ISEC 2018

From July 1 - 11, I attended the International Statistical Ecology Conference at St. Andrews University in Scotland! 

I participated in many of the abundance estimation, capture-recapture modeling, spatial modeling, and some movement modeling sessions. I found that hidden Markov modeling is all the rate, having foremost applications in movement ecology from what I attended, but finding its way into capture-recapture modeling, and time-series-type analyses.

Not having worked with hidden Markov modeling in the past, I understood it as a model of process & observation where the observations are conditional on an unobserved process that is Markov in nature, with a certain number of underlying states. Examples from this conference included modeling of behavioral states of telemetered animals, and Jolly-Seber-type open population modeling. It seemed to me to be an extension of multi-state modeling, if not very similar, with all states being latent.

Machine learning took center stage in all of the image classification projects, to no surprise. Convolutional neural networks, as I understood it, take transformations of neighborhoods of pixels, with weighted values on the transformations to extract key features of the image. These weights can be tuned manually, but allowing the computer to explore the weighting space unsupervised seems to constitute the deep learning portion of the machine learning algorithm.

Spatial capture-recapture saw a huge amount of progress in many facets of its formulation, with a lot of poking, prodding, and recasting of its assumptions, applications, formulation, and estimation. One extremely interesting model conditions observations of animals on a continuous-time movement model for the individual, instead of conditioning on the latent activity center location. In the example shown, the estimates are not markedly improved over ordinary application of SCR on a real dataset, but this is actually a good sign, since the model allows greater insight into the animal's usage of habitat over the ordinary activity-center-based model. Professor Stephen Buckland's Ph.D. student Richard Glennie presented this material.

Another example of improved SCR questioned the ordinary assumption of isotropic detection as a function of distance. Apart from measurement of explicit covariates affecting detection of an animal that can be integrated in the ecological distance model for SCR (Sutherland et al. 2014), can we model unobserved heterogeneity in detection? Dr. Ben Stevenson of the University of Aukland presented an extension to SCR that convolves the ordinary isotropic distance function (whatever it may be) with an underlying Gaussian random field, thereby accounting for unobserved anisotropy in detection. I heard questions regarding the data required to estimate the spatial covariance parameters, hinting that it may be a large task.

Things that I want to pick up in the future that seem popular in the community are thus hidden Markov modeling, machine learning algorithms, and advanced Bayesian analytical software, for example, Nimble.