Kai Wen Ng0000-0001-6565-947XYuk Feng Huang0000-0001-7287-3394Charlene KOO Chai-HoonAhmed El-ShafieAli Najah Ahmed2026-10-162026-10-162025-08-1510.1080/02626667.2025.2539843https://dspace-cris.utar.edu.my/handle/123456789/11764This study investigated the effectiveness of XAI feature selection (FS) techniques, namely Kernel SHAP, VG, and IG, in enhancing the performance of the proposed CNN-GRU early, slow, and late fusion models for daily water level forecasting. The results indicated that VG and IG outperformed Kernel SHAP, as well as conventional filter and embedded FS methods. To facilitate the transfer of model parameters in feature rankings prior to final model calibration, an additional transfer process was introduced to VG and IG, resulting in Transfer-VG (T-VG) and Transfer-IG (T-IG). Instead of reinitializing model parameters during calibration on selected features, T-VG and T-IG removed unimportant features by reshaping the model input layers while transferring the existing model parameters for final calibration. The findings show that T-IG improved the early- and late fusion models by 5.81% and 3.39%, respectively. Meanwhile, T-VG, which is more computationally efficient, enhanced the slow fusion model by 2.58%. © 2025 IAHS.en-USdeep learningfusionmachine learningmultivariate dataXAICalibrationFeature extractionForecastingFusion reactionsLearning systemsTransfer learningDeep learningEarly fusionFeatures selectionFusion modelLate fusionMachine-learningModeling parametersMultivariate dataWater level forecastingartificial intelligenceforecasting methodhydrological modelingmultivariate analysiswater leveljournal-article