Hybrid approach to seabed characterization using shipping noise
Abstract
This proposal contains basic research needed to develop a through-the-sensor statistical inference module (TTS-SIM) that combines de,ep learning and Bayesian maximum entropy (BME) to characterize the seabed using spectrograms of shipping noise. This hybrid approach, allows for a quick estimate of the seabed type in real-time from deep learning and a more thorough estimate of seabed parameters fr,om the BME optimization. Because of the lack of labeled field data, the neural networks are trained for seabed classification on syn,thetic data simulated using a catalog of distinct seabed classes. The trained networks are then applied to measured data samples an,d link each data sample to one of the seabed classes. This work seeks to answer questions about the best way to classify seabeds an,d the network architectures that best generalize. Instead of relying on a single prediction, an ensemble approach is proposed in wh,ich different networks make real-time predictions on different slices of the measured spectrograms. An ensemble voting algorithm ca,n then be used to identify the most probable sediment as well as provide information as to the consistency of that seabed prediction, for different frequency bands. The predicted seabed type also informs the priors for the Bayesian maximum entropy (BME) optimizat,ion designed to fine tune seabed parameter estimations. The proposed work also strives to improve the BME by finding a memory-effic,ient approach to operate on multiple sensors, find appropriate reduced-order models for the seabed, and also find characteristics of, the ship motion. The entire TTS-SIM algorithm needs to be tested on additional measured data to ensure robustness and generalizabi,lity. My groups basic research is necessary to develop, test, and evaluate the TTS-SIM system and ultimately yield a robust hybrid,approach to seabed classification.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 16, 2022
- Source ID
- N000142212402
Entities
People
- Traci Neilsen
Organizations
- Brigham Young University
- Office of Naval Research
- United States Navy