Fang Liu

Autonomous underwater glider fleets are increasingly deployed to observe mesoscale eddies, yet a methodology for evaluating observation quality under uncertainty remains lacking. This paper presents SCOPE, a framework integrating uncertainty propagation, multi-dimensional assessment, and objective metric selection. A 27-metric evaluation system spanning seven quality categories is constructed; an Adaptive Core Metric Selection (ACMS) algorithm compresses these to a compact core subset. Two-stage sensitivity analysis identifies the velocity ratio as the dominant parameter with a non-monotonic effect. The optimal velocity ratio decreases with eddy intensity in both an analytical model (0.58 to 0.47) and four real HYCOM eddies (2.35 to 0.35). ACMS converges on all four real eddies, and parameter rankings are consistent across environments. An ablation experiment in the analytical model validates the effectiveness of the two sensitivity-derived design rules (25% and 45.8%) gap closure and reveals their non-additive interaction. SCOPE offers a framework and diagnostic workflow for uncertainty-aware evaluation of autonomous ocean observation systems.

Autonomous underwater glider fleets are increasingly deployed to observe mesoscale eddies, yet a methodology for evaluating observation quality under uncertainty remains lacking. This paper presents SCOPE, a framework integrating uncertainty propagation, multi-dimensional assessment, and objective metric selection. A 27-metric evaluation system spanning seven quality categories is constructed; an Adaptive Core Metric Selection (ACMS) algorithm compresses these to a compact core subset. Two-stage sensitivity analysis identifies the velocity ratio as the dominant parameter with a non-monotonic effect. The optimal velocity ratio decreases with eddy intensity in both an analytical model (0.58 to 0.47) and four real HYCOM eddies (2.35 to 0.35). ACMS converges on all four real eddies, and parameter rankings are consistent across environments. An ablation experiment in the analytical model validates the effectiveness of the two sensitivity-derived design rules (25% and 45.8%) gap closure and reveals their non-additive interaction. SCOPE offers a framework and diagnostic workflow for uncertainty-aware evaluation of autonomous ocean observation systems.

Autonomous underwater vehicles (AUVs) have been increasingly applied in marine environmental monitoring. Their outstanding capability of performing tasks without human intervention makes them a popular tool for environmental data collection, especially in unknown and remote regions. This paper addresses the path planning problem when AUVs are used to perform plume source tracing in an unknown environment. The goal of path planning is to locate the plume source efficiently. The path planning approach is developed using the Double Deep Q-Network (DDQN) algorithm in the deep reinforcement learning (DRL) framework. The AUV gains knowledge by interacting with the environment, and the optimal direction is extracted from the mapping obtained by a deep neural network. The proposed approach was tested by numerical simulation and on a real ground vehicle. In the numerical simulation, several initial sampling strategies were compared on the basis of survey efficiency. The results show that direct learning based on the interaction with the environment could be an appropriate survey strategy for plume source tracing problems. The comparison with the canonical lawnmower path used in practice showed that path planning using DRL algorithms could be potentially promising for large-scale environment exploration.