Uncertainty Estimation for Detectability Modelling without Ground Truth

Kurzfassung

Ship wakes are produced by the interaction of the ship’s hull with the ocean water and are result of multiple interacting wave systems closely beneath and on the ocean surface. The ship wake signatures in SAR imagery consists of various components. The most frequently encountered wake components are Kelvin wake arms, V-narrow wake arms and two parts of the turbulent wake: the near field and the far field. The detectability of these four most important wake components in SAR imagery is influenced by several physical variables, which are in the following called influencing parameters. The influencing parameters can be categorized into ship properties, environmental conditions and SAR acquisition settings. In a series of preceding studies of the author, the characteristics of the effects of influencing parameters on the detectability of individual wake components have been modelled using machine learning, and categorized and contrasted against the published state-of-the-art. The presented results are based on a large dataset of TS-X/TD-X acquisitions with a total of 2881 samples of candidate ship wakes. Ground truth is gained by collocation with AIS data and followed by manual inspection procedure. However, the ground truth cannot be applied to estimate the uncertainty of the detectability models, as the models are intended to provide insight into physical dependencies between influencing parameters and wake detectability. For uncertainty estimation the following assumption is used: With an infinite amount of samples in the datasets used for modelling the detectability, when the dataset is divided into infinite subsets and an independent detectability model is built for each subset, then the independent detectability models should be identical. Therefore, the finite real dataset is divided into finite subsets and the deviation between the respective independent detectability models based on the finite subsets is used for uncertainty estimation. The contribute concludes with an analysis of wake detectability taking the estimated model’s uncertain-ties into account.