Accounting for observation uncertainty in species-habitat models A case study using bird survey data from Poyang Lake, China

Accounting for observation uncertainty in species-habitat models: A case study using bird survey data from Poyang Lake, China by Kyle S Kwaiser A thesis submitted in partial fulfillment of the requirements for the degree of Master of Science Environmental Informatics School of Natural Resources and Environment University of Michigan August, 2009 Thesis Committee: Professor Daniel Brown Professor Kathleen Bergen Acknowledgements First and foremost, I must thank my loving wife, Zana, for her selfless support of my work toward the completion of this Thesis I also owe a great debt of gratitude to my advisors, Dr Daniel Brown and Dr Kathleen Bergen In conjunction with the School of Natural Resources & Environment, Dr Brown supported me financially during my tenure as a Master’s student He also deserves credit for conceiving the original ideas that became central to my Thesis work To Dr Bergen I owe the debt of many hours spent working one-on-one to classify the imagery used in this project The experience of working closely with an accomplished scientist on a near daily basis is not lost on me I would like to thank the many members of the ESA Lab who tolerated my seemingly endless use of the lab computers for my simulation runs More specifically, Shannon Brines, Derek Robinson, Karly Pence-Wentzloff, Hugh Stimson and Qing Tian deserve mention The International Crane Foundation supplied important first-hand knowledge of the Poyang Lake system Finally, I would like to recognize Dr Rick Riolo, Dr Inés Ibáñez, and Dr William Currie, with whom I took courses that greatly expanded my intellectual and analytical abilities Table of Contents Acknowledgements Table of Contents .3 List of Tables .4 List of Figures Abstract Introduction Methods .9 2.1 Study Site .9 2.2 Bird survey and species analyzed .12 2.3 Wetland land-cover classification 14 2.4 Variable-radius simulation: representing uncertainty 16 2.5 Fixed-radius models: a point of departure and comparison .19 2.6 Statistical analysis 19 Results 22 3.1 Wetland classification 22 3.2 Variable-radius simulations 25 3.3 Fixed-radius models 31 Discussion 35 4.1 Interpreting habitat models 36 4.2 Interpreting uncertainty 37 4.3 The role of fixed-radius models 38 4.4 Ramifications for Poyang Lake 39 4.5 Limitations 41 Conclusions 42 Literature Cited 43 Appendix A Python code used to create the simulated landscapes .48 Appendix B R code used to process and analyze outputs from the Python scripts .64 Appendix C Accuracy assessment contingency table for the 2006 classification .101 List of Tables Table Land-cover classes, descriptions and accuracies for the Poyang Lake Basin classification 15 Table Landscape metrics used as potential explanatory variables 19 Table Fixed-radius and simulated hurdle model results for tundra swans 28 Table Fixed-radius and simulated hurdle model results for white-fronted geese .29 Table Comparison of fixed-radius and simulated model performance for the tundra swan dataset 34 Table Comparison of fixed-radius and simulated model performance for the whitefronted goose dataset 35 List of Figures Figure Map of study site: Poyang Lake, Jiangxi Province, China 11 Figure Conceptual model of variable-radius simulation approach .17 Figure Wedge dimensions randomized to create variable-radius landscapes .18 Figure Land-cover map of Poyang Lake 23 Figure Mean land-cover around each sample point .24 Figure Variance in model coefficients for the tundra swan dataset as a function of the number of simulations conducted .26 Figure Variance in model coefficients for the white-fronted goose dataset as a function of the number of simulations conducted .28 Figure Proportion of coefficient estimates depicting a positive or negative relationship for tundra swans 30 Figure Proportion of coefficient estimates depicting a positive or negative relationship for white-fronted geese 31 Figure 10 Histograms of the simulated coefficient values for the tundra swans models 33 Figure 11 Histograms of the simulated coefficient values for the white-fronted goose models 34 Abstract Uncertainty in the location of in situ wildlife observations may impair the performance of habitat models based on those observations In this thesis, I explore the effects of location uncertainty on inference in species-habitat models using a simulation approach to propagate uncertainty in habitat models and quantify its effects Using a point survey of overwintering migratory birds at Poyang Lake, China, I applied Monte Carlo methods to characterize the uncertainty that results when the observer locations and species abundances are known, but the specific directions and distances of the observations (i.e the specific location of the observed landscape) are unknown Habitat models of a tuberfeeding swan (tundra swan; Cygnus columbianus) indicated that the abundance of this species increased with the amount of shallow water, a land-cover class likely to contain submerged aquatic vegetation communities; and that of a grazing goose (white-fronted goose; Anser albifrons albifrons) increased with the amount of sparse live wetland vegetation, a land-cover class found in the transition space between mudflats and perennial wetland vegetation The incorporation of location uncertainty into the habitat models of tundra swans produced uncertainties in the inferred relationship to shallow water, as indicated by changes from positive to negative parameter coefficient estimates, in 15% of simulated models while the relationship between white-fronted goose abundance and sparse live wetland vegetation was positive for 98% of model runs The causes of these changes in inference were highly dependent on landscape configurations and therefore difficult to predict or generalize I found that fixed-radius models (i.e., models constructed assuming that the observed area was a uniform-sized circular area around each sample point) were consistent in terms of direction of effects with the simulation results but should be used with caution when interpreting the strength of the species-habitat relationship from these models due to high variability resulting from observation uncertainty Introduction Species-habitat models relate the presence, absence and/or abundance of species observations to features of the environment in which the individuals are observed Managers and conservationists have used these models to identify the role of particular habitat characteristics in the success of a species (Hines and Hendrix 2005), to facilitate reserve site-selection (Newbold and Eadie 2004) and map the geographic distributions of species (Fielding and Bell 1997) Location uncertainty, defined here as errors or uncertainties in the purported locations of observation data, can impair model performance (Johnson and Gillingham 2008) and exploration of its effects on inference in species-habitat relationships is needed (Graham et al 2008) In this paper, I present a simulation-based approach to incorporate location uncertainty into habitat models, as well as a case study to quantify the effects of this uncertainty on the inference about species-habitat relationships Location uncertainty is endemic to species-habitat studies Sources of uncertainty include a) movement of animals within and between preferred habitat at the time of observation, b) imprecision of observed species geographic coordinates, and c) landscape and habitat complexity (Graham et al 2008) For example, studies of large land mammals conducted with global positioning system (GPS) tracking devices must contend with location uncertainty that can vary with instrument precision, habitat type and topographic diversity Lewis et al (2007) reported a 95% circular error probable (CEP) radius around GPS locations ranging from 20 to 280 m depending on the habitat setting In a study based on satellite telemetry tracking data of right whales (Eubalaena glacialis), the authors averaged environmental variables in a 7.5 km radius around the recorded point of observation to account for potential sensor error and species movement during the observation period (Baumgartner and Mate 2005) Location uncertainty can affect plant habitat studies as well Historical vegetation records of observations used to map the distribution of select chaparral species were assumed to lie with a 600 m radius of an identified sample point (Franklin 2002) Using sensitivity analysis to compare multiple sources of uncertainty in speciesdistribution models of woodland caribou (Rangifer tarandus caribou) in British Columbia, Canada, Johnson and Gillingham (2008) found location uncertainty to be more influential than sampling bias and thematic classification error The authors reported that habitat types that were most selected for (as indicated by higher coincidence rates of the species observation and a given habitat type) tended to show the highest sensitivity to location uncertainty and are the most susceptible to change in the direction of the inferred effect Tests of species distribution models constructed using perturbed observation data for 40 plant and animal species located in Australia, New Zealand or Switzerland showed that model predictions were robust to moderate amounts of location error when compared to predictions gained from the original (unperturbed) location records (Graham et al 2008) Using a point survey of overwintering migratory birds at Poyang Lake, China in conjunction with a wetland classification derived from satellite imagery, I applied Monte Carlo methods to characterize the uncertainty that results when the observer locations and species abundances are known, but the specific directions and distances of the observations (i.e the specific location of the observed landscape) are unknown and potentially variable Specifically, this ‘variable-radius simulation’ creates a series of hypothetical observation viewing windows that were simulated for each sample point to generate equally likely sets of values for landscape variables hypothesized to be predictive of species abundance My aim was to evaluate whether changes in the inferred species-habitat relationships, evaluated through regression coefficient values, occurred when uncertainties related to observer look-directions were simulated In the event that such changes did occur, I also aimed to characterize the frequency of such switches and the distributional features of coefficient values Finally, I contrasted the results from models built using the variable-radius simulations to models constructed using fixedradius areas around each sample point to evaluate the usefulness of such fixed radius models in the face of location uncertainty of species observations Methods In this section I first describe the Poyang Lake study area, what is known of the bird survey used in this study, and the process of selecting and analyzing the two species selected for demonstrating the simulation approach I then describe the classification process employed to characterize the land covers and habitats and the methods used to simulate the lack of knowledge pertaining to look direction and field-sampling methods The section finishes with a presentation of the statistical methods used to model specieshabitat relationships and assess the degree of uncertainty associated with these models 2.1 Study Site Poyang Lake is a seasonally dynamic lake and wetland complex located in China’s Jiangxi Province (Figure 1) The lake, its surrounding lowlands and seasonal wetlands are important for fisheries, agricultural production and for supplying overwintering habitat for over 300 avian species respectively (Shankman and Liang 2003; Ji et al 2007) The Lake, which is connected hydrologically with the Yangtze River, responds to the monsoonal climate and is at low-water level during the winter months from November – February (Shankman and Liang 2003) It is during this time that the Lake becomes a complex ensemble of hydrologically distinct rivers and sublakes interspersed with wetland vegetation It is also at this time that overwintering migratory bird species are present The endangered status of many of these species, such as the Siberian crane (Grus leucogeranus) and the oriental white stork (Ciconia boyciana), motivates a national and international interest in the protection of Poyang Lake’s aquatic systems and habitats The availability of plant communities within the Poyang Lake Basin (PLB) is spatially and temporally heterogeneous and strongly related to terrain and water levels Low elevation areas of shallow water within the main lake basin plus sublakes are populated with submerged aquatic vegetation including Vallisneria spiralis, Potamogeton malaianus and Hydrilla verticillata (Cui et al 2000, Li 2001, Burnham 2007) The tubers of these plants are an important source of food for bird species such as the Siberian crane (Meine and Archibald 1996) and tundra swan (Cygnus columbianus; Zhang and Lu 1999) Slightly higher areas within the main lake basin (Chen et al 2007) are often populated by sedges (Carex spp.) which are grazed by several species of geese such as the white-fronted goose (Anser albifrons albifrons; Zhang and Lu 1999) and the Bean goose (Anser fabalis; Markkola et al 1999) The zone above the sedges is most commonly associated with Miscanthus spp and Cynodon spp (Zeng et al 2007) and is commonly used for grazing water buffalo (Davis et al 2002) 10 plot( ctCumVar[,1], ylab = paste(ctFigLab[1], " Variance", sep = ""), xaxt="n" , mgp =mgpSet, cex.axis = textCEX-.25, cex.lab = textCEX ) plot( ctCumVar[,2], ylab = paste(ctFigLab[2], " Variance",sep = ""), xaxt="n" , mgp =mgpSet, cex.axis = textCEX-.25, cex.lab = textCEX ) plot( ctCumVar[,4], ylab = paste(ctFigLab[4], " Variance",sep = ""), xaxt="n" , mgp =mgpSet, cex.axis = textCEX-.25, cex.lab = textCEX ) par ( mar = c(4, 4.6, 5, 1)) plot( ctCumVar[,3], ylab = paste(ctFigLab[3], " Variance",sep = ""), mgp =mgpSet, cex.axis = textCEX-.25, cex.lab = textCEX, xaxt = "s", xlab= "Number of Simulations" ) # plot ZERO model variance par( mfrow = c(4,1 )) par( oma = c(0, 0, 0, 0), # c(bottom, left, top, right) mar = c(2.5, 4.6, 5, 1)) # c(bottom, left, top, right) plot( zCumVar[,1], ylab = paste(zFigLab[1], " Variance", sep = ""), xaxt="n" , mgp =mgpSet, cex.axis = textCEX-.25, cex.lab = textCEX ) par ( mar = c(4, 4.6, 0, 1)) plot( zCumVar[,2], ylab = paste(zFigLab[2], " Variance",sep = ""), mgp =mgpSet, cex.axis = textCEX-.25, cex.lab = textCEX, xaxt = "s", xlab= "Number of Simulations" ) par(mfrow = c(1,1)) # reset the plot window ##################################################### ########### Do some histogram plots ############### ##################################################### ## BRING IN THE STATIC-RADIUS MODELS (NEED TO HIT regressAnserA_3-18 or regressCynus first) c1km