On-site estimation of sea state parameters is crucial for ship navigation systems' accuracy, stability, and efficiency. Extensive research has been conducted on model-based estimating methods utilizing only ship motion responses. Model-free approaches based on machine learning (ML) have recently gained popularity, and estimation from time-series of ship motion responses using deep learning (DL) methods has given promising results. Accordingly, in this study, we apply the novel, attention-based neural network (AT-NN) for estimating sea state parameters (wave height, zero-crossing period, and relative wave direction) from raw time-series data of ship pitch, heave, and roll motions. Despite using reduced input data, it has been successfully demonstrated that the proposed approaches by modified state-of-the-art techniques (based on convolutional neural networks (CNN) for regression, multivariate long short-term memory CNN, and sliding puzzle neural network) reduced estimation MSE by 23% and MAE by 16% compared to the original methods. Furthermore, the proposed technique based on AT-NN outperformed all tested methods (original and enhanced), reducing estimation MSE by up to 94% and MAE by up to 70%. Finally, we also proposed a novel approach for interpreting the uncertainty estimation of neural network outputs based on the Monte-Carlo dropout method to enhance the model's trustworthiness.