This paper proposes a convolutional neural network model called Convolutional-Convolutional Neural Network (CCNN) for passive underwater acoustic source localization, using simulated time-domain pressure signals from a hydrophone array as input. The model architecture first applies two-dimensional convolutions along the temporal axis to extract long-term waveform features, followed by deeper cross-channel convolutional layers that jointly model spatial-temporal relationships while simultaneously regressing source distance and azimuth. The training data is generated through numerical simulations incorporating an 18-element foldable robotic sonar array configuration and measured sound velocity profiles from Thousand-Island Lake. Comparative experiments demonstrate CCNN’s superior performance over traditional feedforward neural network (FNN) and convolutional FNN (CFNN) models in localization accuracy and robustness within complex underwater acoustic environments, validating the approach’s practical effectiveness.