The RSS Glasgow local group and the Environmental Statistics section of RSS co-hosted an event on 9 February 2022 on spatio-temporal modelling and anomaly detection in river networks by Kerrie Mengersen and Edgar Santos-Fernandez from the Centre for Data Science at the Queensland University of Technology.
The speakers presented their most recent research to an audience of 67 people and there was a very lively discussion at the end of the presentation.
A recording of the talk is available on the YouTube channel of the RSS Glasgow local group.
The event started with Kerrie presenting the work in the Centre for Data Science and looking for collaborators for many different applications. Next, different sources of data such as citizen science and satellite data were discussed as well as their incorporation with sensor data for better coverage and prediction. Next, Edgar spoke about the model variables in space-time and specifically about the increasing number of spatial locations, highly correlated observations and the fact that multiple parameters are measures with a single sensor (placed at branches of the river network). It was noted that traditional geostatistical models do not work well for such data with repeated measurements.
Following that the general Bayesian space-time model applied for analysis by Kerrie and Edgar was presented. Different covariance matrices and combinations of them were presented such as AR, VAR and VAR2NN (looking at the two nearest neighbours based on stream or flow). Spatial autocorrelation is incorporated using tail-up and tail-down models as well as Euclidian distance. Subsequently, an application of the model to the Boise River, USA data was presented. The proposed model had a much better prediction than the current published best model based on a split of 80% training data and 20% test data.
The model was used to estimate the exceedance probability for river water temperature above 13o as that is of importance to trout. The model helps identify locations of high temperature where the temperature can be lowered by either planting trees or introducing beavers who build dams and hence, increase water levels which results in reduced temperature.
Next, anomalies in the data were discussed. Anomalies can be caused by battery failure of the sensors, the sensor being outside the water and calibration issues. Anomalies are identified by comparing to k-nearest neighbours both in space and time. Three methods could be used for anomaly detection – extreme value theory, time-series outlier detection and hidden Markov models. A second application on data from the Herbert River, Australia was presented. The application showed excellent performance in detecting anomalies.
Furthermore, the SSNbayes R package, which Kerrie and Edgar develop, was presented. Kerrie and Edgar asked for any feedback from users on it.
The talks were followed by a discussion that focused on different modelling approaches and applications which can be further investigated.
Organisers:
Ben Swallow is currently a Lecturer (Research and Education) in Statistics at the University of Glasgow.
Yoana Napier is a Graduate Teaching Assistant at the University of Glasgow. She has recently defended her PhD thesis on 'High resolution air quality and prediction'.