A Learning-Based Approach to Underwater Direction of Arrival Estimation for Small Samples

Abstract

With the development of deep learning technology, the direction of arrival(DOA) estimation based on it is also booming. However, due to the difficulty in obtaining samples, underwater DOA estimation is hard to achieve the same effect as that on land. Meanwhile, underwater channel is more seriously affected by multipath which makes the neural networks have poor generalization ability. In this paper, we construct new input feature for the neural networks. Then we use transfer learning to utilize simulated data, and skillfully split the output task to make use of the multi-task learning mechanism. Experiments and simulations show that our method has good performance improvement.