Jennifer A. Miller

Modeling the distribution of four vegetation alliances using generalized linear models and classification trees with spatial dependence

Abstract Generalized linear models (GLMs) and classification trees were developed to predict the presence of four vegetation alliances in a section of the Mojave Desert in California. Generalized additive models were used to provide response shapes for parameterizing GLMs. Environmental variables used to model the distribution of the alliances included temperature, precipitation, elevation, elevation-derived terrain variables (slope, transformed aspect, topographic moisture index, solar radiation, and landscape position), and categorical landform/surface composition variables. Vegetation distributions exhibit spatial dependence and therefore we used indicator kriging to derive neighborhood values of “presence” also used as predictors in the models. The models were developed using 2859 observations coded present or absent for each of the four alliances, and assessed using 960 observations. In general, all of the models were improved with the addition of the kriged dependence term. However, models that relied heavily on the kriged dependence term were less generalizable for predictive purposes. Classification tree models had higher classification accuracy with the training data, but were less robust when used for predictions with the test data. Each of the models was used to generate a map of predictions for each alliance and the results were often quite different. The predicted maps with the kriged dependence terms looked unrealistically smooth, particularly in the classification tree models where they were often selected as the most important variables, and therefore heavily influenced the spatial pattern of the resulting map predictions.

Land-Cover Change Monitoring with Classification Trees Using Landsat TM and Ancillary Data

We monitored land-cover change in San Diego County (1990‐1996) using multitemporal Landsat TM data. Change vectors of Kauth Thomas features were combined with stable multitemporal Kauth Thomas features and a suite of ancillary variables within a classification tree classifier. A combination of aerial photointerpretation and field measurements yielded training and validation data. Maps of land-cover change were generated for three hierarchical levels of change classification of increasing detail: change vs. no-change; four classes representing broad increase and decrease classes; and nine classes distinguishing increases or decreases in tree canopy cover, shrub cover, and urban change. The multitemporal Kauth Thomas (both stable and change features representing brightness, greenness, and wetness) provided information for magnitude and direction of land-cover change. Overall accuracies of the land-cover change maps were high (72 to 92 percent). Ancillary variables representing elevation, fire history, and slope were most significant in mapping the most complicated level of land-cover change, contributing 15 percent to overall accuracy. Classification trees have not previously been used operationally with remotely sensed and ancillary data to map land-cover change at this level of thematic detail.

Mapping Paleo-Fire Boundaries from Binary Point Data: Comparing Interpolation Methods

Abstract Fire history studies have traditionally emphasized temporal rather than spatial properties of paleo-fire regimes. In this study we compare four methods of mapping paleo-fires in central Washington from binary point data: indicator kriging, inverse distance weighting, Thiessen polygons, and an expert approach. We evaluate the results of each mapping method using a test (validation) dataset and receiver operating characteristic plots. Interpolation methods perform well, but results vary with fire size and spatial pattern of points. Though all methods involve some subjectivity, automated interpolation methods perform well, are replicable, and can be applied across varying landscapes.

Species distribution models: Spatial autocorrelation and non-stationarity

The main goal of species distribution modeling is to identify important underlying factors related to broad-scale ecological patterns in order to make meaningful explanations or accurate predictions. When standard statistical methods such as regression are used to formulate these models, assumptions about the spatial structure of the data and the model parameters are often violated. Autocorrelation and non-stationarity are characteristics of spatial data and models, respectively, and if present and unaccounted for in model development, they can result in poorly specified models as well as inappropriate spatial inference and prediction. While these spatial issues are addressed here in an ecological context using species distribution models, they are broadly relevant to any statistical modeling applications using spatial data.

Analysis of movement data

The study of movement is progressing rapidly as a subdiscipline in Geographic Information Science (GIScience). At the fulcrum of this new research area in GIScience are movement observations. Movement observations may be understood as spatiotemporal signals, which carry information on the movement of dynamic entities and the underlying mechanisms that drive their movement. These observations are key to the study and understanding of movement. Technological advancements in global positioning systems (GPS) and related satellite tracking technologies have resulted in significant increases in the availability of highly accurate data on moving phenomena, dramatically outpacing the development of appropriate methods with which to analyze them. In addition to increased spatial accuracy and temporal resolution of the locational information, improvements are being made to accelerometers and ‘biologgers’ that enable the collection of ancillary behavioral and physiological information. This special issue emerged from a pre-conference event associated with the GIScience 2014 conference held in Vienna: a workshop organized by the authors on ‘Analysis of Movement Data’ (AMD 2014). The workshop and this special issue explore recent trends in the study of movement and novel methods for analyzing and contextualizing movement data. A broad range of topics is covered concerning movement analysis, representation, and modeling. The studies presented use movement data from different domains, such as transportation (vehicles, marine traffic), cyclists and athlete tracking data, storm events, and movement ecology (birds, mammals, etc.). This editorial intends to frame and position the papers included in this special issue and to provide recommendations for future directions in the analysis of movement data. In order to frame the work presented here, we use the overarching research framework for the study of movement proposed by Dodge (2015). This framework, shown in an adapted version in Figure 1, posits that the study of movement consists of a continuum of research ranging from understanding movement to construct knowledge of the behavior of dynamic objects, to using this knowledge for modeling and prediction of movement. Visualization facilitates this process through data exploration, hypothesis generation, and communication of the outcomes (Andrienko et al. 2013, Wood et al. 2011, Zhang et al. 2013, Xavier and Dodge 2014). The framework relies on an iterative validation process, where analytics and models are parameterized, calibrated, and improved using real movement observations. Understanding movement, shown on the right side of Figure 1, entails development of methods for quantification of movement and its parameters (Dodge et al. 2008, Long and Nelson 2013, Laube 2014, Demšar et al. 2015); analysis of its context INTERNATIONAL JOURNAL OF GEOGRAPHICAL INFORMATION SCIENCE, 2016 VOL. 30, NO. 5, 825–834 http://dx.doi.org/10.1080/13658816.2015.1132424

Stratified Sampling for Field Survey of Environmental Gradients in the Mojave Desert Ecoregion

Environmental gradients, represented by mapped physical environmental variables within a GIS, were classified and used to allocate a two-stage random stratified sample for field survey of vegetation in the Mojave Desert Ecoregion, California. The first-stage sample was allocated randomly and with unequal proportions among 129 environmental classes defined by the intersection of climate and geology digital maps at 1 km resolution. The second-stage sample was selected for each 1 km cell in the primary sample by defining six terrain classes related to desert vegetation patterns and randomly locating one plot location per class per cell. The total number of observations (1133) was determined by the resources available for the survey. This approach allowed the vegetation survey to be planned efficiently, alternate samples to be located, and vegetation types to be defined quantitatively. The sample allocated surveyed broad scale environmental gradients effectively, and the objective of oversampling rare environmental classes and undersampling common classes was achieved in most cases. It did not succeed, however, in capturing replicates of rarer plant alliances. We suggest sampling efforts should be weighted even more heavily toward rare environments and plant communities for this objective.

Virtual species distribution models Using simulated data to evaluate aspects of model performance

Species distribution models (SDMs) have become a dominant paradigm for quantifying species-environment relationships, and both the models and their outcomes have seen widespread use in conservation studies, particularly in the context of climate change research. With the growing interest in SDMs, extensive comparative studies have been undertaken. However, few generalizations and recommendations have resulted from these empirical studies, largely due to the confounding effects of differences in and interactions among the statistical methods, species traits, data characteristics, and accuracy metrics considered. This progress report addresses ‘virtual species distribution models’: the use of spatially explicit simulated data to represent a ‘true’ species distribution in order to evaluate aspects of model conceptualization and implementation. Simulating a ‘true’ species distribution, or a virtual species distribution, and systematically testing how these aspects affect SDMs, can provide an important baseline and generate new insights into how these issues affect model outcomes.

Incorporating movement in species distribution models

Movement in the context of species distribution models (SDMs) generally refers to a species’ ability to access suitable habitat. Movement ability can be determined by some combination of dispersal constraints or migration rates, landscape factors such as patch configuration, disturbance, and barriers, and demographic factors related to age at maturity, mortality, and fecundity. Including movement ability can result in more precise projections that help to distinguish suitable habitat that is or can be potentially occupied, from suitable habitat that is inaccessible. While most SDM studies have ignored movement or conceptualized it in overly simplistic ways (e.g. no dispersal versus unlimited dispersal), it is increasingly important to incorporate realistic information on movement ability, particularly for studies that aim to project future distributions such as climate change forecasting and invasive species applications. This progress report addresses the increasingly complex ways in which movement has been incorporated in SDM and outlines directions for further study.

Explicitly incorporating spatial dependence in predictive vegetation models in the form of explanatory variables: a Mojave Desert case study

Predictive vegetation modeling is defined as predicting the distribution of vegetation across a landscape based upon its relationship with environmental factors. These models generally ignore or attempt to remove spatial dependence in the data. When explicitly included in the model, spatial dependence can increase model accuracy. We develop presence/absence models for 11 vegetation alliances in the Mojave Desert with classification trees and generalized linear models, and use geostatistical interpolation to calculate spatial dependence terms used in the models. Results were mixed across models and methods, but in general, the spatial dependence terms more consistently increased model accuracy for widespread alliances. GLMs had higher accuracy in general.

Towards a Better Understanding of Dynamic Interaction Metrics for Wildlife: a Null Model Approach

The ability to measure dynamic interactions, such as attraction or avoidance, is crucial to understanding socio-spatial behaviors related to territoriality and mating as well as for exploring resource use and the potential spread of infectious epizootic diseases. In spite of the importance of measuring dynamic interactions, it has not been a main research focus in movement pattern analysis. With very few exceptions (see Benhamou et al. 2014), no new metrics have been developed in the past 20 years to accommodate the fundamental shift in the type of animal movement data now being collected and there have been few comparison or otherwise critical studies of existing dynamic interaction metrics (but see Long et al. 2014; Miller 2012). This research borrows from the null model approach commonly used in community ecology to compare six currently used dynamic interaction metrics using data on five brown hyena dyads in Northern Botswana. There was disconcerting variation among the dynamic interaction results depending on which metric and which null model was used, and these results highlight the need for more extensive research on measuring and interpreting dynamic interactions in order to avoid making potentially misleading inferences about socio-spatial behaviors.