A UUV simulator for generating sidescan training data for ANNs

Artificial Neural Networks (ANNs) have recently returned to popularity for a number of marine-related applications such as habitat mapping, environmental stewardship, population studies, optical measurement and tracking of aquatic populations, and more. A primary bottleneck in ANN development is the training process which can often be arduous. A typical training might involve providing hundreds to thousands of example images, with associated meta-data and classifications, to the algorithm to iteratively calculate internal node weights. The number of marine operations required to build a library of these training data sets can be prohibitively expensive due to length of time at sea (i.e. ship time), equipment purchase or rental costs, and operator costs. In an effort to significantly reduce these costs and eliminate the risks of operating at sea to generate initial training data sets, L3Harris OceanServer has developed an unmanned undersea vehicle (UUV) simulator and associated environmental simulator which can be used to generate ANN training sets without deploying physical assets and associated operators. The environmental simulator software may be generated with previously collected bathymetry and populated with known 3D objects (e.g. fishing gear, piers, sea grasses, animals, etc.). The Iver UUV is sent on a virtual mission through the simulated environment to collect data (captured in the native data formats aboard the vehicle). In this manner, training data that includes optical data, sidescan sonar data, environmental data, vehicle data, etc. can be acquired/generated in a cost-effective and deterministic environment to aid in ANN-based vehicle-behavior development before deploying software on real-world assets. This presentation will include a description of the Iver UUV Simulator and a case study on how to use the simulator to train an ANN to recognize a fish trap from sidescan sonar data.