Adaptive Sampling for Marine Robotics

Sammendrag

Maintaining a healthy ocean is of the utmost importance. Having only a limited set of resources available to study this vast domain requires research and science to focus on more efficient data collection. Determining when and where to sample is, in this regard, a crucial question. The introduction of robotic elements into ocean observation practices have augmented traditional ship-based sampling and provided efficient and reliable sensing platforms for autonomous sampling of oceanographic data, enabling measurements on scales logistically impossible using traditional techniques. However, robotic sampling still relies on deterministic pre-programmed sensing schemes, consisting of sequential waypoints scripted with mission planning tools. In this case, all relevant information is implemented into the mission a priori. This is problematic, since prior knowledge of oceanographic conditions is usually poor leading to substantial uncertainty; consequently, plans for sampling the oceans are suboptimal at best.

Alternatively, the platform can be programmed to adjust the sampling plan online during the mission, capitalizing on both prior and current (in-situ) information. In this setting, sampling happens sequentially over time, according to a conditional plan which changes online during the mission in response to observed data. This type of autonomous sampling scheme is typically referred to as adaptive sampling or data-driven sampling.

Adaptive/data-driven strategies can operate on an a posterior knowledge base and react to current conditions. The impact of this is twofold: i) enabling the sensor platform to divert from the mission if favorable circumstances materialize (opportunistic behavior), and ii) increasing the prospect of retrieving pursued information more effectively. The latter aspect is often considered the most noteworthy, especially for resource intensive environmental sensing applications, having the potential to reduce time and cost.

This thesis presents different methods and applications in adaptive sampling for marine robotics, focusing on exploration of the upper ocean using single platform applications. The coastal ocean and the upper water column are characterized by substantial heterogeneity and spatio-temporal variation. Sampling can therefore benefit from access to synoptic marine data sources such as ocean models and remote sensing, but due to computational limitations and accuracy, these information sources must be used in combination with statistics. Gaussian Processes (GPs) offer a practical probabilistic approach for modeling spatial dependent data and uncertainty. The foundation for the approaches developed here is based on combining GPs with information-theoretic and data-driven criteria to evaluate potential sampling locations. A general problem related to optimization of choosing these sensing locations is the exponential combinatoric increase in dimensionality. The problems are therefore often simplified using heuristics and greedy algorithms.

The principal contributions of this work are related to i) the design and analysis of information-theoretic approaches in upper water column sampling, coupled with intelligent control and ii) testing and validating these methods in the field. This includes a suite of greedy data-driven sampling strategies for the upper water column, developed and tested in full-scale experiments, with applications spanning thermal gradients and internal-waves, assessment of ocean model accuracy, 3-dimensional tracking of sub-surface chlorophyll, and dispersion dynamics in the water column. A proposed methodology for building compact proxy models from remote sensing SST images is also presented using machine learning tools, as well as an application for autonomous mapping of the seafloor. Field testing of these methods presents a considerable challenge given the harsh and dynamic state of the ocean, where large uncertainties and risk are usually the norm rather than the exception. The results show the benefits and potential of using marine data sources and incorporating adaptive sampling routines for exploration of the upper ocean. The emergence of autonomous systems and adaptive sampling does not displace ships or fixed observation stations, however, the introduction of data-driven sampling can greatly augment and increase the observational efficiency and resolution, helping to ensure scientific success.

This thesis is edited as a collection of papers