Graduation Year

2024

Document Type

Dissertation

Degree

Ph.D.

Degree Name

Doctor of Philosophy (Ph.D.)

Degree Granting Department

Integrative Biology

Major Professor

Andrew M. Kramer, Ph.D.

Committee Member

Diego Santiago-Alarcon, Ph.D.

Committee Member

Ryan Carney, Ph.D. MBA, MPH

Committee Member

Joni Downs, Ph.D.

Keywords

Biogeography, Environmental Niche Model, Forecasting, Machine Learning, Spatiotemporal, WorldClim

Abstract

Species distribution models (SDM) are frequently used to gain ecological understanding and guide conservation decisions. These models are developed with a wide variety of procedures - from regression-based approaches to more modern machine learning algorithms - but a requirement almost all share is the use of predictor variables that strongly simplify the temporal variability of driving factors. These static features are chosen by the user, which can incorporate unnecessary bias into a study. Additionally, current data preparation techniques often remove variables from the process entirely, which inherently removes information that may still be of importance. Conversely, novel architectures of deep learning neural networks allow dealing with fully explicit spatio-temporal dynamics, and thus fitting SDMs without the need to simplify the temporal and spatial dimension of predictor data.

This dissertation investigates the theoretical benefits of time series deep learning approaches in SDMs and provides empirical evidence to their predictive capabilities. In the first chapter, I present a deep learning-based SDM approach that uses time series of spatial data as predictors. The method employs AutoML features to create candidate models, making the code accessible to ecologists with minimal experience in deep learning, while still offering flexibility for various datasets and modeling scenarios. I demonstrate the use of time-series predictors for a vetted dataset of 74 species used in other rigorous assessments of SDM methods. I illustrate the suitability of this method by comparing it to standard machine learning methods fit to conventional static predictors. I show the methods here effectively enable deep learning to independently find appropriate features within the time series data that characterize the conditions associated with the likelihood of species occurrences.

For Chapter Two, I extended the efforts found in Chapter One by performing a comprehensive examination of various deep learning methods for classifying species distributions with time series data on the same 74 species. Specifically, I examined the performance of 16 modeling protocols, which combined four widely used deep learning architectures designed for temporal data processing, with different monitored learning measurements and validation criteria. The results of this analysis indicate that using classification accuracy as the model selection metric, along with validation loss as the monitored learning measurement, is likely to yield improved outcomes, particularly for balanced datasets.

Finally, the third and final chapter of this dissertation explores how well time series-based SDMs learn temporal trends for forecasting species distribution. Specifically, I used the top performing protocols from Chapter Two time series classification to evaluate the climate dynamics of current and future conditions to determine the environmental suitability for a heavily monitored game bird, the chukar partridge (Alectoris chukar). I showed that the model predictive accuracy of this approach was moderate to high, and consistently produced useful predictions for both five and ten-year time intervals.

Download

Share

COinS