The identification of underwater acoustic signals has gained relevance across various fields of knowledge, including marine biology, for studying the behavior of aquatic species, geology, for investigating natural events, and maritime security, for vessel detection. However, the underwater environment is characterized by high levels of noise, making the accurate identification of such signals a significant challenge. In this study, we employ time-frequency analysis techniques, such as Mel-Frequency Cepstral Coefficients (MFCC) and the Continuous Wavelet Transform (CWT), to extract features from acoustic signals collected in natural environments. Furthermore, we assess the impact of applying window functions (Hamming, Blackman-Harris, and Kaiser) on the performance of machine learning models for vessel classification. The results demonstrate that the use of windowing significantly enhances model accuracy, providing a robust approach for the automatic classification of underwater signals. These findings contribute to the advancement of acoustic signal processing and machine learning techniques applied to underwater monitoring.