Multiscale sources of variation in ecological variables: modeling spatial dispersion,elaborating sampling designs

Detection of structured spatial variation and identification of spatial scales are important aspects of ecological studies. Spatial structures can correspond to physical features of the environment or to intrinsic characteristics of ecological processes and phenomena. Spatial variability has been approached through several techniques such as classical analysis of variance, or the calculation of fractal dimensions, correlograms or variograms. Under certain assumptions, these techniques are all closely related to one another and represent equivalent tools to characterize spatial structures.Our perception of ecological variables and processes depends on the scale at which variables are measured. We propose simple nested sampling designs enabling the detection of a wide range of spatial structures that show the relationships among nested spatial scales. When it is known that the phenomenon under study is structured as a nested series of spatial scales, this provides useful information to estimate suitable sampling intervals, which are essential to establish the relationships between spatial patterns and ecological phenomena. The use of nested sampling designs helps in choosing the most suitable solutions to reduce the amount of random variation resulting from a survey. These designs are obtained by increasing the sampling intensity to detect a wider spectrum of frequencies, or by revisiting the sampling technique to select more representative sampling units.     

Are similar landscapes the result of similar histories?

This landscape study was based on the sampling of 20 replicated landscape sites (1 km² each) that were located within the floodplain of the river Seine. For each site, 13 landscape variables were measured at three dates (1963–1985–2000). The aim of this study was to investigate the overall landscape variability through its different dimensions (space vs. time) and to assess the relative importance of each dimension. We used a new statistical method, i.e., partial triadic analysis (PTA), which allowed us to assess both (1) the spatial variability of the floodplain landscape and its dynamics in time and (2) the dynamic trajectories of the landscape variables for each site. The results showed, at the floodplain scale, the same landscape pattern has emerged since 1963, although a major trend was observed which consisted in a decrease in meadows resulting from an increase in arable crops. At the site scale, landscape sites, even if they were all influenced by this general trend during the 40-year period, showed contrasting trajectories. These results suggest that similar sites in 2000 do not necessarily share common histories and that contrasting sites in 2000 may have originated from similar patterns in 1963. The issue of biodiversity surrogates is then discussed, suggesting that new landscape metrics should be developed, emphasising spatial variability and (or) temporal dynamics.     

Scaling patterns of biomass and soil properties: an empirical analysis

We argue that studies at multiple scales must necessarilychange the extent of measurements, not just the spacing, in order toeffectivelycapture information regarding processes at multiple scales. We have implementeda multi-scale sampling scheme using transects of 10 cm, 1m, 10 m, 100 m, and 1 km ateach of four sites along an elevational gradient from dry foothills forest toalpine tundra in the Front Range of Colorado; these four sites form anadditional transect of 22 km. Along each of these transects wetookten equally spaced soil cores and measured variables important in determiningboth microbial and plant community structure: soil water content, organicmattercontent, pH, and total soil biomass. With this sampling scheme we are able totreat scale as an independent variable in our analyses, and our data show thatboth particular sites and particular variables can determine whether estimatesof mean values are scale-dependent or not. A geostatistical analysis using allof our data shows common relationships between scales across ecologicallydiverse sites; biomass shows the most complex pattern of distribution acrossscales, as measured by fractal dimension. Our analyses also reveal theinadequacy of several standard geostatistical models when applied to data frommultiple scales of measurement – we recommend the use of the boundedpowerlaw model in such cases. We hypothesize that because biological communitiesmustrespond simultaneously to multiple variables with differing patterns of spatialvariation, the spatial variation of biological communities will be at least ascomplex as the most complex environmental variable at any given site.This revised version was published online in May 2005 with corrections to the Cover Date.     

Lacunarity indices as measures of landscape texture

Lacunarity analysis is a multi-scaled method of determining the texture associated with patterns of spatial dispersion (i.e., habitat types or species locations) for one-, two-, and three-dimensional data. Lacunarity provides a parsimonious analysis of the overall fraction of a map or transect covered by the attribute of interest, the degree of contagion, the presence of self-similarity, the presence and scale of randomness, and the existence of hierarchical structure. For self-similar patterns, it can be used to determine the fractal dimension. The method is easily implemented on the computer and provides readily interpretable graphic results. Differences in pattern can be detected even among very sparsely occupied maps.     

Accuracy of fractal dimension estimates for small samples of ecological distributions

We carry out a simulation study of the estimation of fractal dimension in a grid-based setting typical of ecological species distributions, using null landscape models. We calculate the box-counting dimension for samples taken in various types of sampling geometry. Sampler geometries include simple blocks,Cantor grids and line transects. This method may be used to measure fractal dimension of a species distribution, but the accuracy depends on a number of criteria. The most important is sampling effort: any estimate will be inaccurate if the sampling effort is low. We also find the geometry of the sampler to be important. For a given sampling effort, schemes based on the Cantor grids performed better than either line transects or simple blocks. Sampling effort can be improved either by using a bigger sampler over a larger area or by repeated sampling of a smaller area: optimum performance is often a trade-off between these two mechanisms. However, performance is also highly sensitive to the type of fractal object being sampled, with certain types of object requiring a much greater effort for an accurate estimate of fractal dimension. These results raise the possibilities of using novel sampling techniques to estimate fractal dimension, when confronted with limited resources and time, but underline also the need for an understanding of the “type” of fractality expected in ecological situations. This revised version was published online in July 2006 with corrections to the Cover Date.     

Fractal characterization of Mytilus edulis L. spatial structure in intertidal landscapes using GIS methods

The blue mussel, Mytilus edulis L., forms dense and variable patch mosaics composed of aggregates of mussel individuals. Knowledge of mussel bed spatial pattern at multiple scales is important for understanding the form and function of intertidal systems where mussels are prominent features. This study extends prior work demonstrating fractal patterns of mussel boundaries in soft-bottom systems at the quadrat-scale by investigating fractal structure using GIS methods at both the quadrat- and bed-scales. The study pursues three goals for mussel beds in eastern Maine: (1) to compare quadrat-scale fractal dimensions obtained using manual methods with those obtained using digital imagery and techniques, (2) to determine if fractal patterns identified at the quadrat-scale are also present at the bed-scale, (3) and to evaluate the effectiveness of aerial photography and image analysis techniques. Photographs of randomly located quadrats (2500 cm² each) were scan digitized and classified into mussel presence/absence classes. Fractal dimensions of mussel/non-mussel boundaries were calculated using the box-counting method and compared with results obtained using analog photographs and methods. Digital aerial photographs at low tide were acquired for beds at two sites and classified using image processing techniques, and bed-scale fractal dimensions were calculated. At the quadrat-scale, fractal dimensions and their relationship with percent cover differed consistently in absolute value from results using manual methods but agreed in demonstrating fractal patterns for all quadrats and a parabolic trend with percent cover very similar to the one revealed manually. At the bed-scale, both sites were shown to be fractal, with higher dimension value for the bed that subjectively appeared more fragmented and highly dissected. Because mussels are important soft-bottom ecosystem engineers, i.e., foundation species that regulate species composition and abundances, the fractal spatial distribution identified in this study suggests that the species affected by them also exhibit fractal patterns. These results indicate the effectiveness of archive imagery and GIS methods for characterizing intertidal systems and point to the feasibility of future image acquisition.     

Spatial scaling of wildland fuels for six forest and rangeland ecosystems of the northern Rocky Mountains, USA

Wildland fuels are important to fire managers because they can be manipulated to achieve management goals, such as restoring ecosystems, decreasing fire intensity, minimizing plant mortality, and reducing erosion. However, it is difficult to accurately measure, describe, and map wildland fuels because of the great variability of wildland fuelbed properties over space and time. Few have quantified the scale of this variability across space to understand its effect on fire spread, burning intensity, and ecological effects. This study investigated the spatial variability of loading (biomass) across major surface and canopy fuel components in low elevation northern Rocky Mountain forest and rangeland ecosystems to determine the inherent scale of surface fuel and canopy fuel distributions. Biomass loadings (kg?m^?2) were measured for seven surface fuel components??four downed dead woody fuel size classes (0?C6?mm, 6?C25?mm, 25?C75?mm, and 75?+?mm), duff plus litter, shrub, and herb??using a spatially nested plot sampling design within a 1?km² square sampling grid installed at six sites in the northern US Rocky Mountains. Bulk density, biomass, and cover of the forest canopy were also measured for each plot in the grid. Surface fuel loadings were estimated using a combination of photoload and destructive collection methods at many distances within the grid. We quantified spatial variability of fuel component loading using spatial variograms, and found that each fuel component had its own inherent scale with fine fuels varying at scales of 1?C5?m, coarse fuels at 10?C150?m, and canopy fuels from 100 to 500?m. Using regression analyses, we computed a scaling factor of 4.6?m for fuel particle diameter (4.6?m increase in scale with each cm increase in particle diameter). Findings from this study can be used to design fuel sampling projects, classify fuelbeds, and map fuel characteristics, such as loading, to account for the inherent scale of fuel distributions to get more accurate fuel loading estimations.     

Selection of scale for Everglades landscape models

This article addresses the problem of determining the optimal “Model Grain” or spatial resolution (scale) for landscape modeling in the Everglades. Selecting an appropriate scale for landscape modeling is a critical task that is necessary before using spatial data for model development. How the landscape is viewed in a simulation model is dependent on the scale (cell size) in which it is created. Given that different processes usually have different rates of fluctuations (frequencies), the question of selection of an appropriate modeling scale is a difficult one and most relevant to developing spatial ecosystem models. The question of choosing the appropriate scale for modeling is addressed using the landscape indices (e.g., cover fraction, diversity index, fractal dimension, and transition probabilities) recently developed for quantifying overall characteristics of spatial patterns. A vegetation map of an Everglades impoundment area developed from SPOT satellite data was used in the analyses. The data from this original 20 × 20 m data set was spatially aggregated to a 40 × 40 m resolution and incremented by 40 meters on up to 1000 × 1000 m (i.e., 40, 80, 120, 160 … 1000) scale. The primary focus was on the loss of information and the variation of spatial indices as a function of broadening “Model Grain” or scale. Cover fraction and diversity indices with broadening scale indicate important features, such as tree islands and brush mixture communities in the landscape, nearly disappear at or beyond the 700 m scale. The fractal analyses indicate that the area perimeter relationship changes quite rapidly after about 100 m scale. These results and others reported in the paper should be useful for setting appropriate objectives and expectations for Everglades landscape models built to varying spatial scales.     

Using fractal analysis to assess how species perceive landscape structure

To develop a species-centered definition of landscapes, I suggest using a fractal analysis of movement patterns to identify the scales at which organisms are interacting with the patch structure of the landscape. Significant differences in the fractal dimensions of movement patterns of two species indicate that the species may be interacting with the patch structure at different scales. Fractal analysis therefore permits comparisons of landscape perceptions of different species within the same environment.I tested the utility of this fractal application by analyzing the movement patterns of three species of acridid grasshoppers (Orthoptera) in a grassland mosaic. The largest species moved up to 6 times faster than the two smaller species, and species exhibited different responses to microlandscape structure within 25-m² plots. Further, the largest species exhibited different responses to microlandscape structure in two pastures subjected to different intensities of cattle grazing. This species thus is able to integrate information on landscape structure at broad spatial scales. Fractal analysis of movement patterns revealed that the two small species had significantly more tortuous patterns than the larger species, which suggests that these species are interacting with patch structure at a finer scale of resolution than the large species. Fractal analysis can be used to identify the perceptive resolution of a species; that is, the spatial grain and extent at which they are able to perceive and respond to heterogeneity. Analysis of movement patterns across a range of spatial scale may reveal shifts in fractal dimension that reflect transitions in how species respond to the patch structure of the landscape at different scales.     

Fractal characteristics of two-dimensional images of ‘Fuji’ apple trees trained to two tree configurations after their winter pruning

The structural diagrams of apple trees are the comprehensive reflection of the effects of their training and pruning as well as their physiological and ecological characteristics and yield. However, there have been few research reports on the characteristics of the structural diagrams of apple trees. The study investigated the fractal dimension numbers and fractal characteristics of the two-dimensional images of 5-year-old and 10-year-old ‘Fuji’ apple trees trained to the tall spindle configuration and the open-center configuration employing box-counting in combination with the image processing technology of the Photoshop. The two-dimensional images of apple trees with the different configurations differed and varied with their ages. The fractal dimension numbers of the two-dimensional images of the 10-year-old apple trees with the tall spindle configuration and with the open-center configuration were 1.6625 and 1.6531 respectively while the fractal dimension numbers of the two-dimensional images of the 5-year-old apple trees with the tall spindle configuration and with the open-center configuration were 1.6429 and 1.6377 respectively. As the age of the apple trees increased, the spatial quantities and distributions of trunks and branches got slightly intensified, and the fractal dimension numbers of their two-dimensional images and the apple yield increased correspondingly. The comparison of the fractal characteristics of the apple trees with the same age, which were trained to the different configurations, revealed that under the same age, the branch quantities and the apple yield of the apple trees with the tall spindle configuration were higher than those with the open-center configuration, so that under the same age the fractal dimension numbers with the tall spindle configuration were higher than those with the open-center configuration. These results showed that the fractal dimension number of the two-dimensional images of apple trees depended upon their trunk and bough distribution and at the same time it increased with their apple yield as well. Therefore, the fractal dimension numbers of the two-dimensional images of apple trees could be employed as an indicator for assessing training and pruning effects on apple trees and their fruit yield.     

The r.le programs for multiscale analysis of landscape structure using the GRASS geographical information system

Geographical information systems (GIS) are well suited to the spatial analysis of landscape data, but generally lack programs for calculating traditional measures of landscape structure (e.g., fractal dimension). Standalone programs for calculating landscape structure measures do exist, but these programs do not enable the user to take advantage of GIS facilities for manipulating and analyzing landscape data. Moreover, these programs lack capabilities for analysis with sampling areas of different size (multiscale analysis) and also lack some needed measures of landscape structure (e.g., texture).We have developed the r.le programs for analyzing landscape structure using the GRASS GIS. The programs can be used to calculate over sixty measures of landscape structure (e.g., distance, size, shape, fractal dimension, perimeters, diversity, texture, juxtaposition, edges) within sampling areas of several sizes simultaneously. Also possible are moving window analyses, which enable the production of new maps of the landscape structure within windows of a particular size. These new maps can then be used in other analyses with the GIS.     

Spatial and semantic dimensions of landscape heterogeneity

This paper addresses the challenge of measuring spatial heterogeneity in categorical map data. Spatial heterogeneity is a complex notion that involves both spatial variability and attribute variability, and metrics to capture this are a product of their developers’ simplifying assumptions about both spatial and attribute dimensions. We argue that the predominantly binary treatment of categorical data is frequently an unnecessary oversimplification that can be replaced by ordered measures based on semantic similarity evaluations. We develop a typology of autocorrelation metrics for categorical data that identifies a critical gap: existing measures are limited in their ability to capture variability of both spatial and attribute dimensions simultaneously. We demonstrate an approach to formally characterize the semantic similarity between pairs of categorical data classes as a continuous numeric variable. A series of experiments on synthetic and actual land cover data contrasts the information content provided by metrics representative of all portions of the typology: the recently proposed semantic variogram, the indicator variogram, the contagion index, and the edge contrast index. Experimental results suggest that the typology captures essential qualities of metric information richness. Among our findings is that the commonly used contagion index is directly correlated with Moran’s I for 2-class maps but it fails to distinguish between negatively and positively autocorrelated patterns. We identify the semantic variogram as the only metric that can simultaneously detect both spatial and semantic attribute aspects of categorical autocorrelation. The semantic variogram is also relatively robust to attribute scale changes and therefore less sensitive to class aggregation than the other metrics.     

Determination of ecological scale

We suggest that ecological processes and physical characteristics possess an inherent scale at which the processes or characteristics occur over the landscape. We propose a conceptual spatial response model that describes the nature of this ecological scale. Based on the proposed spatial model, we suggest methods for estimating the size of study plots or transects and the distance between replicate plots needed to approach statistical independence. Using data on percent cover for Agropyron spicatum, a common arid-land bunchgrass, we demonstrated four relationships that should hold if the spatial response model is appropriate. These relationships are sample variance increases as functions of (1) transect segment length and (2) intersegment length (transect segment dispersal), and correlation decreases as functions of (3) intersegment length and (4) transect segment length. Based on evaluation of these four relationships, cover for A. spicatum is correlated over the landscape on a scale of 400 to 700 m, and a segment length of 64 to 128 m is most appropriate for measuring cover for this species.     

Effects of changing spatial resolution on the results of landscape pattern analysis using spatial autocorrelation indices

Understanding the relationship between pattern and scale is a central issue in landscape ecology. Pattern analysis is necessarily a critical step to achieve this understanding. Pattern and scale are inseparable in theory and in reality. Pattern occurs on different scales, and scale affects pattern to be observed. The objective of our study is to investigate how changing scale might affect the results of landscape pattern analysis using three commonly adopted spatial autocorrelation indices,i.e., Moran Coefficient, Geary Ratio, and Cliff-Ord statistic. The data sets used in this study are spatially referenced digital data sets of topography and biomass in 1972 of Peninsular Malaysia. Our results show that all three autocorrelation indices were scale-dependent. In other words, the degree of spatial autocorrelation measured by these indices vary with the spatial scale on which analysis was performed. While all the data sets show a positive spatial autocorrelation across a range of scales, Moran coefficient and Cliff-Ord statistic decrease and Geary Ratio increases with increasing grain size, indicating an overall decline in the degree of spatial autocorrelation with scale. The effect of changing scale varies in their magnitude and rate of change when different types of landscape data are used. We have also explored why this could happen by examining the formulation of the Moran coefficient. The pattern of change in spatial autocorrelation with scale exhibits threshold behavior,i.e., scale effects fade away after certain spatial scales are reached (for elevation). We recommend that multiple methods be used for pattern analysis whenever feasible, and that scale effects must be taken into account in all spatial analysis.     

Statistical treatment and comparative analysis of scale-dependent aquatic transect data in estuarine landscapes

Estuarine ecosystem dynamics have evolved around and respond to landscape-level influences that are dynamic in space and time. The estuarine water column is effectively the physical and biologial integrator of these landscape inputs. In this paper, we present a floating window Analysis of Covariance (ANCOVA) technique to statistically compare and contrast aquatic transect data that were taken at different times and under different tidal conditions, yet were geographically parallel and spatially articulate. The floating window ANCOVA compared two transects by testing whether the means of the dependent variable were significantly different while also testing whether the slopes of patterns in the dependent variable were significantly different. By varying the size of the floating window where the ANCOVA was run, we were able to examine how scale affected the magnitude and spatial pattern of that variable. The percentages of total models run, at a given window size, that generated significantly different magnitudes (means) and patterns (slopes) in the dependent variable were referred to as the degree of dissimilarity. Plots of window size versus degree of dissimilarity elucidated temporal and spatial variability in water column parameters at a range of scales. The advantages of this new statistical method in relation to traditional spatial statistics are discussed.We demonstrated the efficacy of the floating window ANCOVA method by comparing chlorophyll and salinity transect data taken at the North Inlet, SC estuary during flooding and ebbing tides in Winter, Spring, and Summer 1991. Chlorophyll concentrations represented the biological characteristics of the estuarine water column and salinity represented the physical processes affecting that water column. We found total dissimilarity in the magnitude of salinity data from one season to the next at all scales, but inter-seasonal similarity in spatial patterns over both short (hourly) and long (monthly) time scales. We also found a large seasonal dissimilarity in the magnitude of chlorophyll levels, as expected. Spatial patterns in phytoplankton biomass (as chlorophyll concentrations) appeared to be largely controlled by the physical processes represented with the salinity data. Often, we observed greater dissimilarity in biological and physical parameters from one tide to the next [on a given day] than from one season to the next. In these cases, the greatest flood-ebb differences were associated with landscape-level influences - from rivers and the coastal ocean - that varied greatly with direction of tidal flow. We are currently using spatially articulate aquatic transect data and the floating window ANCOVA technique to validate spatial simulation models at different scales. By using this variable-scale statistical technique to determine coherence between the actual transect data and model output from simulations run at different scales, we will test hypotheses about the scale-dependent relationships between data resolution and model predictability in landscape analysis.     

The VFractal: a new estimator for fractal dimension of animal movement paths

Fractal measurements of animal movement paths have been used to analyze how animals view habitats at different spatial scales. One problem has been the absence of error estimates for fractal d estimators. To address this weakness, I present and test 4 new estimators for measuring fractal dimension at different spatial scales, along with estimates of their variation. The estimators are based on dividing the movement path into pairs of steps, forming V's, and then estimating various statistics from each V.I measured the performance of these estimators by comparing them to the traditional divider d method, using data generated by two different animal movement models. The estimator based on the net distance between the two steps and the cos turning angle was most accurate, giving estimates similar to those of the traditionally-used divider d method. Precision increased with longer and straighter paths.Strengths of this new estimator are that it can estimate fractal d at different spatial scales, give an estimate of variation, and combine data from many separate path segments which have been gathered at various spatial scales.     

Taxon-dependent Scaling: Beetles,Birds, and Vegetation at Four North American Grassland Sites

Because organisms respond to the environment at different scales, it is important to develop ways of determining the appropriate scales for a specific ecological process and organism. We consider whether the relative importance of different scales is associated with organism mobility, and whether this relationship is independent of landscape characteristics. We observed abundances of particular species for vascular plants, ground-dwelling beetles and breeding birds along eight 2-km transects of 40 sampling stations each, distributed over four sites along the regional gradient from shortgrass steppe in central Colorado to tallgrass prairie in central Kansas. For each transect and taxonomic group, the relative importance of factors measured at the trap scale (1 m; soil texture and hardness, vegetation height, bare ground), at the local scale (10 m; density of shrubs and cacti) and at the landscape scale (30 m; Landsat 7 TM spectral bands, slope and elevation) was assessed using hierarchical canonical variance partitioning with forward selection of explanatory variables. Plant, beetle and bird community composition was explained by environmental factors measured at all three scales. Factor influence was more consistent between transects and between plants and beetles for the more homogeneous landscapes of the shortgrass steppe than for the more heterogeneous landscapes of the tallgrass prairie. We conclude that, independent of the mobility of a taxonomic group, factors at several scales are important in explaining community composition. The importance of different scales shifts along a regional gradient, and the variability between sites is high even for nearby sites.     

Analysis of landscape patterns in coastal wetlands of Galveston Bay,Texas (USA)

High productivity and accessibility have made coastal wetlands attractive sites for human settlements. This study analyzed the patterns of wetland landscapes in Galveston Bay, Texas, USA. The first objective of the study was to describe the relationships between the fractal dimension of wetland boundaries and those factors which affect the wetland landscapes (e.g., land use, type of vegetation, size, location, and level of human disturbance). The second objective was to construct a historical database to contrast wetland areas which had experienced different levels of disturbance between 1956 and 1989. The fractal dimension, a measure of how much of the geographical space is filled by boundaries, was measured by the perimeter-area method. The fractal dimension of wetlands was significantly affected by land use, type of vegetation, size, and level of anthropogenic disturbance. In addition, increasing the size of buffers around roads did not significantly affect the fractal dimension of wetlands. Landscape indices, such as fractal dimension, dominance, and diversity, were used to characterize spatial heterogeneity in the historical database. Lake Stephenson, an area of low anthropogenic disturbance, experienced no changes in wetland composition and abundance over time. Anahuac, an area of medium disturbance, experienced changes in both wetland composition and abundance. Texas City, an area of high disturbance, experienced a change in wetland composition. These differences can be associated with the type and level of disturbance present; however, more evidence is needed to determine whether certain landscape patterns have stable, intrinsic properties which allow persistence in the face of disturbance. These results will be informative to resource managers determining how wetlands can be managed as natural resources and nature reserves.     

Spatial pattern analysis of pre- and post-hurricane forest canopy structure in North Carolina,USA

Existing spatial patterns of a forest are in part a product of its disturbance history. Using laser altimetry and field measures of canopy top height to represent pre- and post-hurricane canopy topography, respectively, we measured changes in spatial patterns of stand structure of a United States southern mixed coniferous-deciduous for est. Autocorrelative and fractal properties were measured in this opportunistic study to quantify changes in canopy architecture along twelve, 190-250 m transects that were subjected to moderate to high levels of wind disturbance. Prior to the hurricane, canopy heights were autocorrelated at scales <40 m with an average fractal dimension of 1.71. After the disturbance, autocorrelation disappeared; the average fractal dimension rose to 1.94. This shift towards spatial randomness illustrates part of the cyclical nature of ecosystem development. It shows how a catastrophic collapse of biomass accumulation corresponds to a decrease in ecosystem organization across a landscape. This revised version was published online in July 2006 with corrections to the Cover Date.     

Measuring boundary convexity at multiple spatial scales using a linear “moving window” analysis: an application to coastal river otter habitat selection

Landscape metrics have been used to quantify ecological patterns and to evaluate relationships between animal presence/abundance and habitat at multiple spatial scales. However, many ecological flows occur in linear systems such as streams, or across patch/landscape boundaries (ecotones). Some organisms and flows may depend on the boundary shape, but metrics for defining linear boundary characteristics are scarce. While sinuosity and fractal dimension address some elements of shape, they fail to specify the dominate shape direction (convexity/concavity). We propose a method for measuring boundary convexity and assess its utility, along with sinuosity and fractal dimension, for predicting site selection by coastal river otters. First, we evaluate the characteristics of boundary convexity using a hypothetical boundary. Second, to compare convexity with other linear metrics boundary convexity, sinuosity and fractal dimension were calculated for the coastline of a set of islands in Prince William Sound, AK. Finally, we use logistic regression in an information-theoretic framework to assess site selection of river otters as a function of these linear metrics. Boundary convexity, fractal dimension and sinuosity are relatively uncorrelated at all scales. Otter latrine sites occurred at significantly more convex locations on the coastline than random sites. Using logistic regression and convexity values at the 100 m window-size, 69.5% of the latrine sites were correctly classified. Coastal terrestrial convexity appears to be a promising landscape-scale metric for predicting otter latrine sites. We suggest that boundary convexity may be an important landscape metric for describing species use or ecological flows at ecotones.