<p>This study integrates air temperature (Ta) and river water temperature (Tw) data to develop a reliable framework for improving Tw prediction. The primary objective is to investigate how signal decomposition, when used as a preprocessing step, can enhance the performance of machine learning models. Observations of Ta and Tw were collected from four monitoring stations located along Polish rivers. Six machine learning algorithms were employed: Random Forest Regression (RFR), Support Vector Regression (SVR), Regularized Greedy Forest (RGF), Extremely Randomized Trees (ERT), Histogram Gradient Boosting (HistGBRT), and Natural Gradient Boosting (NGBoost). These models were integrated with two signal decomposition techniques, the Time-Varying Filtering Empirical Mode Decomposition (TVF-EMD) and the Empirical Fourier Decomposition (EFD). Model evaluation was conducted under three scenarios: (i) single models using Ta only, (ii) hybrid models incorporating EFD, and (iii) hybrid models incorporating TVF-EMD. The predictive performances were assessed using RMSE, MAE, R, and NSE. The results clearly indicate that the inclusion of signal decomposition improves model accuracy. The hybrid approaches produced smaller RMSE and MAE values compared with the single models. At the Raba station, the mean RMSE and MAE were reduced by 38.73% and 35.90% using EFD, and by 32.75% and 29.18% using TVF-EMD. Even greater improvements were achieved at the Wisła station, where the reductions reached 59.69% and 59.79% for EFD, and 56.65% and 56.68% for TVF-EMD. These findings demonstrate that both EFD and TVF-EMD substantially enhance the accuracy and robustness of river water temperature predictions.</p> Graphical abstract <p></p>