Graduation Year

2021

Document Type

Dissertation

Degree

Ph.D.

Degree Name

Doctor of Philosophy (Ph.D.)

Degree Granting Department

Geography, Environment and Planning

Major Professor

Joni Downs Firat, Ph.D.

Committee Member

Steven Reader, Ph.D.

Committee Member

Mark Rains, Ph.D.

Committee Member

Lawrence Hall, Ph.D.

Keywords

Boosted regression trees, Ecology, Habitat modelling, Random forests

Abstract

Management and conservation initiatives will always be controlled by finite resources, whether financialor temporal. Understanding a species’ spatial ecology, and how its requirements vary across habitats and locations is key to a successful species management plan. During recent decades, it has been noted how many species populations have declined, despite conservation practices working to increase their numbers. The most prevalent impacts affecting fauna populations have come from anthropogenic change in the form of habitat loss and destruction, along with fragmentation, and global climate change. There is a clear need for management practices to now operate on an entire landscape instead of focusing on small sites, with spatial and statistical models being developed to address such issues. This study evaluated the performance of two new methods for species habitat analysis. The methods make use of machine learning algorithms, namely random forests and boosted regression trees.

Two case studies were completed comparing the predictive performance of the selected models, with both finding they performed strongly. The first case study made use of presence-absence data for the gopher tortoise and associated environmental habitat variables. The models provided an improved level of accuracy and prediction power compared to a traditional regression model. The spatial distribution of the top habitat model was mapped out across the initial survey areas, providing land managers with more accurate habitat suitability locations to be used for relocated gopher tortoises.

The second case study utilized count data for the Florida manatee. The machine learning models were developed using manatee count data and habitat predictor variables, divided into a winter manatee season, and summer manatee season. Both models had high accuracy however the random forests models performed marginally better. The predicted abundance maps displayed the concentration of manatees around warm waterbodies in the winter, then they dispersed in the summer months as the waters warmed up. The maps produced also indicated the potential for new manatee protection zones around the coastline of Florida. These studies demonstrated the high performance and accuracy of machine learning models, with the conclusion being to recommend their use for ecological analysis in the future.

Share

COinS