Feature Selection Based on Principal Component Regression for Underwater Source Localization by Deep Learning

Underwater source localization is an important task, especially for real-time operation. Recently, machine learning methods have been combined with supervised schemes. This opens new possibilities underwater localization. However, in many real scenarios, the number of labeled datasets insufficient purely learning, and training time a deep neural network can be huge. To mitigate problem related to low available, we propose two-step framework based on semi-supervised scheme. The first step utilizes convolutional autoencoder extract latent features from whole available dataset. second performs via encoder multi-layer perceptron trained limited portion reduce time, interpretable feature selection (FS) method principal component regression proposed, which by only introducing location without other prior information. proposed approach validated public dataset SWellEx-96 Event S5. results show that has appealing accuracy robustness unseen data, when data used train gradually decreases. After FS, not stage 95% acceleration but performance becomes more robust receiver-depth accurate extremely limited.