2019_programme: BAYESIAN SEAFLOOR CHARACTERIZATION FROM SAS IMAGERY

• Session: 04. Towards Automatic Target Recognition. Detection, Classification and Modelling
  Organiser(s): Groen Johannes, Jans Wolfgang, Pailhas Yan, Myers Vincent
• Lecture: BAYESIAN SEAFLOOR CHARACTERIZATION FROM SAS IMAGERY [invited]
  Paper ID: 875
  Author(s): Gips Bart
  Presenter: Gips Bart
  Presentation type: oral
  Abstract: Synthetic aperture sonar (SAS) allows for high-resolution imagery of the seafloor with almost photo-like quality. From these images a human can identify what type of seabed is in view (e.g. smooth sand, sand ripples, sea grass, rocks). Segmentation into different seabed types can be useful in a setting where autonomous underwater vehicles (AUVs) are used, since bottom type information can dictate the behaviour of the AUV. Specifically in mine countermeasures (MCM) applications, where AUVs are generally tasked to autonomously scan the seabed for mine-like objects on the seafloor. In this situation, recognizing the type of seafloor can inform us about the difficulty the AUV will have when trying to use an automatic target recognition (ATR) algorithm, as different bottom types may lead to a decrease in detector performance or an increase in the number of false alarms. As such this information can be used in a planning and performance evaluation context [1]. Others have shown that seabed characterization is possible using through-the sensor features [2, 3]. In the current work we will focus on predicting the presence of different bottom types based on multifractal analysis of SAS images [4]. Using a Bayesian framework, and a Gaussian process classification (GPC) algorithm [5], we cannot only predict the seabed type, but also provide an uncertainty estimate. Secondly the GPC has the advantage that it will lead to predictions with high uncertainty if “unknown” data are encountered, i.e. when the SAS images contains features that are statistically different from the training dataset.\nWe compare the GPC with a deterministic classification algorithm an highlight the advantageous and disadvantages compared to some of the methodology proposed in earlier work.\n\nReferences\n[1] B. Gips, C. Strode, and S. Dugelay, “Residual risk maps for performance assessment of autonomous mine countermeasures using synthetic aperture sonar,” vol. 40. Institute of Acoustics, 2018, pp. 47–56.\n[2] D. P. Williams, “Fast unsupervised seafloor characterization in sonar imagery using lacunarity,” IEEE transactions on Geoscience and Remote Sensing, vol. 53, no. 11, pp. 6022–6034, 2015.\n[3] D. Kohntopp, B. Lehmann, D. Kraus, and A. Birk, “Seafloor classification for mine countermeasures operations using synthetic aperture sonar images,” in OCEANS 2017-Aberdeen. IEEE, 2017, pp. 1–5.\n[4] D. Carmichael, L. Linnett, S. Clarke, and B. Calder, “Seabed classification through multifractal analysis of sidescan sonar imagery,” IEE Proceedings-Radar, Sonar and Navigation, vol. 143, no. 3, pp. 140–148, 1996.\n[5] C. K. Williams and D. Barber, “Bayesian classification with gaussian processes,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 20, no. 12, pp. 1342–1351, 1998.\n
    Download the full paper
• Corresponding author: Dr Gips Bart
  Affiliation: Centre for Maritime Research and Experimentation NATO STO
  Country: Italy
  e-mail: