Year of Award

2025

Document Type

Thesis

Degree Type

Master of Science (MS)

Degree Name

Wildlife Biology

Committee Chair

Gordon Luikart

Commitee Members

Vicky Dreitz, John Kimball

Keywords

Machine learning, invasive species, remotely sensed data, species distribution modeling, early detection

Subject Categories

Natural Resources and Conservation

Abstract

Effective management of aquatic invasive species (AIS) requires predictive tools that leverage environmental and ecological data to forecast spread into uninvaded areas. However, species distribution models (SDMs) often struggle to capture key invasion processes such as dispersal, propagule pressure, and biotic interactions. This thesis integrates ecological, dispersal, and environmental predictors within a machine learning framework to enhance AIS risk prediction and model transferability across space.

In Chapter 2, 61 novel predictors were developed for 2,203 Minnesota lakes to model zebra mussel (Dreissena polymorpha) occurrence, incorporating variables related to propagule pressure (e.g., boat visitation), community composition (e.g., fish species richness), and water chemistry (e.g., calcium, pH). Models trained with environmental variables alone achieved moderate performance (mean True Skill Statistic, TSS = 0.77). Adding water chemistry and biodiversity predictors modestly improved accuracy (ΔTSS = 0.03–0.05), while including dispersal-based predictors produced the largest gain (mean TSS = 0.95; ΔTSS = 0.22). The most influential predictors were boat visitation, road-network distance to source populations, global human modification, calcium concentration, winter precipitation, and native fish community structure. Chapter 3 evaluated geographic transferability for zebra mussels and Eurasian watermilfoil (Myriophyllum spicatum) using five machine learning algorithms and two ensemble approaches across intrastate (Minnesota) and interstate (Wisconsin) domains. Species identity (η²ₚ = 0.33) and algorithm choice (η²ₚ = 0.18) explained most performance variation, while environmental novelty (MESS) had a smaller but significant effect (η²ₚ = 0.027). Zebra mussel models, led by Random Forest, were highly transferable (TSS = 0.868 intrastate; 0.892 interstate), with strong correspondence between within-extent and transferred predictions (Kulczynski TSS = 0.915–1.000) and minimal uncertainty (SE ≤ 0.032). Eurasian watermilfoil models were more variable (TSS = 0.729–0.831) but improved under ensemble methods. This work demonstrates that integrating species ecology with informative predictors in robust algorithms can substantially improve predictive modeling of aquatic invasions. The resulting modeling pipeline, including scripts for generating predictors and uncertainty maps, provides managers with a transparent and adaptable tool to guide AIS risk assessment and support proactive decision-making.

Download

Share

COinS
 

© Copyright 2025 Leif Eric Howard