Research

My primary interest is in geospatial data. Specifically, I think about spatial point patterns, how to design sampling plans to observe point pattern data, and fitting point process models to incompletely-observed point patterns.

At the present, I am not actively pursuing my own academic research. However, if my work helps you, I would love to hear from you! I am also interested in peer-reviewing manuscripts.

Path-Based Sampling Designs

Chapter 4 of my dissertation used a simulation study to compare several methods of constructing paths to traverse around a site when data collection is such that all events within a radius of the path will be observed and all events beyond that radius will be unobserved. I used a spatial LGCP model fit via INLA and the SPDE approximation to predict the intensity surface over the site. My conclusion is that paths with lots of “zigzagging” (sure to become a popular technical term…) do a good job of ensuring that every location in the site is close to some part of the path and lead to predictions that perform well on most meaningful performance statistics.

I created a gallery of plots to help understand the massive volume of simulation output.

My results were plagued with model fits that failed to converge due to poor choice of finite element mesh. Further investigation of mesh construction is a natural next step, and I would be happy to guide any researchers wishing to tackle this.

Fitting LGCP Models with R-INLA

I wrote a tutorial on fitting spatial log-Gaussian Cox process models using with R-INLA and a finite element (SPDE) approximation. Such models are rather nontrivial to set up and there was very little documentation available before I wrote this. I released accompanying example code.

You may also be interested in a vignette of some exploration that lead up to this work, as well as my three minute thesis slides.

DP Infinite Mixture Models

My first foray into spatial point process models was working with magnetometer data from unexploded ordnance remediation projects. The sites I focused on were domestic training/testing ranges where repeated firing at stationary targets justified a mixture of normally-distributed hotspots. I published a study using a Dirichlet process infinite mixture model. Ultimately, the computing time was prohibitively large, and the posterior mixture contained too many redundant components to be useful (i.e. the model is prone to overfitting). I recommend the equally useful and computationally tenable LGCP model instead.

My R package for the DP mixture is out there if you should want to explore it.