<p>The prediction of sea surface salinity provides a scientific basis for environmental risk assessment, process design optimization and emergency management of industrial activities. To address the nonlinear and nonstationary characteristics of sea surface salinity, prediction model based on modified singular spectrum decomposition (MSSD), refined composite multivariate multiscale fuzzy entropy, complete ensemble empirical mode decomposition with adaptive noise optimized by catch fish optimization algorithm (CFOACEEMDAN), optimized hybrid parallel neural network based on improved frilled lizard optimization (IFLOHPNN) with good point set and adaptive inertia weight and error correction (EC) using bidirectional gated recurrent unit, named MSSD-CFOACEEMDAN-IFLOHPNN-EC, is proposed. The hybrid neural network used in this paper adopts a parallel architecture, and each branch processes the decomposed subsequence independently, which significantly shortens the network training and prediction time. Salinity data collected at two sites in the Gulf of Mexico are used as an example and compared with 12 other models. The proposed model is shown to have good prediction effect of sea surface salinity, among which RMSE, MAE, MAPE and R² of Site1 reached 0.01903, 0.01451, 1.4620% and 0.99954, respectively, which are better than those of the other comparison models.</p>