This work introduces a two-step self-supervised learning scheme, namely contrastive predictive coding (CPC), for underwater source localization. In the first step, a CPC-based self-supervised feature extractor is trained with the acoustic signals. In the second step, the encoder with frozen parameters is taken from the trained feature extractor and connected with a multi-layer perceptron (MLP) trained for source localization on a small labeled dataset. This approach is evaluated on a public dataset, SWellEx-96 Event S5, against an autoencoder (AE) scheme and a purely supervised scheme. The results indicate that the CPC scheme has the best performance and can extract the slow-changing features related to the source.