A Climate Driven Decision Support System for Rice Management Using SPI-3 Prediction and Particle Swarm Optimization

Abstract

Climate variability and irregular rainfall patterns have become critical challenges affecting rice productivity, irrigation planning, and agricultural sustainability. Previous studies have primarily focused on rainfall forecasting or drought monitoring independently, with limited attention given to transforming climate predictions into actionable agricultural management strategies. This study addresses this gap by proposing an integrated climate-driven decision support framework that combines predictive drought-index modeling with optimization-based agronomic decision-making for adaptive rice field management. The proposed framework integrates satellite-based rainfall observations, seasonal climatic characteristics, and large-scale climate variability indicators to predict short-term moisture conditions represented by the three-month standardized precipitation index. The framework consists of three interconnected stages: climate prediction, optimization, and recommendation generation. In the prediction stage, a gradient boosting regression model enhanced with Bayesian hyperparameter optimization was employed to model nonlinear relationships among rainfall accumulation, lag rainfall patterns, seasonal cyclic features, and climate variability indicators. In the optimization stage, particle swarm optimization was applied to determine optimal fertilizer dosage, irrigation allocation, and harvest timing under varying climate conditions. Experimental procedures included comparative evaluations across multiple machine learning models, hyperparameter tuning strategies, and optimization iterations. The research figures and tables demonstrate the complete framework architecture, prediction performance comparisons, optimization convergence behavior, and adaptive rice management recommendations. Experimental results show that the proposed framework achieved strong predictive performance with a coefficient of determination of 0.851, a root mean square error of 0.391, and a mean absolute error of 0.322. Comparative analysis further confirmed that integrating climate variability indicators significantly improved predictive accuracy compared with baseline models using only historical rainfall information. The optimization process also demonstrated stable convergence toward climate-adaptive agronomic recommendations.