This paper addresses the challenge of underwater detection and classification in complex, acoustically cluttered environments, such as harbors, which are critical for security applications. To enhance detection accuracy, the study utilizes data augmentation and deep learning (DL) techniques. A mixed-data approach is applied to the ShipsEar dataset, integrating target vessel noise, ambient sounds, and interference from other vessels to improve classification performance. Both traditional methods, such as Maximum Likelihood Estimation (MLE), and advanced DL models, including ResNet, are used to classify these audio features. The results demonstrate that DL models, especially deep convolutional networks, significantly outperform conventional methods in accurately identifying underwater targets within noisy backgrounds when optimized with spectrogram data. The findings underscore the potential of combining traditional and modern techniques for robust underwater detection, supported by the EU Horizon SMAUG project.