The application of virtual species to investigate how landscape characteristics affect species distribution model transferability

Liam Grimmett

Research output: ThesisDoctoral Thesis

113 Downloads (Pure)

Abstract

Species distribution models (SDM) play a vital role in understanding the relationship between species and their environment. An application of these models is to transfer them across space and time to understand how species distributions respond under novel conditions. This is done by calibrating models within one spatial or temporal extent and predicting them into another. While SDMs typically model species occurrences as a function of abiotic conditions (e.g. temperature), other factors may influence species distributions such as dispersal, demographic and other biotic processes. In this thesis I investigate how these processes interact with the spatial structure of landscapes using virtual species simulations and investigate how this affects SDM transferability.
Using individual based models that incorporate dispersal and demographic processes, I demonstrate how a species observed response to the environment varies between landscapes. This is driven by local variation in probabilities of occurrence that is determined by local colonisation and extinction rates. These rates vary locally due to the spatial aggregation of suitable habitat, with large patches of high suitability having low extinction rates and high colonisation rates. In contrast, small, fragmented patches as well as areas of low suitability often result in low or zero probabilities of occurrence. These patterns are replicated with a cellular automata (CA) developed in this thesis specifically for simulating virtual species for SDM studies. Resulting simulations not only demonstrate variability in species responses to landscapes, but also differences in habitat suitability maps compared to probability of occurrence maps that is relevant to SDM evaluation.
Using the CA, I simulate virtual species occurrences across several landscapes with different spatial characteristics. I compare the ability of SDMs to predict the species niche (habitat suitability) and occurrences under different model transfer scenarios (unique landscape pairs). Model evaluation demonstrates that even under perfect conditions, models are unable to accurately predict the species’ fundamental niche. The degree of error is dependent on both the unique characteristics of the calibration-transfer landscape and how the virtual species occupies its niche. Subsequent analysis identifies that the range of values occupied by a virtual species and its position relative to the expected response to environmental variables is an important factor determining model accuracy. This is because some common modelling algorithms can either not respond to variables outside the observed range of values or may respond erratically. Models are more accurate when predicting habitat suitability compared to occurrences. This is due to many suitable areas being unoccupied due to the local spatial structure, lowering measured accuracy.
These findings demonstrate a need to standardise methods for delineating regions of interpolation and extrapolation, while placing predictions into context by interpreting response curves in relation to model mechanics as well as its ecological plausibility. Accuracy of SDMs may be improved by incorporating variables that describe landscape structure, particularly when measured against occurrences. Virtual species using the CA developed here would be a valid approach to this but our results also demonstrate a need to reconsider how we evaluate model accuracy in virtual species studies. Future work should describe model performance more holistically rather than focusing on global measures of accuracy.
Original languageEnglish
QualificationDoctor of Philosophy
Awarding Institution
  • Charles Sturt University
Supervisors/Advisors
  • Whitsed, Rachel, Principal Supervisor
  • Horta, Ana, Principal Supervisor
Place of PublicationAustralia
Publisher
Publication statusPublished - 2023

Fingerprint

Dive into the research topics of 'The application of virtual species to investigate how landscape characteristics affect species distribution model transferability'. Together they form a unique fingerprint.

Cite this