The emerging field of underwater sensor networks (UWSN) has a vital potential shaping the modern and future landscape that assists in measuring water quality, pollution tracking, and identification of underwater habitats. The challenging conditions in the UWSN environments raise data and security concerns in terms of outliers related to the complicated communication system, poor visibility, and limited resources. The data quality and network efficiency may be affected due to these unwanted conditions, giving rise to certain malicious activities in the network. This study aims to enhance the outlier identification process in terms of security and quality perspectives using the Long Short-Term Memory (LSTM) framework. The focus is to identify the temporal patterns and differentiate between various outliers in critical UWSN conditions. Results reveal that the proposed framework achieved high accuracy up to 95% and surpassed the other traditional machine learning models. It is worth mentioning that underwater sensor data have a complicated pattern that can be more appropriately handled using deep learning models, including LSTM, in comparison to traditional machine learning models.