Underwater optical wireless communication (UOWC) is a nascent technology facilitating communication and data exchange among underwater sensors. However, it faces challenges like limited bandwidth and frequent transmission failures. Modulation classification plays a crucial role in optimising spectrum allocation, ensuring reliable communication, reducing interference, enhancing network security, and enabling diverse applications in UOWC. Deep learning (DL) has succeeded in various domains but has not been extensively explored in UOWC. This study uses a Convolutional Neural Network (CNN) to classify modulation techniques in UOWC. Raw modulated signals are converted into constellation signal images and fed into the CNN for training. The performance is evaluated on a CNN pre-trained model like SqueezeNet. Simulation results show that this method achieves better classification accuracy without selecting features manually.