From dynamics to smartness: a fusion-aware jump-capture framework using information-rich SMAP datasets for high-precision soil moisture variability forecasting across oceanic and continental climates

The richness and continuity of NASA’s SMAP (Soil Moisture Active Passive) datasets provide a vital foundation for understanding soil–water dynamics, which are critical to the development of innovative and sustainable agricultural systems in Atlantic Canada’s Maritime region. However, the inherent complexity and high dimensionality of these multi-source inputs, often characterised by abrupt temporal or spatial fluctuations, can obscure relevant patterns, leading to reduced model accuracy unless they are carefully decomposed or feature optimised. To address this drawback, a meticulous smart forecasting platform has been developed, centred on a fusion-aware jump-capture decomposer (FJMD-EWT), which integrates multivariate jump-plus-mode decomposition (MJMD), multivariate variational mode decomposition (MVMD), and the empirical wavelet transform (EWT). This platform further integrates bootstrap-enhanced ridge regression with cross-validation (BS-RidgeCV) for powerful feature selection and an LSTM network optimised with a gradient-based optimisation algorithm (GBO) to generate short- and long-term surface soil moisture (upper 0–5 cm soil layer) forecasts across three agroclimatic zones in Atlantic Canada: Halifax, Charlottetown, and Moncton. Random Vector Functional Link (DeepRVFL) frameworks, including the Elastic Net regression method (Genet), Kernel Ridge Regression (KRR), GAdaboost, and Deep Ensemble Random Vector Functional Link (DeepRVFL), were integrated with GBO and FJMD-EWT to validate the robustness of the leading platform for forecasting daily soil moisture horizons of t + 7 and t + 10 days. Results revealed that the GLSTM-FJMD-EWT model consistently outperformed all others, resulting in correlation coefficient improvements of 5–16% and skill score enhancements of 30–45% in Halifax, 7–11% and 20–85% in Charlottetown, and 9–13% and up to 770% in Moncton, demonstrating its superior temporal stability and forecasting precision across diverse agroclimatic conditions. The application of BS-RidgeCV feature selection effectively reduced data dimensionality while retaining the most informative soil-climate predictors, with the elimination of (79.77%-85.50%) associated with irrelevant pool features. The GLSTM-FJMD-EWT hybrid framework effectively identified transient discontinuities and nonstationary signal changes that traditional MVMD-based models frequently missed, leading to more dependable and physically coherent predictions across diverse soil-climate systems.