Pantnagar Journal of Research

Abstract

Balancing interpretability, predictive accuracy, and temporal realism is essential for actionable crop analytics. This study integrates Bayesian calibration, Spatial Hierarchical Bayesian Regression (SHBM), latent-variable Structural Equation Modeling (SEM), and Spatio-Temporal State-Space Modeling (ST-SSM) to explain and predict wheat yield across the Tarai Bhabar belt of Uttarakhand (Udham Singh Nagar, Nainital, Champawat, Haridwar). A 20-year panel (2005–2024) of agronomic, soil, and weather variables was analyzed under a time-aware split withholding the last 20 % years for testing. Bayesian calibration of sowing date and nitrogen rate produced posterior distributions subsequently embedded in SHBM and SEM. The integrated model achieved Test R^² = 71% and RMSE = 0.120 t ha^-1, outperforming single-model baselines. Outputs were translated into site specific agronomic and policy recommendations including optimal sowing windows, nitrogen management, and soil-organic carbon augmentation. This establishes a reproducible benchmark for sub-state crop analytics combining interpretability, uncertainty quantification, and temporal decomposition.