̽��ϵ��

Topic

Bayesian Methods of Integrating Multiple Sources of Data to Estimate Wild Population Abundance

Department of Mathematics and Statistics

Date & location

• Thursday, April 16, 2026
• 11:00 A.M.
• Clearihue Building, Room B017

Examining Committee

Supervisory Committee

• Dr. Laura Cowen, Department of Mathematics and Statistics, ̽��ϵ�� (Supervisor)
• Dr. Michelle Miranda, Department of Mathematics and Statistics, UVic (Member)
• Dr. Saman Muthukumarana, Department of Mathematics and Statistics, UVic (Member)
• Dr. Simon Bonner, Department of Mathematics and Statistics, UVic (Member)
• Dr. Bradley Anholt, Department of Biology, UVic (Outside Member)

External Examiner

• Dr. Marie Auger-Méthé, Department of Statistics, University of British Columbia

Chair of Oral Examination

• Dr. Nikki Macdonald, School of Public Administration, UVic

Abstract

Estimating abundance of wildlife species is a fundamental problem in ecological statistics. Ongoing research continues to develop methods that produce more precise and accurate estimates. Data collected from wild populations are often sparse, noisy, or incomplete. This limits the effectiveness of likelihood based methods which require sufficient data to produce reliable estimates. Bayesian methods provides a robust and reliable framework in such settings by incorporating uncertainty and prior information into the model. As a result, ecological statistics, especially hierarchical modeling has increasingly shifted towards Bayesian approaches. In addition, integrating data from different sources has been shown to improve estimation by reducing bias and increasing both accuracy and precision. Although promising, these type of methods still have some shortcomings to the applicability for some specific type of data.

Application of integrated models to capture-recapture and presence-absence data is dependent on the sampling scheme. We explored the applicability of such datasets collected at the same sampling sites and occasions. In chapter 2, we reviewed two integrated models that combine presence-absence data from camera traps and capture-recapture data from hair snares to compare bias and precision to estimate the population abundance of grizzly bears of the central Rocky Mountains of Alberta, Canada. Unlike many other studies, we found that integrating presence-absence data with capture-recapture data does not improve the precision of the density estimates. The potential reasons for such results are discussed in detail. A possible reason is the violation of one of the integrated models assumptions: independence among multiple datasets. To address this, in chapter 3, we proposed an open population integrated population model that explicitly models the dependence of presence-absence data to capture-recapture data. We performed a simulation study to evaluate the model performance and also investigate the effect of different sampling scenarios on model performance. We compared the integrated population model with the spatially explicit capture-recapture (SECR) model as a single dataset model. Later, we apply the model to the grizzly bear data. From both the simulation study and the case study, we found that the proposed integrated model improved the estimates compared to SECR in terms of accuracy, precision, and bias. This model is more effective than SECR in the various sampling scenarios with budget and logistical constraints.

In chapter 4, we explored the applicability of integrating remote sensing data (chlorophyll-a (chl-a) and sea surface temperature (SST) anomaly) as covariates for salmon stock recruitment models. We used spawner-recruitment data of sockeye salmon from Pitt River, British Columbia, Canada. Remote sensing data were extracted from the central and northern region of the Strait of Georgia (SoG), British Columbia, which represents the ocean entry point for the Pitt River stock when juvenile sockeye salmon migrate towards the ocean. For comparison, we also used in-situ SST data from two of the lighthouses in SoG. The spawner-recruitment relationship was modeled using Ricker and Larkin models. To account for potential temporal autocorrelation, an autoregressive lag 1 (AR(1)) component was also considered. We found that remote sensing chl-a data in the Larkin model were useful to predict Fraser River sockeye salmon stock recruitment; however, in-situ sea surface temperature data outperformed remote sensing sea surface temperature data. We concluded that integrating remote sensing data could improve the stock recruitment forecasting, although longer time series might produce better results.