Item talk:Q258117

{

 "USGS Publications Warehouse": {
   "@context": "https://schema.org",
   "@type": "CreativeWork",
   "additionalType": "Conference Paper",
   "name": "Probabilistic models of seafloor composition using multispectral acoustic backscatter: The benthic detectorists",
   "identifier": [
     {
       "@type": "PropertyValue",
       "propertyID": "USGS Publications Warehouse IndexID",
       "value": "70198608",
       "url": "https://pubs.usgs.gov/publication/70198608"
     },
     {
       "@type": "PropertyValue",
       "propertyID": "USGS Publications Warehouse Internal ID",
       "value": 70198608
     }
   ],
   "inLanguage": "en",
   "datePublished": "2018",
   "dateModified": "2018-08-14",
   "abstract": "We describe and compare two probabilistic models\nfor task-specific seafloor characterization based on multispectral\nbackscatter. We examine whether generative or discriminative\napproaches to supervised seafloor characterization do better\nat harnessing the greatly increased information about seafloor\nsubstrate composition that is encoded in the backscattering\nresponse across multiple frequencies. A Gaussian mixture model\n(GMM) is proposed as a generative model, and a fully-connected\nconditional random field (CRF) is proposed as a discriminative\nmodel. Either model uses input data derived from monospectral\nor multispectral backscatter without modification. The CRF\napproach considers both the relative backscatter magnitudes of\ndifferent substrates as well as their relative proximity, and can\nbe optimized using parameters. The GMM model, in contrast,\nincludes no spatial information in its estimates, being based solely\non relative backscatter magnitudes. Both GMM and CRF modeling\napproaches perform better with multispectral backscatter\ncompared to monospectral, significantly outperforming all three\nmonospectral frequencies. With multispectral backscatter inputs,\nbased on average classification accuracies alone, there was little\nto choose between the two modeling approaches (classification\naccuracy of 81% and 83% for GMM and CRF models, respectively,\nevaluated using 50% of available bed observations to\ntrain and 50% to test the models). However, a CRF model that\nhas been optimized with respect to its tunable parameters tends\nto produce higher posterior probabilities (i.e. greater certainty)\nfor its classifications. Using monospectral backscatter inputs, the\nCRF model significantly outperformed the GMM model in terms\nof average classification accuracy. On balance, therefore, based\non the evidence presented here, the CRF is suggested to be\nthe superior approach for task-specific seafloor classification.\nAlthough further work using additional data is required to\nfurther examine this conclusion, the work presented here will\nguide and focus subsequent research efforts as more areas of\nthe seafloor are mapped with the new technology. In order to\nfacilitate these efforts, the algorithms presented here are encoded\nin a freely available python toolbox for Probabilistic acoustic\nSediment Mapping, called PriSM , that can be used for both\nmonospectral and multispectral backscatter. Finally, we show that\napplication of the CRF model to the outputs of a geoacoustical\nmodel of seafloor scattering results in realistic substrate classification\nboundaries. This hybrid CRF and physics-based approach\ncan predict the physical properties of the seafloor at a finer spatial\nresolution than is possible using the geoacoustical model alone.",
   "description": "30 p.",
   "publisher": {
     "@type": "Organization",
     "name": "GeoHab Conference Proceedings"
   },
   "author": [
     {
       "@type": "Person",
       "name": "Buscombe, Daniel",
       "givenName": "Daniel",
       "familyName": "Buscombe",
       "identifier": {
         "@type": "PropertyValue",
         "propertyID": "ORCID",
         "value": "0000-0001-6217-5584",
         "url": "https://orcid.org/0000-0001-6217-5584"
       }
     },
     {
       "@type": "Person",
       "name": "Grams, Paul E. pgrams@usgs.gov",
       "givenName": "Paul E.",
       "familyName": "Grams",
       "email": "pgrams@usgs.gov",
       "identifier": {
         "@type": "PropertyValue",
         "propertyID": "ORCID",
         "value": "0000-0002-0873-0708",
         "url": "https://orcid.org/0000-0002-0873-0708"
       },
       "affiliation": [
         {
           "@type": "Organization",
           "name": "Southwest Biological Science Center",
           "url": "https://www.usgs.gov/centers/southwest-biological-science-center"
         }
       ]
     },
     {
       "@type": "Person",
       "name": "Kaplinski, Matthew",
       "givenName": "Matthew",
       "familyName": "Kaplinski"
     }
   ],
   "funder": [
     {
       "@type": "Organization",
       "name": "Southwest Biological Science Center",
       "url": "https://www.usgs.gov/centers/southwest-biological-science-center"
     }
   ]
 }

}