An innovative sensor-based artificial intelligence system equipped with a twin delayed deep deterministic policy gradient (TD3++) algorithm and a multimodal IoT sensor network was developed to address challenges in the agri-food supply chain of the Guangdong-Hong Kong-Macao Greater Bay Area in China (GBA). Using 5.2 million sensor data points collected from 632 cold chain nodes, the system integrated the following: (1) a TD3++ routing optimizer that fused real-time global positioning system data, thermal imaging from FLIR A65 Image Temperature Sensor, and accelerometer data in a micro-electro-mechanical system; (2) a cross-modal attention mechanism; and (3) an edge-cloud detection system with hyperspectral sensor arrays using a quantized You Only Look Once version 5 model. The results showed a 23% reduction in transportation costs, a 92.7% temperature compliance, and an F1-score of 0.89 in crop disease identification. At the Shenzhen Agricultural Logistics Hub, cold chain breaches were reduced by 53% and carbon dioxide emissions by 18.3% compared with conventional long short-term memory (LSTM)-based systems. These baseline metrics for conventional LSTM performance are derived from industry standards for predictive logistics controllers. The results validate the applicability of sensor-driven AI in achieving sustainable agriculture while providing a scalable roadmap for regional cold chain modernization. Since the present study was limited to GBA and the initial deployment costs of high-precision sensor arrays, further study is required to explore hardware integration and system validation in different regions to enhance the global scalability of the framework.