A Neural-Statistical Hybrid Model for Spatio-temporal Prediction of Groundwater Dynamics in Bangladesh

Dan Pagandam (Commonwealth Research Organisation)

Wednesday 9th April 14:00-15:00 Maths 311B

Abstract

Deep neural networks are powerful models capable of learning useful representations from large complex datasets for the purpose of prediction. Such models may offer great potential in applied statistical applications if they can be assembled into architectures that facilitate interpretability and that can reliably quantify uncertainty. In this talk we present an application where observed groundwater levels at monitoring wells in the Indo-Gangetic Basin, Bangladesh, were modelled using readily available spatial and spatio-temporal predictors. We constructed a model with two deep neural sub-architectures embedded within a log-additive statistical model. The sub-architectures allowed for the partitioning of groundwater dynamics arising from: (i) local hydrological forcings; and (ii) broad-scale spatio-temporal trends. The separability of these two components facilitated interpretation of the dominant groundwater trends in the basin. The log-additive statistical model also allowed for the quantification of uncertainty in predictions of groundwater levels. We observed good coverage of out-of-sample data, indicating that uncertainty quantification from the deep neural statistical model was reliable. From an application standpoint, we demonstrate that groundwater dynamics can be reliably and easily predicted at large spatial and temporal scales that may aid governments in managing water resources. From a methodological perspective, we demonstrate that it is possible to structure models to harness the powerful attributes of deep neural networks, whilst retaining some of the desirable properties of statistical models.

Add to your calendar

Download event information as iCalendar file (only this event)