{"title": "Bayes Networks on Ice: Robotic Search for Antarctic Meteorites", "book": "Advances in Neural Information Processing Systems", "page_first": 988, "page_last": 994, "abstract": null, "full_text": "Bayes  Networks  on  Ice: \n\nRobotic  Search  for Antarctic  Meteorites \n\nLiam Pedersen-,  Dimi Apostolopoulos, Red Whittaker \n\nRobotics Institute \n\nCarnegie Mellon University \n\nPittsburgh, PA 15213 \n\n{pedersen+,  dalv,  red}@ri.cmu.edu \n\nAbstract \n\nA  Bayes  network  based  classifier  for  distinguishing  terrestrial \nrocks  from  meteorites  is  implemented  onboard  the  Nomad  robot. \nEquipped with a camera,  spectrometer and eddy current sensor, this \nrobot searched the  ice  sheets of Antarctica and autonomously made \nthe first robotic identification of a meteorite, in January 2000 at the \nElephant Moraine.  This paper discusses  rock classification from  a \nrobotic platform, and describes the system onboard Nomad. \n\n1 \n\nIntroduction \n\nFigure  1  :  Human  meteorite  search  with  snowmobiles  on  the  Antarctic  ice \nsheets, and on foot in the moraines. \n\nAntarctica  contains  the  most  fertile  meteorite  hunting  grounds  on  Earth.  The \npristine,  dry  and  cold  environment  ensures  that  meteorites  deposited  there  are \npreserved  for  long  periods.  Subsequent  glacial  flow  of the  ice  sheets  where  they \nland  concentrates  them  in  particular  areas.  To  date,  most  meteorites  recovered \nthroughout  history  have  been  done  so \nlast  20  years. \nFurthermore, they are less likely to  be contaminated by terrestrial compounds . \n\nin  Antarctica  in  the \n\n\u2022 http://www.cs.cmu.edu/-pedersen \n\n\fMeteorites  are  of interest  to  space  scientists  because,  with  the  exception  of the \nApollo  lunar  samples,  they  are  the  sole  source  of extra-terrestrial  material  and  a \nwindow  on  the  early  evolution  of the  solar  system.  The  identification  of Martian \nand  lunar  meteorite  samples,  and  the  (controversial)  evidence  of fossil  bacteria  in \nthe  former  underscores  the  importance  of systematically retrieving as  many samples \nas possible. \n\nCurrently,  Antarctic  meteorite  samples  are  collected  by human  searchers,  either  on \nfoot,  or  on  snowmobiles,  who  systematically  search  an  area  and  retrieve  samples \naccording  to  strict  protocols. \nIn  certain  blue  ice  fields  the  only rocks  visible  are \nmeteorites.  At other places (moraines - areas where the  ice flow brings rocks to  the \nsurface) searchers have to contend with many terrestrial rocks (Figure  1). \n\n1.1  Robotic  search  for  Antarctic  meteorites \n\ncolor \ncamera \n\nreflectance \nspectrometer \n\nFigure  2  :  Nomad  robot,  equipped  with  scientific  instruments, \ninvestigates a rock in Antarctica. \n\nWith the goal of autonomously search for  meteorites in Antarctica,  Carnegie Mellon \nUniversity  has  built  and  demonstrated  [1]  a  robot,  Nomad  (Figure  2),  capable  of \nlong  duration  missions  in  harsh  environments.  Nomad  is  equipped  with  a  color \ncamera on a pan-tilt platform to  survey the ice for rocks and acquire close up  images \nof any  candidate  objects, and a manipulator arm to  place the fiber optic probe of a \nspecially  designed  visible  light  reflectance  spectrometer  over  a  sample.  The \nmanipulator arm can also place other sensors, such a metal detector. \n\nThe eventual goal, beyond Antarctic meteorite search,  is  to  develop technologies for \nextended  robotic  exploration  of remote  areas,  including  planetary  surfaces.  One \nparticular  technology  is  the  capacity  to  carry  out  autonomous  science,  including \nautonomous  geology  and  the  ability  to  recognize  a  broad  range  of rock  types  and \nnote exceptions. \n\nIdentifying  meteorites  amongst  terrestrial  rocks  is  the  fundamental  engineering \nproblem  of robotic  meteorite  search  and  is  the  topic  addressed  by the  rest  of this \npaper. \n\n2  Bayes  network  rock  and  meteorite  classifier \n\nClassifying rocks from a mobile robotic vehicle entails several unique issues: \n\n\u2022  The  classifier  must  learn  from  examples.  Human  experts  often  have  trouble \nexplaining how they can identifY many rocks,  and will  refer to  an example.  In \nthe  words  of a  veteran  Antarctic  meteorite  searcher  [2]  \"First  you  find  a  few \nmeteorites, then you know what to look for\". \n\n\fA  complication is  the  difficulty of acquiring  large  sets  of training  data,  under \nrealistic field conditions.  To  date  this  has  required two  earlier expeditions  to \nAntarctica,  as  well  as  visits  to  the  Arctic  and  the  Atacama  desert  in  Chile. \nTherefore,  it  is  necessary  to  constrain  a  classifier  as  much  as  possible  with \navailable prior knowledge,  so  that training can be accomplished with minimum \ndata. \n\n\u2022  The  classifier must  be  able  to  accept  incomplete  data,  and compound  evidence \nfor  different hypotheses as  more information becomes available.  The robot has \nmultiple  sensors,  and  there  is  a  cost  associated  with  using  each  one.  Sensors \nsuch  as  the  spectrometer  are  particularly  expensive  to  use  because  the  robot \nmust  be  maneuvered  to  bring  the  rock  sample  into  the  sensor  manipulator \nworkspace.  Therefore,  it  is  desirable  that  initial  classifications be made  using \ndata  from  cheap  long  range  sensors,  such  as  a  color  camera,  before  final \nverification using expensive sensors on promising rock samples. \n\nA corollary of this is that the classifier should accept prior evidence from other \nsources,  such  as  an  experts  knowledge  on  what  to  expect  in  a  particular \nlocation. \n\n\u2022  Rock  classes  are  often  ambiguous,  and  the  distinctions  between  certain  types \n[3].  The  classifier  must  handle  this  ambiguity,  and  indicate \n\nfuzzy  at  best \nseveral likely hypotheses if a definite classification cannot be achieved. \n\nThese  requirements  for  a  robotic  rock  classifier  argue  strongly in  favor  of a  Bayes \nnetwork  based  approach,  which  can  satisfy  them  all.  The  intuitive  graphical \nstructure  of a Bayes network makes  it easier to  encode physical constraints  into  the \nnetwork topology,  thus  reducing  the  intrinsic  dimensionality.  Bayesian update  is  a \nprincipled  way \nis  naturally \nrepresented by prior probabilities. \nAdditionally,  with  a  Bayes  network  it  is  simple  to  compute  the  likelihood  of any \nnew  data,  and  thus  conceivably  recognize  bad  sensor  readings.  Furthermore,  the \nnetwork can be  queried to  estimate  the  information  gain  of further  sensor readings, \nenabling active sensor selection. \n\nto  compound  evidence,  and  prior  information \n\n2.1  Network  architecture \n\nThe  (simplified)  network  architecture \nfor  distinguishing  rocks  from  meteorites, \nusing features  from sensor data,  is  shown in Figure 3.  It is a compromise between a \nfully  connected  network \n(no  constraints  whatsoever,  and  computationally \nintractable)  and  a  naive  Bayes  classifier  (can  be  efficiently  evaluated,  but  lacks \nsufficient representational power).  Sensor features  are  only weakly  (conditionally) \ndependent  on  each  other  because  of a  careful  choice  of suitable  features,  and  the \nintermediate  node  Rock-type,  whose  states  include  all  possible  rock  and meteorite \ntypes likely to be encountered by the classifier. \n\nA  complication  is  that  the  sensor  features  are  continuous  quantities,  yet  the  Bayes \nnetwork  implementation  can  only  handle  discrete  variables. \nTherefore  the \ncontinuous variables need to be suitably quantized. \n\n\fRock/Meteorite \ntype \nIron meteorite \nSandstone \n\nMeteorite? \nTrue \nFalse \n\nFigure 3  :  Bayes network for discriminating meteorites and rocks based on features \ncomputed from sensor data. \n\n2.2  Sensors  and  feature  vectors \n\n<: \n\n\u2022 u \n~ \u2022  0.5 \n-e \n\u2022 \n~ ..  0 \n! \n\npeak \n,....----, \n\nstrengt  of  peak \n(+) or trough (-)  at \ngiven wavelength \n\n400 \n\n600 \n\n800 \n\nwavelength /[nm] \n\n1000 \n\nFigure 4  : Example spectrum (with extracted features)  and color images of rocks \non ice.  One of the rocks in the image is  meteorite. \n\nIn  Antarctica Nomad  acquired  reflectance  spectra  and  color  images  (Figure  4)  of \nsample  rocks.  Spectra  are  obtained  by  shining  white  light  on  the  sample  and \nanalyzing the  reflected light to  determine  the fraction  of light reflected at a  series of \nwavelengths. \nThe  relevant  features  in  a  spectrum,  for  the  purpose  of identifying  rocks,  are  the \npresence, location and size of peaks and troughs in the  spectrum (Figure 4), and the \naverage  magnitude  (albedo)  of the  spectrum  over  certain  wavelengths.  Spectral \ntroughs  and peaks  are  detected by computing the  correlation of the  spectrum with a \nset of 10 templates over a finite region of support (50 nm).  Restricting the  degree of \noverlap between  templates  minimizes  statistical dependencies between the  resulting \nspectral  features  (Figure  3).  Normalizing  the  correlation  coefficients  makes  them \n(conditionally)  independent  of the  average  spectral  intensity  and  robust  to  changes \nto  scale  (important,  because  in  practice,  when  making  a  field  measurement  of a \nspectrum it is  difficult to  accurately determine the  scale).  A  13  element real valued \nfeature  vector  (each component corresponding to  a  sensor feature  node  in Figure  3) \nis thus obtained from the original 1000+ element spectrum. \n\nColor  images  are  harder  to  interpret  (one  of the  rocks  in  Figure  4  is  a  meteorite). \nFirst the  rock  needs  to  be  segmented  from  the  background  of snow  and  ice  in  the \n\n\fimage,  using  a  partially observable  Markov  model  [4].  Features  of interest are  the \nrock cross  sectional  area (used as  a proxy for  size,  and requiring that the  scaling of \nthe  images  be  known),  average  color,  and  simple  texture  and  shape  metrics  [4]. \nMeteorites  tend  to  be  small  and  dark  compared  to  terrestrial  rocks.  An  8  element \nreal valued feature vector is computed from each image. \n\nAll real valued features  are  quantized prior to  being entered into the Bayes network, \nwhich cannot handle continuous quantities. \n\n2.3  Network  training \n\nThe  conditional  probability  matrices \nthe  probability \ndistributions  of network  sensor  feature  nodes  given  Rock  type  (and  other  parent \n\n(CPM's)  describing \n\nnodes)  are  learned  from  examples  (of rock types  along  with  the  associated  feature \nvectors  derived  from  sensor  readings  on  rock  samples  of the  given  type)  using  the \nalgorithm in  [5].  If X  is  a node  (with N  states)  with parent Y,  and  with CPM Pij = \nP(X=iIY=j),  then  each  column  is  represented  by  a  Dirichlet  distribution  (initially \nuniform)  and  assumed  independent  of  the  others. \nIf  (X) \" (XN  are  the  Dirichlet \nparameters for P(XIY=j)  then  lij =a;{fPk [6].  Given a new example  {X=i,Y=j}with \n(Xi + w.  This is a true Bayesian \nweight w the Dirichlet parameters are updated:  (Xi  7 \nlearning algorithm,  and  is  stable.  Furthermore, it is  possible to  weight each training \nsample  to  reflect  its  frequency  of occurrence  for  the  rock  type  that  generated  it. \nThis  is  especially  important  if multiple  sensor  readings  are  taken  from  a  single \nsample \n\n1 \n\n.... \nCD \n!!! \nc \no \n:;: \n'2: \nCI o \nu \nI!! \n\n,  ~  , \n\n, \n\ni\n, \n\ni \n, \n\ni i i  ..... i i i   i\n, \n, \n\n-~-2 \n.... . .. 1 ............. t ............ l ............. l~ .... 1. ........... t ............ l ............ t ............. i ........... . \n, \n.\u2022 \u00b7\u00b7\u00b7\u00b7\u00b7\u00b7t\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7i\u00b7~\u00b7\u00b7\u00b7\u00b7 \u00b7r\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7t\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7~\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7 ; \u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7i\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7i\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7j\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7 \n\u00b7\u00b7 .. \u00b7\u00b7,r .. \u00b7\u00b7\u00b7 .. \u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7 .. \u00b7\u00b7\u00b7 .. \u00b7\u00b7 .. \u00b7\u00b7\u00b7 .. \u00b7\u00b7\u00b7J\u00b7\u00b7\u00b7\u00b7 .. \u00b7\u00b7\u00b7 .. \u00b7I\u00b7 .. \u00b7\u00b7\u00b7 .. \u00b7 ... .j.. ........... .\\. ............ j .............. j ........... . \n~ ............ ~ .............. ! ............. ~ .............. ~ ........... . \n\u2022......... ~ ...... .\"..~ ............. ~ ............. ~........... \n.. r .. t ............ \u00b7I ............ i .......... \u00b7l ............ t ..........  \u2022  \u2022  \u2022  \u2022\u2022  spectrum  .. . \nr ...... t ............ i ............ t ............ \u00b7t ............ l ...... \u00b7 ..  -\n.......... ' ............ I ............ \u00b7t ............ \u00b7t ............ y........ \n\n- Image \n\n::~~ors \n\no \n\nfalse positives \n\n1 \n\nFigure  5  :  Classifier  rate  of classification  curves  using  laboratory  data  for \ntraining and testing (25% cross validation), for different sensors. \n\nThe  training  data  (gathered  from  previous  Antarctic  expeditions,  and  from  US \nlaboratory  collections\u00b7  of meteorites  and  Antarctic  rocks)  is  insufficient  to  fully \npopulate  the  (quantized)  space  on  which  the  CPM's  are  defmed,  unless  the  real \nvalued  feature  nodes are  very coarsely quantized.  To avoid this,  more  spectral data \nwas  generated  from  each  sample  spectra  by  adding  random  noise  (generated  by a \n\n\u2022 Johnson Space Center, Houston and Ohio State University, Columbus. \n\n\fnon-linear spectrometer noise model) to  it.  (This is  analogous to  the  approach used \nby [7]  for training neural networks). \n\nUsing  meteorite  and  terrestrial  rock  data  acquired  in  the  lab,  partitioned  into  75% \ntraining,  25%  testing  cross  validation  sets,  the  Rate  of Classification (ROC)  curves \nin  Figure  5  are  generated.  Note  the  superior  classification  with  spectra  versus \nclassification with color images  only.  In fact,  given a spectrum,  a color image does \nnot  improve  classification.  However,  because  it  is  easier  to  acquire  color  images \nthan spectra, they are  still useful as a sensor for preliminary screening. \n\n3  Antarctica  2000  field  results \n\nIn  January  2000  the  Nomad  robot  was  deployed  to  the  Elephant  moraine  in \nAntarctica  for  robotic  meteorite  searching  trials.  Nomad  searched  areas  known  to \ncontain  meteorites,  autonomously  acquiring  color  images  and  reflection  spectra  of \nboth  native  terrestrial  rocks  and  meteorites,  and  classifying  them.  On  January  22, \n2000  Nomad successfully identified a meteorite  amongst terrestrial rocks  on the ice \nsheet (http://www. frc.ri.cmu.edulproj ects/meteorobot2000/). \n\n1 \n\n: \n\n: \n\n: \n\n: \n\n. \n\n. \n\n. \n\n. \n\n: \n\nc \no \n~ 0.5 \nC) o u \nf \n\n;;~ :\n\n~  .~ \n\n: \n\n\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7i\u00b7\u00b7\u00b7\u00b7 .\"\" ... ~ .\u2022 ~ ........  ' \u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7~\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7i\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7~ \n\n:\u00b7\u00b7 jlrrI - \u00b7 :::1) ~=:~.d \n\n: \u2022\u2022\u2022\u2022 .(i) a prio.ri \n\n: \n\n: \n\no 0 \n\n0.5 \n\nfalse  positive  rate \n\n1 \n\nFigure  6:  Rate  of classification  curves  for  the  Nomad  robot  searching  for \nmeteorites in Antarctica, 2000 A.D. \n\nOverall  performance  (using  spectra  only,  due  to  a  problem  that  developed  with \ncamera  zoom  control)  is  indicated  by  the  ROC  performance  curves  in  Figure  6. \nThese  were  generated  from  a  test  set  of rocks  and  meteorites  (40  and  4  samples \nrespectively,  with  multiple  readings  of each)  in  a  particular  area  of the  moraine. \nFigure  6(i)  is  using  the  a priori classifier built  from  the  lab  data  (used  to  generate \nFigure  5),  acquired  prior  to  arrival  in  Antarctica.  Performance  clearly  does  not \nmatch  that  in  Figure  5.  There  is  a  notable  improvement in (ii),  the  ROC  curve  for \nthe  same  classifier  further  trained  with  field  data  acquired  by the  robot  in  the  area \n(from 8 rocks and 2 meteorites not in this test set). \n\nEven  with  retraining,  classification  is  systematically  bad  for  a  particular  class  of \nrocks  (hydro-thermally  altered  dolerites  and  basalts)  that  occurred  in  the  Elephant \nmoraine.  These  rocks  are  stained  red with  iron oxide  (rust)  whose  spectrum has  a \nvery prominent peak at 900  nm,  precisely where many meteorite  spectra also have a \npeak.  This  is  not  surprising,  given  that  most meteorites  contain  metallic  iron,  and \n\n\ftherefore  can  have  rust on the  surface.  However,  these  rocks  were  absent from the \ninitial  training  set  and  not  initially  expected  in  this  area.  Performance  is  much \nbetter if these  rocks  are  removed from the  test set (iii) and the retrained classifier is \nused. \n\n4  Conclusions \n\nWith  the  caveat that  training  be  continued  using  data  acquired by the  robot  in  the \nfield,  the Bayes network approach to  robotic rock classification is  a viable approach \nto  this  task.  Nomad  did  autonomously  identify  several  meteorites.  However,  in \nareas  with  hydro-thermally  altered  rocks  (iron-oxide  stained) \nthe  reflection \nspectrometer  must  be  supplemented  by  other  sensors,  such  as  metal  detectors, \nmagnetometers  or  more  exotic  spectrometers  (thermal  emission  or  Raman), \nobviously at greater cost. \n\nSensor  noise  and  systematic  effects  due  to  autonomous  robot placement of sensors \non  samples  in  the  unstructured  and  uncontrolled polar  environment  are  significant. \nThey  are  hard  to  know  a  priori  and  need  to  be  learned  from  data  acquired  by  the \nrobot,  and  in  field  conditions,  as  demonstrated  by  the  significant  improvement  in \nclassification achieved after field retraining. \n\nFurther work needs  to  be  done  in  selective  sensor selection,  active  modeling  of the \nlocal  geographical  distribution  of rocks,  and  recognizing  bad  sensor  readings,  but \nindications  are  that  this  can  be  done  in  a  principled  way  with  the  Bayes  network \nclassifier and will be addressed in future papers. \n\nAcknowledgments \n\nThe  authors  gratefully acknowledge  the  invaluable  assistance  of Professor William \nCassidy  of  the  University  of Pittsburgh,  Professor  Gunter  Faure  of  Ohio  State \nUniversity,  Marilyn  Lindstrom  and  the  staff  at  the  Antarctic  meteorite  curation \nfacility  of  NASA's  Johnson  Space  Center,  and  Drs.  Martial  Hebert  and  Andrew \nMoore of Carnegie Mellon University. \n\nThis  work  was  funded  by  NASA,  and  supported  in  Antarctica  by  the  National \nScience Foundation's Office of Polar Programs. \n\nReferences \n\n[1]  D.  Apostolopoulos,  M.  Wagner,  W.  Whittaker,  \"Technology  and  Field  Demonstration \nResults  in  the  Robotic  Search  for  Antarctic  Meteorites\",  Field  and  Service  Robotics \nConference, Pittsburgh, USA,  1999 \n\n[2]  Cassidy,  William,  University  of  Pittsburgh  Department  of  Geology,  personal \ncommunication,  1997. \n\n[3]  R.  Dietrich and B.  Skinner, Rocks and Minerals,  Wiley  1979. \n\n[4]  L.  Pedersen, D.  Apostolopoulos, W. Whittaker,  T.  Roush,  G.  Benedix, \"Sensing and Data \nClassification  for  Robotic  Meteorite  Search\",  Proceedings  of  SPIE  Photonics  East \nConference,  Boston,  1998. \n\n[5]  SpiegelhaJter,  David  I.,  A.  Philip  Dawid,  Steffen  L.  Lauritzen  and  Robert  G.  Cowell, \n\"Bayesian analysis in expert systems\" in Statistical Science,  8(3), p219-283.,  1993. \n\n[6]  A.  Gelman,  I.  Carlin,  H.  Stem,  D.  Rubin,  Bayesian  Data  AnalYSiS,  Chapman  &  Hall, \n1995. \n\n[7]  D.  Pomerleau,  \"Efficient  Training  of  Artificial  Neural  Networks  for  Autonomous \nNavigation\", NeurComp vol.  3 no.  1 p  88-97,  1991 \n\n\f", "award": [], "sourceid": 1798, "authors": [{"given_name": "Liam", "family_name": "Pedersen", "institution": null}, {"given_name": "Dimitrios", "family_name": "Apostolopoulos", "institution": null}, {"given_name": "William", "family_name": "Whittaker", "institution": null}]}