Adaptive sampling of transient environmental phenomena with autonomous mobile platforms

Abstract

In the environmental and earth sciences, hypotheses about transient phenomena have been universally investigated by collecting physical sample materials and performing ex situ analysis. Although the gold standard, logistical challenges limit the overall efficacy: the number of samples are limited to what can be stored and transported, human experts must be able to safely access or directly observe the target site, and time in the field and subsequently the laboratory, increases overall campaign expense. As a result, the temporal detail and spatial diversity in the samples may fail to capture insightful structure of the phenomenon of interest. The development of in situ instrumentation allows for near real-time analysis of physical phenomenon through observational strategies (e.g., optical), and in combination with unmanned mobile platforms, has considerably impacted field operations in the sciences.

In practice, mobile platforms are either remotely operated or perform guided, supervised autonomous missions specified as navigation between human-selected waypoints. Missions like these are useful for gaining insight about a particular target site, but can be sample-sparse in scientifically valuable regions, particularly in complex or transient distributions. A skilled human expert and pilot can dynamically adjust mission trajectories based on sensor information. Encoding their insight onto a vehicle to enable adaptive sampling behaviors can broadly increase the utility of mobile platforms in the sciences. This thesis presents three field campaigns conducted with a human-piloted marine surface vehicle, the ChemYak, to study the greenhouse gases methane (CH₄) and carbon dioxide (CO₂) in estuaries, rivers, and the open ocean.

These studies illustrate the utility of mobile surface platforms for environmental research, and highlight key challenges of studying transient phenomenon. This thesis then formalizes the maximum seek-and-sample (MSS) adaptive sampling problem, which requires a mobile vehicle to efficiently find and densely sample from the most scientifically valuable region in an a priori unknown, dynamic environment. The PLUMES algorithm -- Plume Localization under Uncertainty using Maximum-ValuE information and Search --

is subsequently presented, which addresses the MSS problem and overcomes key technical challenges with planning in natural environments. Theoretical performance guarantees are derived for PLUMES, and empirical performance is demonstrated against canonical uniform search and state-of-the-art baselines in simulation and field trials. Ultimately, this thesis examines the challenges of autonomous informative sampling in the environmental and earth sciences. In order to create useful systems that perform diverse scientific objectives in natural environments, approaches from robotics planning, field design, Bayesian optimization, machine learning, and the sciences must be drawn together. PLUMES captures the breadth and depth required to solve a specific objective within adaptive sampling, and this work as a whole highlights the potential for mobile technologies to perform intelligent autonomous science in the future.