IMPROVING THE DEVELOPMENT, EVALUATION, AND INTERPRETATION OF SPECIES DISTRIBUTION MODELS

Abstract

Species distribution models (SDMs) encompass a variety of statistical approaches to model the potential geographic distributions of species based on functional relationships between the values of spatial environmental covariates at a set of focal observations (Elith & Leathwick, 2009; Guisan & Zimmermann, 2000). Under the assumption that a species has had sufficient time to reach equilibrium with its environment, SDMs represent an expression of a species’ fundamental environmental niche as formulated by Hutchinson (1957). However, that assumption rarely holds, and thus the distributions characterized by SDMs only represent a partial expression of a species' realized niche. Fully capturing the realized niche means also understanding and incorporating the barriers to dispersal that species face (Barve et al., 2011; Soberón & Peterson, 2005). Implementing movement into SDM workflow has continued to be a difficult task. As a result, models are unconstrained and predict to every available location. For SDMs of invasive species, unconstrained predictions can lead to unrealistic estimates of the amount of suitable habitat due to overestimation (can an organism arrive at the location?) or underestimation (is the species still undergoing range expansion and/or niche infilling?) (Atwater & Barney, 2021; Guisan et al., 2014). Given the current impacts of invasive species to biodiversity, economic activity, human health, and ecosystem services (Simberloff et al., 2013) and the potential for these impacts to worsen under future climate conditions (Bertelsmeier et al., 2013; Finch et al., 2021), there is urgency to understand the processes driving invasions while producing ecological research that can also complement and enhance ongoing prevention, mitigation, and eradication efforts. Mapped predictions of potential establishment of invasive species from SDMs offer land managers information that can be used for spatially prioritization and to allocate limited resources within affected or at-risk landscapes (Reaser et al., 2020). To ensure the efficacy and applicability of SDMs as predictive tools across geographies and taxa, ongoing improvements are needed to advance to the capabilities and methodology of SDMs. The research presented in this dissertation describes how advances in the development, evaluation, and interpretation of SDMs have the potential to improve both ecological understanding and management outcomes. Specifically, the following chapters address three current challenges in species distribution modeling: integrating dispersal constraints, independently validating performance, and understanding the drivers of model predictions. These avenues of investigation demonstrate how SDMs can be enhanced both as tools to explain or identify the potential processes underlying the spatial patterns of distributions or as resources to applied research where predictions serve practical land management and decision-making needs. Overview Chapter one addresses the limitations of unconstrained SDM predictions by augmenting mapped outputs with spatial representations of where species could realistically move within the landscape. By combining current and future habitat suitability with invasion pathway-level dispersal information for over 100 freshwater invaders, this research represents a more realistic and pragmatic depiction of spatial invasion risk. The chapter is also a novel, large-scale application that provides a template for future regional invasion risk assessments focused on supporting management activities like early detection and rapid response. The utility of SDMs for invasive species management is underpinned by the confidence that they are built with the strongest and most relevant measures of model performance. In chapter two, I focus on enhancing existing spatial sampling techniques to address the challenge of model evaluation. Due to bias in species occurrence data from non-random aggregations of observations, traditional random data-splitting can produce misleading and erroneously optimistic performance metrics. Using the inherent gridded geometry of the hashtag (#) shape, I introduce a novel implementation of spatial blocking that provides a robust and ecologically informed methodology that can evaluate spatial ecological models like SDMs for spatial transferability, a key consideration when predicting suitable habitat for invasive species that continue to spread. Even a carefully constructed SDM can be subject to the "black box" problem (Molnar et al., 2020) wherein the behavior and decision making behind model predictions becomes more opaque as methods increase in complexity. In chapter three, I explore the ability of Shapley Additive exPlanations to improve model interpretation and look beyond the predicted values alone. I use a focal SDM to better understand how and why a model generates individual predictions, given the spatial environmental covariates present in the model and the biological context of a focal species on the landscape. By presenting methodological improvements related to the broad-scale integration of invasion pathways, rigorous spatial testing, and out-of-sample importance of model drivers, this work argues that species distribution modeling can move forward as both a tool of explanation and prediction while balancing predictive capacity, spatial transferability, and utility for applied ecological research.