Developing Hybrid Decoposition Based Machine Learning model for Soil Moisture Estimation

Soil moisture estimation remains challenging due to the inherent non-linearity and non-stationarity of hydroclimatic systems. This study develops a novel hybrid framework integrating Multivariate Empirical Mode Decomposition (MEMD) with machine learning models (SVR, RF, AdaBoost) to address these challenges. The methodology decomposes ERA-5 climatic variables using MEMD to isolate multiscale temporal patterns, followed by Boruta-RF feature selection to identify optimal predictors. Comprehensive evaluation using daily soil moisture data from Ranchi demonstrates that the MEMD-based hybrid models significantly outperform standalone approaches, with MEMD-SVR achieving the highest accuracy (MSE reduction of $\mathbf{3 5 - 6 0 \%}$, NSE improvement of $\mathbf{1 5 - 2 5 \%}$). The framework’s effectiveness is validated through multiple statistical metrics (MSE, MAPE, NSE, WI, UCI and KGE) and graphical analyses (Taylor diagrams, radar plots, scatter plots, Violin plots and time series plot), confirming its potential for precise soil moisture estimation in agricultural and hydrological applications.