Abstract: To address issues such as production loss of control due to cloud command interruptions in weak-network environments for smart agricultural management systems, as well as the difficulty of traditional deep learning models with large parameter sizes in meeting real-time inference demands at the edge, a smart agricultural multi-platform management system based on cloud-edge collaboration and lightweight AI is proposed.In the cloud service layer, an environmental time-series prediction model is constructed using the LightGBM algorithm, and a lightweight YOLOv8n object detection model incorporating the convolutional block attention module (CBAM) is introduced.Channel pruning reduces weights to 6.8MB while maintaining an average detection accuracy of 89.2%, significantly lowering computational overhead.In the edge control layer, a state-machine-based dual-mode switching mechanism and a threshold mirroring caching strategy are designed to enable logical autonomy of edge gateways under network disconnections.Combined with the IAP breakpoint-resume protocol, the system ensures the reliability of firmware remote updates.At the application level, a three-dimensional visualization management platform is built using WebGL and levels of detail (LOD) optimization strategies to achieve digital twin mapping of physical greenhouse states.Experimental results demonstrate that the system can trigger edge autonomy within 2 seconds after network disconnection, achieving a firmware update success rate of over 99%.The improved YOLO model achieves a 3.0% increase in detection accuracy compared to the native YOLOv8n model, with a single-frame inference time of only 20.5ms, meeting real-time monitoring requirements.The system exhibits high practical deployment and scalability value.