Maritime safety is crucial, and automating hostile vessel detection can greatly enhance it. Recent success with Recurrent Neural Networks (RNN) in similar tasks suggests that these models can help improve maritime safety. Feature selection is essential for machine learning models as it improves accuracy, reduces overfitting, and speeds up training. This research investigates the effectiveness of different feature selection methods combined with various RNN-based deep learning models for predicting nearby vessel behaviors. We employed univariate, decision-tree-based, random forest-based, and recursive feature elimination methods to select the top 70%, 80%, and 90% of relevant features. Using these subsets, we trained six RNN-based deep learning models: unidirectional and bidirectional RNNs, Gated Recurrent Units (GRU), and Long Short-Term Memory (LSTM) networks. We obtained the best performance using the bidirectional LSTM model combined with decision tree-based feature selection at 70% of the features. It recorded an accuracy of 92.2%, a precision of 93.0%, a recall of 92.6%, and an F1-score of 92.7%. This combination outperformed the use of the full dataset and all other feature selection methods across multiple deep-learning models. These results highlight the significant impact of feature selection on enhancing the predictive abilities of deep learning models in maritime security. The study emphasizes the importance of choosing optimal features and model architectures to improve maritime defense systems.