Species distribution models (SDMs) are an important tool for biodiversity assessments, ecosystem management and conservation planning, assessments of the impacts of climate change and accelerated land-use change on species distributions, and invasive species monitoring. These models predict the spatial distribution of plant or animal species across a landscape by quantifying the relationship between species distributions and environmental characteristics at known locations. A range of statistical and machine-learning modeling approaches have been employed in SDM but there remains uncertainty about which model types perform best in different modeling situations and which input variables are the most important predictor variables. The objective of this dissertation research is to perform a systematic assessment of the performance of a suite of different SDMs for modeling a range of plant and animal species distributions in Massachusetts (northeastern USA). The study area presents a unique opportunity to assess the performance of the models in a region that is characterized by significant ecological legacies of past human land uses. SDMs are commonly applied in ecosystems characterized by steep ecological gradients, where species reach physiological tolerances or resource limits commonly and there is a high correlation between environmental variables and species distribution. Environmental gradients in the study region are moderate, yet there is a well-established relationship between past land-use and the distribution of plant and animal species.
The results of the study will provide insights into why different models perform differently in different modeling situations, including plant versus animal distributions and highly habitat specific species versus generalists. The distribution of six forest vegetation species will be modeled using a statistic modeling approach (generalized linear model) and two machine-learning approaches (decision trees and the genetic algorithm for rule-set prediction). Habitat distribution modeling will be performed for six rare herptiles (three salamanders and three turtles), using three machine-learning approaches (ecological niche factor analysis, maximum entropy, and genetic algorithm for rule-set prediction). The study will result in improved maps of forest vegetation and rare reptiles and amphibian habitats in the study area. The potential species distributions produced through the modeling exercises will be combined with land-cover maps dating from 1971, 1985, and 1999 to assess the impact of land-cover changes on species distributions. Additionally, the research will provide the basis for outreach programs that will introduce citizen scientists and students at a neighborhood school to mapping techniques and to the extreme importance of monitoring the impacts of landscape change in the region.
