Intelligent Feature Extraction Using Frozen Dictionary Learning

Modern spacecraft collect and store massive amounts of information. Analyzing this data can lead to incredible scientific discoveries and technological improvements. However, spacecraft are still limited in their ability to transmit large amounts of data. As a result, most data collected in space never makes the trip to Earth and is never analyzed or used in any meaningful way by scientists on the ground. The development of intelligent onboard algorithms will allow for more interesting data to be collected and transmitted to Earth. One way to improve real-time decision-making  by  onboard  algorithms  is  by  supplementing  the  algorithms  with  ground-based  analysis.  For  example,  computationally  expensive  machine  learning  algorithms  can  extract  meaningful  features  from  data  using  high  performance  computers  that  are  infeasible  to  send  into  space.  However, the results of the ground-based calculations can subsequently be transmitted to onboard systems, allowing them to collect and transmit data more relevant to their missions. This research project will focus on improving ground-based intelligent data understanding by developing an algorithm to intelligently learn anomalous or interesting features from synthetic aperture  radar  (SAR)  images  based  on a  technique  called  ‘frozen  dictionary  learning.’  The effectiveness  of  the  algorithm  will  be  verified  by  using  the  extracted  features  to  determine  the presence of volcanoes in SAR images of the planet Venus.  This 1-year seed project will result in a  software  library  of  feature  extraction  tools  that can  interface  with  common  machine  learning  classifiers.  The  infrastructure  resulting  from  this  project  will  lay  the  foundation  for  subsequent  research  into  ground-based feature  extraction  methods  which  can  be  used  to  improve  onboard decision making.

Contact Info