NEURAL NETWORK FEATURE CLASSIFICATION

GIVEN A SET OF IMAGES OF SEA SURFACE PARAMETERS, IT IS POSSIBLE TO CLASSIFY LAND AND OCEAN AREAS. HIGH CHLOROPHYLLCONCENTRATIONS IN ESTUARIES CAN INDICATE PLANKTON BLOOM CONDITIONS. BY EXAMINING SUCH IMAGES, WETLAND, ESTUARY, AND COASTAL RESOURCES CAN BE MONITORED. HOWEVER, THE OPERATIONAL USE OF THIS INFORMATION IS HINDERED BY THE OBSCURING EFFECTS OF CLOUDS AND BY THE MASSIVE AMOUNT OF DATA TO BE REVIEWED BY A HUMAN ANALYST. ARTIFICIAL INTELLIGENCE CAN MITIGATE THESE EFFECTS. WE ENVISION A SYSTEM WHICH WILL USE AN INITIAL TRAINING SET TO BOTH COMPUTE PRINCIPAL COMPONENTS OF IMAGE VARIABILITY AND TRAIN A NEURAL NETWORK. THE NEURAL NETWORK WILL THEN OPERATE ON IMAGERY AVAILABLE FROM THE EARTH OBSERVING SYSTEM. AN EXPERT SYSTEM WILL MODIFY AND/OR COMBINE THOSE CLASSIFICATIONS AND THE RESULTS STORED AS OCEAN FEATURE OBJECTS WITHIN A GIS. THE MODIFIED CLASSIFICATION WILL BE ADDED TO THE NEURAL NETWORK TRAINING SET. THEN THE NETWORK WILL BE RETRAINED, THUS ADAPTING THE NETWORK TO THE LESSONS LEARNED BY EXPERIENCE. THE WORK PROPOSED DURING PHASE I WILL TEST THE FEASIBILITY OF THE RISKIEST PART OF SUCH A SYSTEM, NAMELY, THE ABILITY OF A NEURAL NETWORK TO GENERATE USEFUL CLASSIFICATION IN THEPRESENCE OF CLOUDS AND NOISE.