Evolving Gaussian Process for modeling Spatiotemporally Varying Processes

Distributed Autonomous System Laboratory

Collaborators and Students: 

Harshal Maske

Project Description: 

We address the key challenges in monitoring spatiotemporal phenomena, which are to estimate its current state, predict its future evolution, and infer the initial conditions utilizing limited sensor measurements. The problem manifests due to the fact that it is typically infeasible or expensive to deploy sensors on a large scale across vast spatial domains. We demonstrate a novel approach, by layering a dynamical systems prior over temporal evolution of weights of a kernel model to perform spatiotemporal modeling. The model is then used to determine the number of sensors as well as their locations to fully predict the evolution of a spatiotemporal process.


Figures: Feasibility of this approach was tested on a very large dataset: the 4 km AVHRR Pathfinder project, which is a satellite monitoring climate data for sea surface temperature. The dataset is challenging, with measurements at over 37 million coordinates, and several missing pieces of data. The goal was to learn the day and night temperature models. Below, the original data is on the left and the estimated data is on the right.

Figure: Pathfinder, Day, Original Data Figure: Pathfinder, Day, Estimated Data
Orginal Data from Pathfinder Estimated Data
Figure: Pathfinder, Night, Original Data Figure: Pathfinder, Night, Estimated Data
Original Pathfinder Data, Nigth Estimated Data, Night