCOGR.APHY 24:5 (21)UI) abundance in this kind of population structure (e.g. Thomas and Kunin 1999, Tischendorf and Fahrig 2000). It has often been emphasized that the applicability of the metapopulation approach is restricted to cases in which the landscape can be meaningfully divided into suitable and unsuitable habitat (e,g. Hanski and Kuussaari 1995, Hanski and Simberloff 1997). Nevertheless, empirical studies on the factors affecting distribution and abundance at the landscape scale ( 5 100 km-^) of the many species with more complex habitat availability patterns and habitat requirements are rather scarce (Wiens 1997). A landscape ecological approach based on a spatial grid system may prove to be a widely applicable tool for studying these kinds of common situations. Acknowledgements - We thank Kari Haapaia, Julia Poyry and Heikki Tuominen for their help in the field in 1999, Roger L, H. Deniii.s, Risto Heikkinen. Marko Nieminen, Juha Poyry and Raimo Virkkala gave helpful comments on Ihc manuscript. We thank Antii Ponttinen for useful hints in spatial adjustment, Riitta Teiniianta and Tuuli Toivonen helped with several GIS problems, Jussi Kaukoranta produced the wind speed estimations in the Finnish Meteorological Institute, This study was financially supported by ihe Finnish Biodiversity Research Programme (FIBRE). The Finnish Ministry of the Environment and Forest and Park Service.

Journal

Ecography – Wiley

Published: Oct 1, 2001

There are no references for this article.