Sensor-Based Cross-Modal Spatiotemporal Alignment and Causal Profit Modeling for Agricultural Input Optimization

As smart agriculture gradually shifts from single-yield monitoring toward the coordinated optimization of input efficiency, resource conservation, and farm profitability, the use of multi-source agricultural data to support precise water, fertilizer, and pesticide inputs has become an important issue. To address the inconsistent sampling frequencies, heterogeneous semantic scales, and difficulty in directly modeling input–profit relationships among field images, meteorological environments, soil states, and agricultural management records, a Multimodal Agricultural Input–Output Optimization Network, termed MAION, is proposed. In this framework, a unified plot–time-window agricultural state representation is constructed through a cross-modal spatiotemporal alignment module. The potential effects of different input behaviors on yield, cost, and net profit are estimated through an input–output causal profit modeling module, and profit-driven reinforcement learning is further used to generate input strategies oriented toward long-term net profit maximization. Since the task belongs to yield, cost, and profit regression prediction and continuous agricultural decision optimization rather than classification or recognition, classification metrics such as accuracy, precision, and recall were not adopted. Instead, RMSE, MAE, $R^2$, Net Profit Improvement, Input–Output Ratio, Cumulative Reward, Policy Stability, and Regret were used for evaluation. Experimental results show that MAION achieves the best performance in yield, cost, and net profit prediction, with RMSE values of $0.587$, $0.531$, and $0.648$, respectively, and corresponding $R^2$ values of $0.914$, $0.891$, and $0.883$. These results are markedly superior to those of Random Forest, XGBoost, LSTM, GRU, Transformer, Multimodal Transformer, and reinforcement learning baseline models. In the economic decision-making experiment, MAION achieves a Net Profit Improvement of $17.68\%$, an Input–Output Ratio of $1.71$, a Cost Efficiency Gain of $15.46\%$, and a Cumulative Reward of $301.27$, while obtaining the lowest policy fluctuation and regret. The results indicate that the proposed framework can provide effective data-driven decision support for precision input, cost control, and profit optimization in smart agriculture.