Mapping echoes: the acoustics of riverbed and shallow seafloors

Mapping echoes: the acoustics of riverbed and shallow seafloors

NAU School of Computing, Informatics and Cyber Systems

D612e176cb18a5190a722dc0313cb583?s=128

Daniel Buscombe

April 04, 2018
Tweet

Transcript

  1. 3 April 2018 Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 1/41

    1/41
  2. Co-conspirators • Paul Grams, U.S. Geological Survey Grand Canyon Monitoring

    & Research Center • Matt Kaplinski, Sandbar Studies Unit, NAU-SESES Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 2/41 2/41
  3. Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 3/41 3/41

  4. Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 3/41 3/41

  5. Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 3/41 3/41

  6. Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 3/41 3/41

  7. • 1999: 100% surface topography of Mars by NASA’s Mars

    Orbiter Laser Altimeter • 2018: estimated 7 —18% of Earth’s ocean floor mapped at same resolution • 2018: launch of ‘Seabed 2030’ (GEBCO / Nippon Foundation) Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 3/41 3/41
  8. Acoustic remote sensing Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 4/41

    4/41
  9. Acoustic remote sensing Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 4/41

    4/41
  10. Acoustic remote sensing Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 4/41

    4/41
  11. Acoustic remote sensing Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 4/41

    4/41
  12. Acoustic benthic habitat mapping Fonseca et al. (2009) Applied Acoustics

    Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 5/41 5/41
  13. Acoustic benthic habitat mapping Fonseca et al. (2009) Applied Acoustics

    Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 5/41 5/41
  14. Acoustic benthic habitat mapping Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18

    5/41 5/41
  15. Shallow water was technically difficult Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar,

    4/3/18 6/41 6/41
  16. Now, only logistics stand in our way Daniel Buscombe. daniel.buscombe@nau.edu

    NAU-SICCS Seminar, 4/3/18 7/41 7/41
  17. Now, only logistics stand in our way Daniel Buscombe. daniel.buscombe@nau.edu

    NAU-SICCS Seminar, 4/3/18 7/41 7/41
  18. The world’s shallow water: continental shelf • ≈10% < 150

    m average depth • water absorbs > 50% of visible light energy within 10 m Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 8/41 8/41
  19. The world’s shallow water: lakes • Cæl et al. (2017)

    • 278,966,646 lakes • ≈99.9% < 10 m average depth Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 9/41 9/41
  20. The world’s shallow water: rivers • Andreadis et al. (2013)

    • 3,458,785 rivers • ≈85% < 1 m average depth Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 10/41 10/41
  21. Shallow water is disproportionately important • Biodiversity Daniel Buscombe. daniel.buscombe@nau.edu

    NAU-SICCS Seminar, 4/3/18 11/41 11/41
  22. Shallow water is disproportionately important • Biodiversity • Sediment transport

    Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 11/41 11/41
  23. Shallow water is disproportionately important • Biodiversity • Sediment transport

    • Economy Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 11/41 11/41
  24. Shallow water is disproportionately important • Biodiversity • Sediment transport

    • Economy • Human society: 40% global pop. within 10m of sea level (UN, 2007) Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 11/41 11/41
  25. Acoustics in shallow water are different Classical scattering theory not

    applicable: • Bed is acoustically rough Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 12/41 12/41
  26. Acoustics in shallow water are different Classical scattering theory not

    applicable: • Bed is acoustically rough Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 12/41 12/41
  27. Acoustics in shallow water are different Classical scattering theory not

    applicable: • Small beams Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 12/41 12/41
  28. Acoustics in shallow water are different Classical scattering theory not

    applicable: • Small footprints = too few scatterers Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 12/41 12/41
  29. Acoustics in shallow water are different Classical scattering theory not

    applicable: • Significant topographic ‘contamination’ of the backscatter signal Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 12/41 12/41
  30. Acoustics in shallow water are different Classical scattering theory not

    applicable: • Deterministic physics out of favour Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 12/41 12/41
  31. Data-driven substrate classification Hassan et al. (2014) PLoS-ONE Daniel Buscombe.

    daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 13/41 13/41
  32. Data-driven substrate classification Hassan et al. (2014) PLoS-ONE Daniel Buscombe.

    daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 13/41 13/41
  33. Probabilistic approaches • y∗ = arg max y∈M P(x|y)P(y) =

    arg max y∈M P(x, y) • Modeling P(x, y) → generate x Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 14/41 14/41
  34. Probabilistic approaches • y∗ = arg max y∈M P(y|x, θ)

    • Bypasses P(x, y) (complicated) but task-specific • ANN, SVM, RF, etc Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 14/41 14/41
  35. Colorado River in Grand Canyon Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar,

    4/3/18 15/41 15/41
  36. Colorado River in Grand Canyon Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar,

    4/3/18 15/41 15/41
  37. Colorado River in Grand Canyon Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar,

    4/3/18 15/41 15/41
  38. Gaussian Mixture Model P(x|λ) = M ∑ m=1 wm g(x|µm

    , Σm ) • conditional probability of backscatter given λ • per-substrate weight • Gaussian pdf Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 16/41 16/41
  39. Gaussian Mixture Model Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 16/41

    16/41
  40. Filtering the topographic contamination Buscombe et al. (2017) Journal of

    Geophysical Research Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 17/41 17/41
  41. Filtering the topographic contamination Buscombe et al. (2017) Journal of

    Geophysical Research Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 17/41 17/41
  42. Filtering → ‘compositional backscatter’ Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18

    18/41 18/41
  43. Filtering → ‘compositional backscatter’ Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18

    18/41 18/41
  44. Backscatter vs. substrate Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 19/41

    19/41
  45. Backscatter vs. substrate Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 19/41

    19/41
  46. Backscatter vs. substrate Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 19/41

    19/41
  47. GMM results Buscombe et al. (2017) Journal of Geophysical Research

    Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 20/41 20/41
  48. GMM performance Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 21/41 21/41

  49. Colorado River, RM - 4.6 Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar,

    4/3/18 22/41 22/41
  50. Multispectral Backscatter Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 23/41 23/41

  51. Multispectral Backscatter Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 23/41 23/41

  52. Multispectral Backscatter Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 23/41 23/41

  53. Bedford Basin, Nova Scotia Buscombe & Grams (in review) IEEE

    Geo. & Rem. Sens. Letters Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 24/41 24/41
  54. Bedford Basin, Nova Scotia Buscombe & Grams (in review) IEEE

    Geo. & Rem. Sens. Letters Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 24/41 24/41
  55. Patricia Bay, British Columbia Buscombe & Grams (in review) IEEE

    Geo. & Rem. Sens. Letters Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 25/41 25/41
  56. Patricia Bay, British Columbia Buscombe & Grams (in review) IEEE

    Geo. & Rem. Sens. Letters Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 25/41 25/41
  57. Backscatter vs. Substrate Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 26/41

    26/41
  58. Conditional Random Field PΦ (x, y) = I ∏ i=1

    ϕi Di ZΦ (x) = ∑ y PΦ (x, y) P(y|x) = 1 ZΦ (x) PΦ (x, y) P(y|x, θ) = 1 ZΦ (x) exp( −E(y|x, θ) ) Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 27/41 27/41
  59. Fully-Connected Conditional Random Field E(y|x, θ) = ∑ i ψi

    (yi, xi |θ) + ∑ i<j ψij (yi, yj, fi, fj |θ) • unary potentials • pairwise potentials Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 28/41 28/41
  60. Unary and pairwise potentials ψij (yi, yj, fi, fj |θ)

    = Λ(yi, yj |θ) L ∑ l=1 kl (fl i , fl j ) kl (fl i , fl j ) = exp      − |xi − xj |2 2θ2 β      +exp      − |pi − pj |2 2θ2 γ      • compatibility function • proximity tolerance Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 29/41 29/41
  61. CRF result, Bedford Basin Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18

    30/41 30/41
  62. Multispectral vs. Monospectral Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 31/41

    31/41
  63. CRF performance Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 32/41 32/41

  64. CRF performance Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 32/41 32/41

  65. Generative vs. Discriminative Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 33/41

    33/41
  66. Generative vs. Discriminative Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 33/41

    33/41
  67. Deterministic Unary Potentials Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 34/41

    34/41
  68. CRF refinement Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 35/41 35/41

  69. Summary • Acoustic RS poised to explode in shallow water

    Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 36/41 36/41
  70. Summary • Acoustical problems in shallow water can be overcome

    Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 36/41 36/41
  71. Summary • Machine learning for substrate classification Daniel Buscombe. daniel.buscombe@nau.edu

    NAU-SICCS Seminar, 4/3/18 36/41 36/41
  72. Summary • New discriminative algorithm Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar,

    4/3/18 36/41 36/41
  73. Summary • Hybrid physical-statistical modeling Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar,

    4/3/18 36/41 36/41
  74. Thanks! Questions? Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 37/41 37/41

  75. This slide is intentionally blank Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar,

    4/3/18 38/41 38/41
  76. Computing backscatter from echœs • Raw echo (what we measure)

    BS(θ) = EL − SL + 2TL − Af • 10 log 10 of ratios between a quantity and a reference quantity of acoustic pressure of 1 µ Pa • Source level [MEASURED] • Transmission losses [ESTIMATED] • True area of beam footprint [ESTIMATED] Amiri-Simkoœi et al., JASA, 2009; Buscombe et al., JGR, 2014 Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 39/41 39/41
  77. Modeling the signal footprint • model using the sonar geometry

    Buscombe et al., JGR, 2014 Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 40/41 40/41
  78. Modeling the signal footprint • correct using scaling factor =

    f(small scale roughness) Buscombe et al., JGR, 2017 Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 40/41 40/41
  79. What’s next? Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 41/41 41/41

  80. What’s next? Daniel Buscombe. daniel.buscombe@nau.edu NAU-SICCS Seminar, 4/3/18 41/41 41/41