A novel approach for predicting soil water storage in agricultural areas: An optimized stacked ensemble model based on variational mode decomposition and empirical wavelet transform

To reconcile model accuracy and data quality in soil water storage (SWS) prediction, this study developed the VMD-EWT-NSGA-III-LSTM-CNN-TFT (VE-NLCT) model by integrating variational mode decomposition, empirical wavelet transform, and a heterogeneous stacked ensemble algorithm. Using monthly data from the Songnen Plain (1982–2024), the model was applied to forecast storage trends over 2024–2039 and analyze spatiotemporal evolution. Results showed increased uncertainty in Stage 4 (2012–2024): Coefficient of Variation rose to 0.0854, Mann-Kendall’s Tau = 0.3838, and Approximate Entropy increased progressively. The VE-NLCT decomposed raw data into six intrinsic mode functions (IMFs); IMF1/2 achieved Signal-to-Noise Ratios of 41.81 dB and 25.16 dB. Decomposed data exhibited improved stationarity (Kwiatkowski-Phillips-Schmidt-Shin reduced by 0.427–1.181) and lower overfitting risk. Posterior evaluation confirmed high accuracy (mean C = 0.085; 96.93% of sites rated “excellent” and all p-values exceeded 0.95). Forecasts revealed persistent spatial heterogeneity: southwest had the lowest storage (min 2760 m³/ha), central region the highest. Northern storage increased but remained medium low. Storage stability declined over three 5 year stages, with August–September standard deviations reaching 951 and 956 in Stage 3. The VE-NLCT outperformed benchmarks across all metrics (R² = 0.9883; error reductions of 30.72–71.09% in mean absolute error, root mean square error, mean absolute scaled error, dynamic threshold error and shape-based error; adversarial validation Area Under the Curve = 0.5342; noise immunity score = 0.911). The VE-NLCT lessens reliance on massive training data, fitting data-sparse agricultural regions, and delivering an accurate, robust SWS prediction tool for black soil agriculture areas.