Optimized Random Forest with Explainable AI for Accurate Boro Rice Yield Prediction and Real-Time Web App

A precise prediction of rice yield is vital when it comes to planning of food security and decisions making in agricultural fields. This paper involves an elaborate machine learning model to forecast the Boro rice yield by using a data set consisting of$\text{4, 4 8 0}$records and 43 agricultural, environmental, and soil characteristics. We compared five regression models: Linear Regression, Random Forest, Gradient Boosting, Support Vector Regression and K-Nearest Neighbors, optimizing hyperparameters using cross-validation grid search. The selection of feature was done using a mutual information analysis, and the most significant indicators are considered the inputs fertilizers (urea, muriate of potash, DAP), the characteristics of the soil (pH, texture, moisture) and the rainfall. The optimized Random Forests model performed better, with$\mathrm{R} \hat{\mathrm{2}}=\text{0. 9 9 9 9}$,$\text{R M S E}=\text{0. 0 0 0}$, and the computational time of 0.01 seconds. Explainable AI methods such as SHAP and Morris sensitivity analysis offered an insight into the model decision process, where it was found that yield estimates are strongly affected by soil category, cultivated area, and fertilizer usage data. A practical web application was designed to ease its use in the field, where farmers and agricultural planners can have real-time estimation of their yields. The presented model shows a great enhancement of the available methods, as it yields high scores in accuracy, as well as availability of interpretation on sustainable agricultural management.