Informative spatial sampling with autonomous underwater vehicles
To understand complex spatio-temporal phenomena in our ocean, there have recently been increased efforts in using numerical process modeling, methods for data assimilation, novel computing and sensor technology. Autonomous robots such as AUVs with onboard computing resources provide rich opportunities for oceanographic sampling, and fill in the gap by adjusting ocean models with in-situ observations. Ideas from statistical sampling design are highly useful in this field, because they enable coherent data assimilation and can help guide AUVs to informative spatial locations. We present approaches of AUV sampling in coastal ocean domains, using a Gaussian process proxy model onboard the AUV. This model can be trained from numerical ocean models, and the Gaussian process is easy to update with in-situ data over the AUV sampling time. With the limited time and resources, we build on the onboard proxy model to develop adaptive sampling design algorithms that enable effective AUV exploration of the ocean domain of interest. Common design criteria here include variance reduction and entropy-based sampling. We further suggest more targeted criteria aiming to find hotspots or map excursion sets where the variables are above a threshold. Such criteria relate the sampling efforts to value of information or active learning approaches. We illustrate these approaches on various real-world experiment example cases. We show how the AUV track the maximum concentration depth within a phytoplankton volume. Another case illustrates river plumes characterisation, where the goal is to use the AUV to explore the boundary between fresh water and the saline ocean water. Yet another application shows mine tailings monitoring efforts using informed AUV sampling. Kanna Rajan will give a connected presentation following this one.