Agricultural systems in Maharashtra, India, encounter substantial uncertainty driven by climatic variability, heterogeneous soil characteristics, and stochastic biological dynamics. Conventional artificial intelligence methodologies in precision agriculture predominantly utilize black-box prediction mechanisms that compromise interpretability and lack rigorous theoretical foundations. This investigation presents a mathematically transparent framework where crop performance dynamics are represented as discrete-time Markov chains, with finite states characterizing varying agricultural health conditions. Drawing upon empirical crop production and yield datasets from Maharashtra, transition probabilities governing movement between discrete crop states—Degraded, Suboptimal, Healthy, and Optimal—are quantified through historical observation analysis. A constrained optimization formulation is constructed to maximize long-term probabilities of favorable agricultural states while maintaining probabilistic coherence. Theoretical analysis establishes existence guarantees, asymptotic convergence properties, and optimality characterizations under convex feasibility constraints. Empirical results validate substantial performance enhancements: degraded-state persistence probability diminished from 72% to 60%, concurrent with healthy-to-optimal transition probability increases from 26% to 36%, and optimal-state retention rising from 71% to 76%. This framework delivers actionable guidance for agricultural stakeholders, synthesizing data-driven learning capabilities with mathematical rigor to strengthen crop system stability and enhance farmer welfare under environmental uncertainty.