Source localization with temperature variation in a laboratory tank

Environmental variability complicates deep learning applications to ocean acoustics. Transfer learning has the potential to improve the deep learning model’s ability to perform in a new environment. This work illustrates the potential of transfer learning for the case of predicting source-receiver range in a laboratory tank with changing water temperature. A four-layer convolutional neural network is trained to predict the source-receiver range. The training dataset consists of time-averaged spectral levels from chirps (50–100 kHz) measured at room temperature. Due to time-dependent ambient noise, data measured on multiple days are combined to form training datasets with 1806 samples. Early stopping is employed, and the trained models are validated on test datasets also measured in room temperature water. The trained models are then applied to data measured in warmer water, and the error in the predicted ranges increases. Transfer learning is performed using a dataset with 50 samples measured at warmer temperatures, and new testing is performed. Transfer learning refines the trained model and leads to reduced errors in the predictions. This work provides an example of how transfer learning could be applied to accommodate environmental variation in ocean applications of deep learning. [Work supported by ONR.]