Transformer-Enhanced Soft Actor-Critic with EV-Aware Reward Shaping for Maize Optimization

Optimizing fertilization and irrigation strategies is essential for improving productivity and resource efficiency in precision agriculture. Artificial intelligence (AI), particularly reinforcement learning (RL), has been increasingly explored for adaptive crop management under uncertain environmental conditions. However, many existing approaches rely on single-action formulations that struggle with joint input control, leading to economically unstable outcomes and limited policy interpretability. This study proposes a Transformer-enhanced Soft Actor-Critic (SAC) framework with expected value (EV)-aware reward shaping for maize optimization in a Decision Support System for Agrotechnology Transfer (DSSAT) Gym environment, enabling simultaneous control of fertilization and irrigation under dynamic crop-environment interactions. Unlike standard SAC implementations, the proposed framework incorporates a transformer-based state encoder for richer agronomic state representation and an EV-aware reward shaping mechanism to guide economically stable long-horizon decision-making. The proposed AI-driven approach improves economic profitability and profit stability compared with the prior state-of-the-art (SOTA) large language model (LLM)-enhanced Deep Q-Network (DQN) baseline. Behavioral analysis shows that the learned policy exhibits temporally structured decision patterns characterized by smaller-magnitude, higher-frequency actions and an associated input-efficiency trade-off. Furthermore, Shapley Additive Explanations (SHAP)-based explainable AI (XAI) analysis identifies growth-stage and crop-development variables as dominant drivers of long-horizon control decisions. Overall, the results demonstrate that the Transformer-enhanced SAC with EV-aware reward shaping provides a more profitable, financially stable, and interpretable AI-based decision-making framework for maize optimization in the DSSAT Gym environment.