The digital rural industry has problems such as complex multi-source heterogeneous data, conflicting decision-making objectives, and weak cross-regional adaptability in the production-tourism-e-commerce chain. The requirements of intelligent development are far too high for traditional single-modality or single-objective optimisers. Therefore, this paper proposes a reinforcement learning model for industrial intelligent decision-making that is driven by multimodal data fusion in digital rural areas. In terms of methods, a multimodal dataset of remote sensing images, meteorological time series, market transactions and policy texts is first constructed, and a unified state space is generated through feature modeling and fusion. At the same time, a multi-objective reward function that combines parallel benefits, costs, risks and sustainability is used. Combine with the Transformer-Actor-Critic policy network and adaptive parameter regulation mechanism to achieve generalised decision optimisation across industries and regions. Experimentally verified on a set of 90,000 samples that covered agricultural production, rural tourism and e-commerce transactions. Based on the above results, the method proposed in this paper has improved the accuracy (to 91.2%), F1 score (to 89.6%), RMSE (to 0.118), and inference delay (to 0.031 s/step) of the baseline models such as Random Forest, DQN and Transformer RL; at the same time, the long-term cumulative reward has increased by more than 20%. Based on the above research results, the model proposed in this paper can enhance the intelligent decision-making capacity of the digital rural economy and also shows strong generalisation ability across regions. This way can provide a realistic technical route for promoting the efficient, green and sustainable development of the digital rural industry, and is feasible to be implemented in the agricultural, tourism and e-commerce sectors.
